R/adamGeneral.R
In config-i1/smooth: Forecasting Using State Space Models
Defines functions
parametersChecker
parametersChecker <- function(data, model, lags, formulaToUse, orders, constant=FALSE, arma,
outliers=c("ignore","use","select"), level=0.99,
persistence, phi, initial,
distribution=c("default","dnorm","dlaplace","dalaplace","ds","dgnorm",
"dlnorm","dinvgauss","dgamma"),
loss, h, holdout, occurrence,
ic=c("AICc","AIC","BIC","BICc"), bounds=c("traditional","usual","admissible","none"),
regressors, yName,
silent, modelDo, ParentEnvironment,
ellipsis, fast=FALSE){
# The function checks the provided parameters of adam and/or oes
##### data #####
# yName is the name of the object. It might differ from response if matrix is provided
responseName <- yName;
# If this is simulated, extract the actuals
if(is.adam.sim(data) || is.smooth.sim(data)){
data <- data$data;
}
# If this is Mdata, use all the available stuff
else if(inherits(data,"Mdata")){
h <- data$h;
holdout <- TRUE;
if(modelDo!="use"){
lags <- frequency(data$x);
}
data <- ts(c(data$x,data$xx),start=start(data$x),frequency=frequency(data$x));
}
# Extract index from the object in order to use it later
### tsibble has its own index function, so shit happens because of it...
if(inherits(data,"tbl_ts")){
yIndex <- data[[1]];
if(any(duplicated(yIndex))){
warning(paste0("You have duplicated time stamps in the variable ",yName,
". I will refactor this."),call.=FALSE);
yIndex <- yIndex[1] + c(1:length(data[[1]])) * diff(tail(yIndex,2));
}
}
else{
yIndex <- try(time(data),silent=TRUE);
# If we cannot extract time, do something
if(inherits(yIndex,"try-error")){
if(!is.data.frame(data) && !is.null(dim(data))){
yIndex <- as.POSIXct(rownames(data));
}
else if(is.data.frame(data)){
yIndex <- c(1:nrow(data));
}
else{
yIndex <- c(1:length(data));
}
}
}
yClasses <- class(data);
# If this is something like a matrix
if(!is.null(ncol(data)) && ncol(data)>1){
xregData <- data;
# Get rid of the bloody tibble class. Gives me headaches!
if(inherits(data,"tbl_df") || inherits(data,"tbl")){
data <- as.data.frame(data);
}
if(!is.null(formulaToUse)){
responseName <- all.vars(formulaToUse)[1];
y <- data[,responseName];
}
else{
responseName <- colnames(xregData)[1];
# If we deal with data.table / tibble / data.frame, the syntax is different.
# We don't want to import specific classes, so just use inherits()
if(inherits(data,"tbl_ts")){
# With tsibble we cannot extract explanatory variables easily...
y <- data$value;
}
else if(inherits(data,"data.table") || inherits(data,"data.frame")){
y <- data[[1]];
}
else if(inherits(data,"zoo")){
if(ncol(data)>1){
xregData <- as.data.frame(data);
}
y <- zoo(data[,1],order.by=time(data));
}
else{
y <- data[,1];
}
}
# Give the indeces another try
yIndex <- try(time(y),silent=TRUE);
# If we cannot extract time, do something
if(inherits(yIndex,"try-error")){
if(!is.null(dim(data))){
yIndex <- try(as.POSIXct(rownames(data)),silent=TRUE);
if(inherits(yIndex,"try-error")){
yIndex <- c(1:nrow(data));
}
}
else{
yIndex <- c(1:length(y));
}
}
else{
yClasses <- class(y);
}
}
else{
xregData <- NULL;
if(!is.null(ncol(data)) && !is.null(colnames(data)[1])){
responseName <- colnames(data)[1];
y <- data[,1];
}
else{
y <- data;
}
}
# Make the response a secure name
responseName <- make.names(responseName);
# Define obs, the number of observations of in-sample
obsAll <- length(y) + (1 - holdout)*h;
obsInSample <- length(y) - holdout*h;
if(obsInSample<=0){
stop("The number of in-sample observations is not positive. Cannot do anything.",
call.=FALSE);
}
# Interpolate NAs using fourier + polynomials
yNAValues <- is.na(y);
if(any(yNAValues)){
warning("Data contains NAs. The values will be ignored during the model construction.",call.=FALSE);
X <- cbind(1,poly(c(1:obsAll),degree=min(max(trunc(obsAll/10),1),5)),
sinpi(matrix(c(1:obsAll)*rep(c(1:max(max(lags),10)),each=obsAll)/max(max(lags),10), ncol=max(max(lags),10))));
# If we deal with purely positive data, take logarithms to deal with multiplicative seasonality
if(any(y[!yNAValues]<=0)){
lmFit <- .lm.fit(X[!yNAValues,,drop=FALSE], matrix(y[!yNAValues],ncol=1));
y[yNAValues] <- (X %*% coef(lmFit))[yNAValues];
}
else{
lmFit <- .lm.fit(X[!yNAValues,,drop=FALSE], matrix(log(y[!yNAValues]),ncol=1));
y[yNAValues] <- exp(X %*% coef(lmFit))[yNAValues];
}
if(!is.null(xregData)){
xregData[yNAValues,responseName] <- y[yNAValues];
}
rm(X);
# Clean memory if have a big object
if(obsInSample>10000){
gc(verbose=FALSE);
}
}
# If this is just a numeric variable, use ts class
if(all(yClasses=="integer") || all(yClasses=="numeric") ||
all(yClasses=="data.frame") || all(yClasses=="matrix")){
if(any(class(yIndex) %in% c("POSIXct","Date"))){
yClasses <- "zoo";
}
else{
yClasses <- "ts";
}
}
yFrequency <- frequency(y);
yStart <- yIndex[1];
yInSample <- matrix(y[1:obsInSample],ncol=1);
if(holdout){
yForecastStart <- yIndex[obsInSample+1];
yHoldout <- matrix(y[-c(1:obsInSample)],ncol=1);
yForecastIndex <- yIndex[-c(1:obsInSample)];
yInSampleIndex <- yIndex[c(1:obsInSample)];
yIndexAll <- yIndex;
}
else{
yInSampleIndex <- yIndex;
yIndexDiff <- diff(tail(yIndex,2));
yForecastStart <- yIndex[obsInSample]+yIndexDiff;
if(any(yClasses=="ts")){
yForecastIndex <- yIndex[obsInSample]+as.numeric(yIndexDiff)*c(1:max(h,1));
}
else{
yForecastIndex <- yIndex[obsInSample]+yIndexDiff*c(1:max(h,1));
}
yHoldout <- NULL;
yIndexAll <- c(yIndex,yForecastIndex);
}
if(!is.numeric(yInSample)){
stop("The provided data is not numeric! Can't construct any model!", call.=FALSE);
}
# If the user asked for trend, but it's not in the data, add it
if(!is.null(formulaToUse) &&
any(all.vars(formulaToUse)=="trend") && all(colnames(xregData)!="trend")){
if(!is.null(xregData)){
xregData <- cbind(xregData,trend=c(1:obsAll));
}
else{
xregData <- cbind(y=y,trend=c(1:obsAll));
}
}
# Number of parameters to estimate / provided
parametersNumber <- matrix(0,2,5,
dimnames=list(c("Estimated","Provided"),
c("nParamInternal","nParamXreg","nParamOccurrence","nParamScale","nParamAll")));
#### Check what is used for the model ####
if(!is.character(model)){
stop(paste0("Something strange is provided instead of character object in model: ",
paste0(model,collapse=",")),call.=FALSE);
}
# Predefine models pool for a model selection
modelsPool <- NULL;
if(!fast){
# Deal with the list of models. Check what has been provided. Stop if there is a mistake.
if(length(model)>1){
if(any(nchar(model)>4) || any(nchar(model)<3)){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[nchar(model)>4 | nchar(model)<3],collapse=",")),call.=FALSE);
}
else if(any(substr(model,1,1)!="A" & substr(model,1,1)!="M" & substr(model,1,1)!="C")){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[substr(model,1,1)!="A" & substr(model,1,1)!="M"],collapse=",")),
call.=FALSE);
}
else if(any(substr(model,2,2)!="N" & substr(model,2,2)!="A" &
substr(model,2,2)!="M" & substr(model,2,2)!="C")){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[substr(model,2,2)!="N" & substr(model,2,2)!="A" &
substr(model,2,2)!="M"],collapse=",")),call.=FALSE);
}
else if(any(substr(model,3,3)!="N" & substr(model,3,3)!="A" &
substr(model,3,3)!="M" & substr(model,3,3)!="d" & substr(model,3,3)!="C")){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[substr(model,3,3)!="N" & substr(model,3,3)!="A" &
substr(model,3,3)!="M" & substr(model,3,3)!="d"],collapse=",")),
call.=FALSE);
}
else if(any(nchar(model)==4 & substr(model,4,4)!="N" & substr(model,4,4)!="A" &
substr(model,4,4)!="M" & substr(model,4,4)!="C")){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[nchar(model)==4 & substr(model,4,4)!="N" &
substr(model,4,4)!="A" & substr(model,4,4)!="M"],collapse=",")),
call.=FALSE);
}
else{
modelsPoolCombiner <- (substr(model,1,1)=="C" | substr(model,2,2)=="C" |
substr(model,3,3)=="C" | substr(model,4,4)=="C");
modelsPool <- model[!modelsPoolCombiner];
modelsPool <- unique(modelsPool);
if(any(modelsPoolCombiner)){
if(any(substr(model,nchar(model),nchar(model))!="N")){
model <- "CCC";
}
else{
model <- "CCN";
}
}
else{
model <- c("Z","Z","Z");
if(all(substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="N")){
model[3] <- "N";
}
if(all(substr(modelsPool,2,2)=="N")){
model[2] <- "N";
}
model <- paste0(model,collapse="");
}
}
}
}
# If chosen model is "AAdN" or anything like that, we are taking the appropriate values
if(nchar(model)==4){
Etype <- substr(model,1,1);
Ttype <- substr(model,2,2);
Stype <- substr(model,4,4);
damped <- TRUE;
if(substr(model,3,3)!="d"){
message(paste0("You have defined a strange model: ", model,
". Switching to ", paste0(Etype,Ttype,"d",Stype)));
model <- paste0(Etype,Ttype,"d",Stype);
}
}
else if(nchar(model)==3){
Etype <- substr(model,1,1);
Ttype <- substr(model,2,2);
Stype <- substr(model,3,3);
if(any(Ttype==c("Z","X","Y"))){
damped <- TRUE;
}
else{
damped <- FALSE;
}
}
else{
message(paste0("You have defined a strange model: ",model));
message("Switching to 'ZZZ'");
model <- "ZZZ";
Etype <- "Z";
Ttype <- "Z";
Stype <- "Z";
damped <- TRUE;
}
# Define if we want to select or combine models... or do none of the above.
if(is.null(modelsPool)){
if(any(unlist(strsplit(model,""))=="C")){
modelDo <- "combine";
if(Etype=="C"){
Etype <- "Z";
}
if(Ttype=="C"){
Ttype <- "Z";
}
if(Stype=="C"){
Stype <- "Z";
}
}
else if(any(unlist(strsplit(model,""))=="Z") ||
any(unlist(strsplit(model,""))=="X") ||
any(unlist(strsplit(model,""))=="Y") ||
any(unlist(strsplit(model,""))=="F") ||
any(unlist(strsplit(model,""))=="P")){
modelDo <- "select";
# The full test, sidestepping branch and bound
if(any(unlist(strsplit(model,""))=="F")){
modelsPool <- c("ANN","AAN","AAdN","AMN","AMdN",
"ANA","AAA","AAdA","AMA","AMdA",
"ANM","AAM","AAdM","AMM","AMdM",
"MNN","MAN","MAdN","MMN","MMdN",
"MNA","MAA","MAdA","MMA","MMdA",
"MNM","MAM","MAdM","MMM","MMdM");
# Remove models from pool if specific elements are provided
if(Etype!="F"){
modelsPool <- modelsPool[substr(modelsPool,1,1)==Etype];
}
else{
Etype[] <- "Z"
}
if(Ttype!="F"){
modelsPool <- modelsPool[substr(modelsPool,2,2)==Ttype];
}
else{
Ttype[] <- "Z"
}
if(Stype!="F"){
modelsPool <- modelsPool[substr(modelsPool,nchar(modelsPool),nchar(modelsPool))==Stype];
}
else{
Stype[] <- "Z"
}
model <- "FFF";
}
# The test for pure models only
if(any(unlist(strsplit(model,""))=="P")){
modelsPool <- c("ANN","AAN","AAdN","ANA","AAA","AAdA",
"MNN","MMN","MMdN","MNM","MMM","MMdM");
# Remove models from pool if specific elements are provided
if(Etype!="P"){
modelsPool <- modelsPool[substr(modelsPool,1,1)==Etype];
}
else{
Etype[] <- "Z"
}
if(Ttype!="P"){
modelsPool <- modelsPool[substr(modelsPool,2,2)==Ttype];
}
else{
Ttype[] <- "Z"
}
if(Stype!="P"){
modelsPool <- modelsPool[substr(modelsPool,nchar(modelsPool),nchar(modelsPool))==Stype];
}
else{
Stype[] <- "Z"
}
model <- "PPP";
}
}
else{
modelDo <- "estimate";
}
if(Etype=="X"){
Etype <- "A";
}
else if(Etype=="Y"){
Etype <- "M";
}
}
else{
if(any(unlist(strsplit(model,""))=="C")){
modelDo <- "combine";
}
else{
modelDo <- "select";
}
}
#### Check the components of model ####
### Check error type
if(all(Etype!=c("Z","X","Y","A","M","C","N"))){
warning(paste0("Wrong error type: ",Etype,". Should be 'Z', 'X', 'Y', 'C', 'A', 'M' or 'N'. ",
"Changing to 'Z'"),call.=FALSE);
Etype <- "Z";
modelDo <- "select";
}
# If the error is "N", then switch off the ETS model
if(Etype=="N"){
componentsNamesETS <- NULL;
componentsNumberETS <- 0;
Ttype[] <- "N";
damped[] <- FALSE;
Stype[] <- "N";
modelDo[] <- "estimate";
modelsPool[] <- NULL;
etsModel <- FALSE;
}
else{
componentsNamesETS <- "level";
componentsNumberETS <- 1;
etsModel <- TRUE;
}
if(etsModel){
### Check trend type
if(all(Ttype!=c("Z","X","Y","N","A","M","C"))){
warning(paste0("Wrong trend type: ",Ttype,". Should be 'Z', 'X', 'Y', 'C' 'N', 'A', or 'M'. ",
"Changing to 'Z'"),call.=FALSE);
Ttype <- "Z";
modelDo <- "select";
}
modelIsTrendy <- (Ttype!="N");
if(modelIsTrendy){
componentsNamesETS <- c(componentsNamesETS,"trend");
componentsNumberETS[] <- componentsNumberETS+1;
}
}
else{
modelIsTrendy <- FALSE;
}
#### Check the lags vector ####
if(any(c(lags)<0)){
stop("Right! Why don't you try complex lags then, mister smart guy?",call.=FALSE);
}
# If there are zero lags, drop them
if(any(lags==0)){
lags <- lags[lags!=0];
}
# Form the lags based on the provided stuff. Get rid of ones and leave unique seasonals
# Add one for the level
lags <- c(1,unique(lags[lags>1]));
# Warning if the lags length is higher than the sample size
if(max(lags) >= obsInSample){
warning("The maximum lags value is ", max(lags),
", while the sample size is ", obsInSample,
". I cannot guarantee that I'll be able to fit the model.",
call.=FALSE);
}
#### ARIMA term ####
# This should be available for pure models only
if(is.list(orders)){
arOrders <- orders$ar;
iOrders <- orders$i;
maOrders <- orders$ma;
select <- orders$select;
if(is.null(select)){
select <- FALSE;
}
}
else if(is.vector(orders)){
arOrders <- orders[1];
iOrders <- orders[2];
maOrders <- orders[3];
select <- FALSE;
}
else{
select <- FALSE;
}
# If there is arima, prepare orders
if(sum(c(arOrders,iOrders,maOrders))>0){
arimaModel <- TRUE;
# See if AR is needed
arRequired <- FALSE;
if(sum(arOrders)>0){
arRequired[] <- TRUE;
}
# See if I is needed
iRequired <- FALSE;
if(sum(iOrders)>0){
iRequired[] <- TRUE;
}
# See if I is needed
maRequired <- FALSE;
if(sum(maOrders)>0){
maRequired[] <- TRUE;
}
# Define maxOrder and make all the values look similar (for the polynomials)
maxOrder <- max(length(arOrders),length(iOrders),length(maOrders),length(lags));
if(length(arOrders)!=maxOrder){
arOrders <- c(arOrders,rep(0,maxOrder-length(arOrders)));
}
if(length(iOrders)!=maxOrder){
iOrders <- c(iOrders,rep(0,maxOrder-length(iOrders)));
}
if(length(maOrders)!=maxOrder){
maOrders <- c(maOrders,rep(0,maxOrder-length(maOrders)));
}
if(length(lags)!=maxOrder){
lagsNew <- c(lags,rep(0,maxOrder-length(lags)));
arOrders <- arOrders[lagsNew!=0];
iOrders <- iOrders[lagsNew!=0];
maOrders <- maOrders[lagsNew!=0];
}
# Define the non-zero values. This is done via the calculation of orders of polynomials
ariValues <- list(NA);
maValues <- list(NA);
for(i in 1:length(lags)){
ariValues[[i]] <- c(0,min(1,arOrders[i]):arOrders[i])
if(iOrders[i]!=0){
ariValues[[i]] <- c(ariValues[[i]],1:iOrders[i]+arOrders[i]);
}
ariValues[[i]] <- unique(ariValues[[i]] * lags[i]);
maValues[[i]] <- unique(c(0,min(1,maOrders[i]):maOrders[i]) * lags[i]);
}
# Produce ARI polynomials
ariLengths <- unlist(lapply(ariValues,length));
ariPolynomial <- array(0,ariLengths);
for(i in 1:length(ariValues)){
if(i==1){
ariPolynomial <- ariPolynomial + array(ariValues[[i]], ariLengths);
}
else{
ariPolynomial <- ariPolynomial + array(rep(ariValues[[i]],each=prod(ariLengths[1:(i-1)])),
ariLengths);
}
}
# Produce MA polynomials
maLengths <- unlist(lapply(maValues,length));
maPolynomial <- array(0,maLengths);
for(i in 1:length(maValues)){
if(i==1){
maPolynomial <- maPolynomial + array(maValues[[i]], maLengths);
}
else{
maPolynomial <- maPolynomial + array(rep(maValues[[i]],each=prod(maLengths[1:(i-1)])),
maLengths);
}
}
# What are the non-zero ARI and MA polynomials?
### What are their positions in transition matrix?
nonZeroARI <- unique(matrix(c(ariPolynomial)[-1],ncol=1));
nonZeroMA <- unique(matrix(c(maPolynomial)[-1],ncol=1));
# Lags for the ARIMA components
lagsModelARIMA <- matrix(sort(unique(c(nonZeroARI,nonZeroMA))),ncol=1);
nonZeroARI <- cbind(nonZeroARI+1,which(lagsModelARIMA %in% nonZeroARI));
nonZeroMA <- cbind(nonZeroMA+1,which(lagsModelARIMA %in% nonZeroMA));
# Number of components
componentsNumberARIMA <- length(lagsModelARIMA);
# Their names
componentsNamesARIMA <- paste0("ARIMAState",c(1:componentsNumberARIMA));
# If all orders are zero, drop ARIMA part
if(all(c(arOrders,iOrders,maOrders)==0)){
arOrders <- NULL;
iOrders <- NULL;
maOrders <- NULL;
arimaModel <- FALSE;
arRequired <- arEstimate <- FALSE;
iRequired <- FALSE;
maRequired <- maEstimate <- FALSE;
lagsModelARIMA <- initialArimaNumber <- 0;
componentsNumberARIMA <- 0;
componentsNamesARIMA <- NULL;
nonZeroARI <- NULL;
nonZeroMA <- NULL;
}
else{
# Number of initials needed. This is based on the longest one. The others are just its transformations
initialArimaNumber <- max(lagsModelARIMA);
}
}
else{
arOrders <- NULL;
iOrders <- NULL;
maOrders <- NULL;
arimaModel <- FALSE;
arRequired <- arEstimate <- FALSE;
iRequired <- FALSE;
maRequired <- maEstimate <- FALSE;
lagsModelARIMA <- initialArimaNumber <- 0;
componentsNumberARIMA <- 0;
componentsNamesARIMA <- NULL;
nonZeroARI <- NULL;
nonZeroMA <- NULL;
select <- FALSE;
}
if(etsModel){
modelIsSeasonal <- Stype!="N";
}
else{
modelIsSeasonal <- FALSE;
}
# Lags of the model used inside the functions
lagsModel <- matrix(lags,ncol=1);
# If we have a trend add one more lag
if(modelIsTrendy){
lagsModel <- rbind(1,lagsModel);
}
# If we don't have seasonality, remove seasonal lag
if(!modelIsSeasonal && any(lagsModel>1)){
lagsModel <- lagsModel[lagsModel==1,,drop=FALSE];
}
# If this is non-seasonal model and there are no seasonal ARIMA lags, trim the original lags
if(!modelIsSeasonal && all(c(arOrders[lags>1],iOrders[lags>1],maOrders[lags>1])==0) && any(lags>1)){
arOrders <- arOrders[lags==1];
iOrders <- iOrders[lags==1];
maOrders <- maOrders[lags==1];
lags <- lags[lags==1];
}
# Lags of the model
lagsModelSeasonal <- lagsModel[lagsModel>1];
lagsModelMax <- max(lagsModel);
lagsLength <- length(lagsModel);
if(etsModel){
#### Check the seasonal model vs lags ####
if(all(Stype!=c("Z","X","Y","N","A","M","C"))){
warning(paste0("Wrong seasonality type: ",Stype,". Should be 'Z', 'X', 'Y', 'C', 'N', 'A' or 'M'. ",
"Setting to 'Z'."),call.=FALSE);
if(lagsModelMax==1){
Stype <- "N";
modelIsSeasonal <- FALSE;
}
else{
Stype <- "Z";
modelDo <- "select";
}
}
if(all(modelIsSeasonal,lagsModelMax==1)){
if(all(Stype!=c("Z","X","Y"))){
warning(paste0("Cannot build the seasonal model on data with the unity lags.\n",
"Switching to non-seasonal model: ETS(",substr(model,1,nchar(model)-1),"N)"));
}
Stype <- "N";
modelIsSeasonal <- FALSE;
substr(model,nchar(model),nchar(model)) <- "N";
}
# Check the pool of models to combine if it was decided that the data is not seasonal
if(!modelIsSeasonal && !is.null(modelsPool)){
modelsPool <- modelsPool[substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="N"];
}
# Check the type of seasonal
if(Stype!="N"){
componentsNamesETS <- c(componentsNamesETS,"seasonal");
componentsNumberETS[] <- componentsNumberETS+1;
componentsNumberETSSeasonal <- 1;
}
else{
componentsNumberETSSeasonal <- 0;
}
# Check, whether the number of lags and the number of components are the same
if(lagsLength>componentsNumberETS){
if(Stype!="N"){
componentsNamesETS <- c(componentsNamesETS[-length(componentsNamesETS)],
paste0("seasonal",c(1:(lagsLength-componentsNumberETS-1))));
componentsNumberETSSeasonal[] <- lagsLength-componentsNumberETS+1;
componentsNumberETS[] <- lagsLength;
}
else{
lagsModel <- lagsModel[1:componentsNumberETS,,drop=FALSE];
lagsModelMax <- max(lagsModel);
lagsLength <- length(lagsModel);
}
}
else if(lagsLength<componentsNumberETS){
stop("The number of components of the model is smaller than the number of provided lags", call.=FALSE);
}
}
else{
componentsNumberETSSeasonal <- 0;
}
outliers <- match.arg(outliers);
if(outliers!="ignore"){
select <- TRUE;
}
if(!fast){
#### Distribution selected ####
distribution <- match.arg(distribution[1], c("default","dnorm","dlaplace","dalaplace","ds","dgnorm",
"dlnorm","dinvgauss","dgamma"));
}
if(select){
assign("distribution",distribution,ParentEnvironment);
assign("outliers",outliers,ParentEnvironment);
# This stuff is needed for switch to auto.adam.
return(list(select=select));
}
#### Loss function type ####
if(is.function(loss)){
lossFunction <- loss;
loss <- "custom";
multisteps <- FALSE;
}
else{
loss <- match.arg(loss[1],c("likelihood","MSE","MAE","HAM","LASSO","RIDGE",
"MSEh","TMSE","GTMSE","MSCE",
"MAEh","TMAE","GTMAE","MACE",
"HAMh","THAM","GTHAM","CHAM","GPL",
"aMSEh","aTMSE","aGTMSE","aMSCE","aGPL"));
if(any(loss==c("MSEh","TMSE","GTMSE","MSCE","MAEh","TMAE","GTMAE","MACE",
"HAMh","THAM","GTHAM","CHAM","GPL",
"aMSEh","aTMSE","aGTMSE","aMSCE","aGPL"))){
if(!is.null(h) && h>0){
multisteps <- TRUE;
}
else{
stop("The horizon \"h\" needs to be specified and be positive in order for the multistep loss to work.",
call.=FALSE);
multisteps <- FALSE;
}
}
else{
multisteps <- FALSE;
}
lossFunction <- NULL;
}
#### Explanatory variables: xregModel and regressors ####
regressors <- match.arg(regressors,c("use","select","adapt"));
xregModel <- !is.null(xregData);
#### Persistence provided ####
# Vectors for persistence of different components
persistenceLevel <- NULL;
persistenceTrend <- NULL;
persistenceSeasonal <- NULL;
persistenceXreg <- NULL;
# InitialEstimate vectors, defining what needs to be estimated
persistenceEstimate <- persistenceLevelEstimate <- persistenceTrendEstimate <-
persistenceXregEstimate <- TRUE;
# persistence of seasonal is a vector, not a scalar, because we can have several lags
persistenceSeasonalEstimate <- rep(TRUE,componentsNumberETSSeasonal);
if(!is.null(persistence)){
if(all(modelDo!=c("estimate","use"))){
warning(paste0("Predefined persistence can only be used with ",
"preselected ETS model.\n",
"Changing to estimation of persistence values."),call.=FALSE);
persistence <- NULL;
persistenceEstimate <- TRUE;
}
else{
# If it is a list
if(is.list(persistence)){
# If this is a named list, then extract stuff using names
if(!is.null(names(persistence))){
if(!is.null(persistence$level)){
persistenceLevel <- persistence$level;
}
else if(!is.null(persistence$alpha)){
persistenceLevel <- persistence$alpha;
}
if(!is.null(persistence$trend)){
persistenceTrend <- persistence$trend;
}
else if(!is.null(persistence$beta)){
persistenceTrend <- persistence$beta;
}
if(!is.null(persistence$seasonal)){
persistenceSeasonal <- persistence$seasonal;
}
else if(!is.null(persistence$gamma)){
persistenceSeasonal <- persistence$gamma;
}
if(!is.null(persistence$xreg)){
persistenceXreg <- persistence$xreg;
}
else if(!is.null(persistence$delta)){
persistenceXreg <- persistence$delta;
}
}
else{
if(!is.null(persistence[[1]])){
persistenceLevel <- persistence[[1]];
}
if(!is.null(persistence[[2]])){
persistenceTrend <- persistence[[2]];
}
if(!is.null(persistence[[3]])){
persistenceSeasonal <- persistence[[3]];
}
if(!is.null(persistence[[4]])){
persistenceXreg <- persistence[[4]];
}
}
# Define estimate variables
if(!is.null(persistenceLevel)){
persistenceLevelEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
if(!is.null(persistenceTrend)){
persistenceTrendEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
if(!is.null(persistenceSeasonal)){
if(is.list(persistenceSeasonal)){
persistenceSeasonalEstimate[] <- length(persistenceSeasonal)==length(lagsModelSeasonal);
}
else{
persistenceSeasonalEstimate[] <- FALSE;
}
parametersNumber[2,1] <- parametersNumber[2,1] + length(unlist(persistenceSeasonal));
}
if(!is.null(persistenceXreg)){
persistenceXregEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + length(persistenceXreg);
}
}
else if(is.numeric(persistence)){
# If it is smaller... We don't know the length of xreg yet at this stage
if(length(persistence)<lagsLength){
warning(paste0("Length of persistence vector is wrong! ",
"Changing to estimation of persistence vector values."),
call.=FALSE);
persistence <- NULL;
persistenceEstimate <- TRUE;
}
else{
# If there ARIMA elements, remove them
if(any(substr(names(persistence),1,3)=="psi")){
persistence <- persistence[substr(names(persistence),1,3)!="psi"];
}
j <- 0;
if(etsModel){
j <- j+1;
persistenceLevel <- as.vector(persistence)[1];
names(persistenceLevel) <- "alpha";
if(modelIsTrendy && length(persistence)>j){
j <- j+1;
persistenceTrend <- as.vector(persistence)[j];
names(persistenceTrend) <- "beta";
}
if(Stype!="N" && length(persistence)>j){
persistenceSeasonal <- as.vector(persistence)[j+c(1:length(lagsModelSeasonal))];
names(persistenceSeasonal) <- paste0("gamma",c(1:length(persistenceSeasonal)));
j <- j+length(lagsModelSeasonal);
}
}
if(xregModel && length(persistence)>j){
if(j>0){
persistenceXreg <- persistence[-c(1:j)];
}
else{
persistenceXreg <- persistence;
}
# If there are names, make sure that only deltas are used
if(!is.null(names(persistenceXreg))){
persistenceXreg <- persistenceXreg[substr(names(persistenceXreg),1,5)=="delta"];
}
else{
names(persistenceXreg) <- paste0("delta",c(1:length(persistenceXreg)));
}
}
persistenceEstimate[] <- persistenceLevelEstimate[] <- persistenceTrendEstimate[] <-
persistenceXregEstimate[] <- persistenceSeasonalEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + length(persistence);
# bounds <- "none";
}
}
else{
warning(paste0("Persistence is not a numeric vector!\n",
"Changing to estimation of persistence vector values."),call.=FALSE);
persistence <- NULL;
persistenceEstimate <- TRUE;
}
}
}
else{
persistenceEstimate <- TRUE;
}
# Make sure that only important elements are estimated.
if(!etsModel){
persistenceLevelEstimate[] <- FALSE;
}
if(!etsModel || Ttype=="N"){
persistenceTrendEstimate[] <- FALSE;
persistenceTrend <- NULL;
}
if(!etsModel || Stype=="N"){
persistenceSeasonalEstimate[] <- FALSE;
persistenceSeasonal <- NULL;
}
if(!xregModel){
persistenceXregEstimate[] <- FALSE;
persistenceXreg <- NULL;
}
#### Phi ####
if(etsModel){
if(!is.null(phi)){
if(all(modelDo!=c("estimate","use"))){
warning(paste0("Predefined phi can only be used with preselected ETS model.\n",
"Changing to estimation."),call.=FALSE);
phi <- 0.95;
phiEstimate <- TRUE;
}
else{
if(!is.numeric(phi) & (damped)){
warning(paste0("Provided value of phi is meaningless. phi will be estimated."),
call.=FALSE);
phi <- 0.95;
phiEstimate <- TRUE;
}
else if(is.numeric(phi) & (phi<0 | phi>2)){
warning(paste0("Damping parameter should lie in (0, 2) region. ",
"Changing to the estimation of phi."),call.=FALSE);
phi[] <- 0.95;
phiEstimate <- TRUE;
}
else{
phiEstimate <- FALSE;
if(damped){
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
}
}
}
else{
if(damped){
phiEstimate <- TRUE;
phi <- 0.95;
}
else{
phiEstimate <- FALSE;
phi <- 1;
}
}
}
else{
phi <- 1;
phiEstimate <- FALSE;
}
#### Lags for ARIMA ####
if(arimaModel){
lagsModelAll <- rbind(lagsModel,lagsModelARIMA);
lagsModelMax <- max(lagsModel);
}
else{
lagsModelAll <- lagsModel;
}
#### This needs to be amended after developing the first prototype! ####
# If we have the zoo class and weird lags, amend profiles
# Weird lags (dst and fractional): 24, 24*2 (half hour), 52, 24*4 (15 minutes), 7*24, 7*48, 365, 24*52, 24*365
# if(any(yClasses=="zoo") && any(lags %in% c(24, 48, 52, 96, 168, 336, 365, 1248, 8760))){
# For hourly, half-hourly and quarter hour data, just amend the profiles for DST.
# For daily, repeat profile of 28th on 29th February.
# For weekly, repeat the last week, when we have 53 instead of 52.
# }
#### Occurrence variable ####
if(is.occurrence(occurrence)){
oesModel <- occurrence;
occurrence <- oesModel$occurrence;
if(occurrence=="provided"){
occurrenceModelProvided <- FALSE;
}
else{
occurrenceModelProvided <- TRUE;
}
pFitted <- matrix(fitted(oesModel), obsInSample, 1);
}
else{
occurrenceModelProvided <- FALSE;
oesModel <- NULL;
pFitted <- matrix(1, obsInSample, 1);
}
pForecast <- rep(NA,h);
# If it is logical, convert to numeric
if(is.logical(occurrence)){
occurrence <- occurrence*1;
}
if(is.numeric(occurrence)){
# If it is data, then it should correspond to the in-sample.
if(all(occurrence==1)){
occurrence <- "none";
}
else{
if(any(occurrence<0,occurrence>1)){
warning(paste0("Parameter 'occurrence' should contain values between zero and one.\n",
"Converting to appropriate vector."),call.=FALSE);
occurrence[] <- (occurrence!=0)*1;
}
# "provided", meaning that we have been provided the values of p
pFitted[] <- occurrence[1:obsInSample];
# Create forecasted values for occurrence
if(h>0){
if(length(occurrence)>obsInSample){
pForecast <- occurrence[-c(1:obsInSample)];
}
else{
pForecast <- rep(tail(occurrence,1),h);
}
if(length(pForecast)>h){
pForecast <- pForecast[1:h];
}
else if(length(pForecast)<h){
pForecast <- c(pForecast,rep(tail(pForecast,1),h-length(pForecast)));
}
}
else{
pForecast <- NA;
}
occurrence <- "provided";
oesModel <- list(fitted=pFitted,forecast=pForecast,occurrence="provided");
}
}
occurrence <- match.arg(occurrence[1],c("none","auto","fixed","general","odds-ratio",
"inverse-odds-ratio","direct","provided"));
otLogical <- yInSample!=0;
# If the data is not occurrence, let's assume that the parameter was switched unintentionally.
if(all(otLogical) & all(occurrence!=c("none","provided"))){
occurrence <- "none";
occurrenceModelProvided <- FALSE;
}
# If there were NAs and the occurrence was not specified, do something with it
# In all the other cases, NAs will be sorted out by the model
if(any(yNAValues) && (occurrence=="none")){
otLogical <- (!yNAValues)[1:obsInSample];
occurrence[] <- "provided";
pFitted[] <- otLogical*1;
pForecast[] <- 1;
occurrenceModel <- FALSE;
oesModel <- structure(list(y=matrix((otLogical)*1,ncol=1),fitted=pFitted,forecast=pForecast,
occurrence="provided"),class="occurrence");
}
else{
if(occurrence=="none"){
occurrenceModel <- FALSE;
otLogical <- rep(TRUE,obsInSample);
}
else if(occurrence=="provided"){
occurrenceModel <- TRUE;
oesModel$y <- matrix(otLogical*1,ncol=1);
}
else{
occurrenceModel <- TRUE;
}
# Include NAs in the zeroes, so that we avoid weird cases
if(any(yNAValues)){
otLogical <- !(!otLogical | yNAValues[1:obsInSample]);
}
}
if(any(yClasses=="ts")){
ot <- ts(matrix(otLogical*1,ncol=1), start=yStart, frequency=yFrequency);
}
else{
ot <- ts(matrix(otLogical*1,ncol=1), start=c(0,0), frequency=lagsModelMax);
}
obsNonzero <- sum(ot);
obsZero <- obsInSample - obsNonzero;
# If occurrence is provided, use it as is
if(occurrence=="provided"){
ot[] <- pFitted;
}
# Check if multiplicative models can be fitted
allowMultiplicative <- !((any(yInSample<=0) && !occurrenceModel) || (occurrenceModel && any(yInSample<0)));
if(etsModel){
# Clean the pool of models if only additive are allowed
if(!allowMultiplicative && !is.null(modelsPool)){
modelsPoolMultiplicative <- ((substr(modelsPool,1,1)=="M") |
substr(modelsPool,2,2)=="M" |
substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="M");
if(any(modelsPoolMultiplicative)){
modelsPool <- modelsPool[!modelsPoolMultiplicative];
# This is needed, because PPP and FFF use pool, not Branch and bound
if(!any(c(any(unlist(strsplit(model,""))=="P"),any(unlist(strsplit(model,""))=="F")))){
warning("Only additive models are allowed for your data. Amending the pool.",
call.=FALSE);
}
}
}
if((any(model==c("PPP","FFF","YYY")) || any(unlist(strsplit(model,""))=="Z")) && !allowMultiplicative){
model <- "XXX";
Etype[] <- "A";
Ttype[] <- switch(Ttype,"Y"=,"Z"=,"P"=,"F"="X",Ttype);
Stype[] <- switch(Stype,"Y"=,"Z"=,"P"=,"F"="X",Stype);
modelsPool <- NULL;
warning("Only additive models are allowed for your data. Changing the selection mechanism.",
call.=FALSE);
}
else if(!allowMultiplicative && any(c(Etype,Ttype,Stype)=="M")){
warning("Your data contains non-positive values, so the ETS(",model,") might break down.",
call.=FALSE);
}
else if(any(model==c("PPP","FFF")) && allowMultiplicative){
model <- "ZZZ";
}
}
#### Initial values ####
# Vectors for initials of different components
initialLevel <- NULL;
initialTrend <- NULL;
initialSeasonal <- NULL;
initialArima <- NULL;
initialXreg <- NULL;
# InitialEstimate vectors, defining what needs to be estimated
# NOTE: that initial==c("optimal","complete") means initialEstimate==TRUE!
initialEstimate <- initialLevelEstimate <- initialTrendEstimate <-
initialArimaEstimate <- initialXregEstimate <- TRUE;
# initials of seasonal is a vector, not a scalar, because we can have several lags
initialSeasonalEstimate <- rep(TRUE,componentsNumberETSSeasonal);
# This is an initialisation of the variable
initialType <- "optimal"
# initial type can be: "o" - optimal, "b" - backcasting, "p" - provided.
if(any(is.character(initial))){
initialType[] <- match.arg(initial, c("optimal","backcasting","complete"));
}
else if(is.null(initial)){
if(!silent){
message("Initial value is not selected. Switching to optimal.");
}
initialType[] <- "optimal";
}
else if(!is.null(initial)){
if(all(modelDo!=c("estimate","use"))){
warning(paste0("Predefined initials vector can only be used with preselected ETS model.\n",
"Changing to estimation of initials."),call.=FALSE);
initialType[] <- "optimal";
initialEstimate[] <- initialLevelEstimate[] <- initialTrendEstimate[] <-
initialSeasonalEstimate[] <- initialArimaEstimate[] <- initialXregEstimate[] <- TRUE;
}
else{
# If the list is provided, then check what this is.
# This should be: level, trend, seasonal[[1]], seasonal[[2]], ..., ARIMA, xreg
if(is.list(initial)){
# If this is a named list, then extract stuff using names
if(!is.null(names(initial))){
if(!is.null(initial$level)){
initialLevel <- initial$level;
}
if(!is.null(initial$trend)){
initialTrend <- initial$trend;
}
if(!is.null(initial$seasonal)){
initialSeasonal <- initial$seasonal;
}
if(!is.null(initial$arima)){
initialArima <- initial$arima;
}
if(!is.null(initial$xreg)){
initialXreg <- initial$xreg;
}
}
else{
if(!is.null(initial[[1]])){
initialLevel <- initial[[1]];
}
if(!is.null(initial[[2]])){
initialTrend <- initial[[2]];
}
if(!is.null(initial[[3]])){
initialSeasonal <- initial[[3]];
}
if(!is.null(initial[[4]])){
initialArima <- initial[[4]];
}
if(!is.null(initial[[5]])){
initialXreg <- initial[[5]];
}
}
}
else{
if(!is.numeric(initial)){
warning(paste0("Initial vector is not numeric!\n",
"Values of initial vector will be estimated."),call.=FALSE);
initialType[] <- "optimal";
}
else{
# If this is a vector, then it should contain values in the order:
# level, trend, seasonal1, seasonal2, ..., ARIMA, xreg
# if(length(initial)<(sum(lagsModelAll))){
# warning(paste0("The vector of initials contains only values for several components. ",
# "We will use what we can."),call.=FALSE);
# }
# else{
j <- 0;
if(etsModel){
initialLevel <- initial[1];
initialLevelEstimate[] <- FALSE;
if(modelIsTrendy){
j <- 2;
# If there is something in the vector, use it
if(all(!is.na(initial[j]))){
initialTrend <- initial[j];
initialTrendEstimate[] <- FALSE;
}
}
if(Stype!="N"){
# If there is something in the vector, use it
if(length(initial[-c(1:j)])>0){
initialSeasonal <- vector("list",componentsNumberETSSeasonal);
m <- 0;
for(i in 1:componentsNumberETSSeasonal){
if(all(!is.na(initial[j+m+1:lagsModelSeasonal[i]]))){
initialSeasonal[[i]] <- initial[j+m+1:lagsModelSeasonal[i]];
m <- m + lagsModelSeasonal[i];
}
else{
break;
}
}
j <- j+m;
initialSeasonalEstimate[] <- FALSE;
}
}
}
if(arimaModel){
# If there is something else left, this must be ARIMA
if(all(!is.na(initial[j+c(1:initialArimaNumber)]))){
initialArima <- initial[j+c(1:initialArimaNumber)];
j <- j+max(lagsModelARIMA);
initialArimaEstimate[] <- FALSE;
}
}
if(xregModel){
# Something else? xreg for sure!
if(length(initial[-c(1:j)])>0){
initialXreg <- initial[-c(1:j)];
initialXregEstimate[] <- FALSE
}
}
parametersNumber[2,1] <- parametersNumber[2,1] + j;
}
# }
}
}
}
#### Check the provided initials and define initialEstimate variables ####
if(etsModel){
# Level
if(!is.null(initialLevel)){
if(length(initialLevel)>1){
warning("Initial level contains more than one value! Using the first one.",
call.=FALSE);
initialLevel <- initialLevel[1];
}
initialLevelEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
# Trend
if(!is.null(initialTrend)){
if(length(initialTrend)>1){
warning("Initial trend contains more than one value! Using the first one.",
call.=FALSE);
initialTrend <- initialTrend[1];
}
initialTrendEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
# Seasonal
if(!is.null(initialSeasonal)){
# The list means several seasonal lags
if(is.list(initialSeasonal)){
# Is the number of seasonal initials correct? If it is bigger, then remove redundant
if(length(initialSeasonal)>componentsNumberETSSeasonal){
warning("Initial seasonals contained more elements than needed! Removing redundant ones.",
call.=FALSE);
initialSeasonal <- initialSeasonal[1:componentsNumberETSSeasonal];
}
# Is the number of initials in each season correct? Use the correct ones only
if(any(!(sapply(initialSeasonal,length) %in% lagsModelSeasonal))){
warning(paste0("Some of initial seasonals have a wrong length, ",
"not corresponding to the provided lags. I will estimate them."),
call.=FALSE);
initialSeasonalToUse <- sapply(initialSeasonal,length) %in% lagsModelSeasonal;
initialSeasonal <- initialSeasonal[initialSeasonalToUse];
}
initialSeasonalEstimate[] <- !(lagsModelSeasonal %in% sapply(initialSeasonal,length));
# If there are some gaps in what to estimate, reform initialSeason to make sense in the future creator function
if(!all(initialSeasonalEstimate) && !all(!initialSeasonalEstimate)){
initialSeasonalCorrect <- vector("list",componentsNumberETSSeasonal);
initialSeasonalCorrect[which(!initialSeasonalEstimate)] <- initialSeasonal;
initialSeasonal <- initialSeasonalCorrect;
}
}
# The vector implies only one seasonal
else{
if(all(length(initialSeasonal)!=lagsModelSeasonal)){
warning(paste0("Wrong length of seasonal initial: ",length(initialSeasonal),
"Instead of ",lagsModelSeasonal,". Switching to estimation."),
call.=FALSE)
initialSeasonalEstimate[] <- TRUE;
}
else{
initialSeasonalEstimate[] <- FALSE;
}
# Create a list from the vector for consistency purposes
initialSeasonal <- list(initialSeasonal);
}
parametersNumber[2,1] <- parametersNumber[2,1] + length(unlist(initialSeasonal));
}
}
else{
initialLevel <- initialTrend <- initialSeasonal <- NULL;
initialLevelEstimate <- initialTrendEstimate <- initialSeasonalEstimate <- FALSE
# If ETS is switched off, set error to whatever, based on the used distribution
Etype[] <- switch(distribution,
"dinvgauss"=,"dlnorm"=,"dllaplace"=,"dls"=,"dlgnorm"=,"dgamma"="M",
"A");
}
# ARIMA
if(!is.null(initialArima)){
if(length(initialArima)!=initialArimaNumber){
warning(paste0("The length of ARIMA initials is ",length(initialArima),
" instead of ",initialArimaNumber,". Estimating initials instead!"),
call.=FALSE);
initialArimaEstimate[] <- TRUE;
}
else{
initialArimaEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + length(initialArima);
}
}
# xreg
if(!is.null(initialXreg)){
initialXregEstimate[] <- FALSE;
}
#### Check ARIMA parameters, if they are provided ####
if(arimaModel){
# Check the provided parameters for AR and MA
if(!is.null(arma)){
arEstimate <- arRequired;
maEstimate <- maRequired;
# If this is a proper list, extract parameters separately
if(is.list(arma)){
armaParameters <- vector("numeric",sum(sapply(arma,length)));
j <- arIndex <- maIndex <- 0;
# The named list (proper thing)
if(!is.null(names(arma))){
# Check if the length of the provided parameters is correct
if(arRequired && !is.null(arma$ar) && length(arma$ar)!=sum(arOrders)){
warning(paste0("The number of provided AR parameters is ",length(arma$ar),
"while I need ",sum(arOrders),". ",
"Switching to estimation."),call.=FALSE);
arEstimate[] <- TRUE;
}
if(maRequired && !is.null(arma$ma) && length(arma$ma)!=sum(maOrders)){
warning(paste0("The number of provided MA parameters is ",length(arma$ma),
"while I need ",sum(maOrders),". ",
"Switching to estimation."),call.=FALSE);
maEstimate[] <- TRUE;
}
for(i in 1:length(lags)){
if(arRequired && !is.null(arma$ar) && arOrders[i]>0){
armaParameters[j+c(1:arOrders[i])] <- arma$ar[arIndex+c(1:arOrders[i])]
names(armaParameters)[j+c(1:arOrders[i])] <- paste0("phi",1:arOrders[i],"[",lags[i],"]");
j[] <- j+arOrders[i];
arIndex[] <- arIndex+arOrders[i];
arEstimate[] <- FALSE;
}
if(maRequired && !is.null(arma$ma) && maOrders[i]>0){
armaParameters[j+c(1:maOrders[i])] <- arma$ma[maIndex+c(1:maOrders[i])]
names(armaParameters)[j+c(1:maOrders[i])] <- paste0("theta",1:maOrders[i],"[",lags[i],"]");
j[] <- j+maOrders[i];
maIndex[] <- maIndex+maOrders[i];
maEstimate[] <- FALSE;
}
}
}
# Just a list. Assume that the first element is ar, and the second is ma
else{
for(i in 1:length(lags)){
k <- 1;
if(arRequired && !is.null(arma[[1]]) && arOrders[i]>0){
armaParameters[j+c(1:arOrders[i])] <- arma[[1]][arIndex+c(1:arOrders[i])]
names(armaParameters)[j+c(1:arOrders[i])] <- paste0("phi",1:arOrders[i],"[",lags[i],"]");
j[] <- j+arOrders[i];
arIndex[] <- arIndex+arOrders[i];
arEstimate[] <- FALSE;
k[] <- 2;
}
if(maRequired && !is.null(arma[[k]]) && maOrders[i]>0){
armaParameters[j+c(1:maOrders[i])] <- arma[[k]][maIndex+c(1:maOrders[i])]
names(armaParameters)[j+c(1:maOrders[i])] <- paste0("theta",1:maOrders[i],"[",lags[i],"]");
j[] <- j+maOrders[i];
maIndex[] <- maIndex+maOrders[i];
maEstimate[] <- FALSE;
}
}
# Check if the length of the provided parameters is correct
if(length(armaParameters)!=sum(arOrders)+sum(maOrders)){
warning(paste0("The number of provided ARMA parameters is ",length(armaParameters),
"while I need ",sum(arOrders)+sum(maOrders),". ",
"Switching to estimation."),call.=FALSE);
maEstimate <- arEstimate[] <- TRUE;
armaParameters <- NULL;
}
}
}
else if(is.vector(arma)){
# Check the length of the vector
if(length(arma)!=sum(arOrders)+sum(maOrders)){
warning(paste0("The number of provided ARMA parameters is ",length(arma),
"while I need ",sum(arOrders)+sum(maOrders),". ",
"Switching to estimation."),call.=FALSE);
maEstimate <- arEstimate[] <- TRUE;
armaParameters <- NULL;
}
else{
armaParameters <- arma;
j <- 0;
for(i in 1:length(lags)){
if(arOrders[i]>0){
names(armaParameters)[j+1:arOrders[i]] <- paste0("phi",1:arOrders[i],"[",lags[i],"]");
j <- j + arOrders[i];
}
if(maOrders[i]>0){
names(armaParameters)[j+1:maOrders[i]] <- paste0("theta",1:maOrders[i],"[",lags[i],"]");
j <- j + maOrders[i];
}
}
maEstimate <- arEstimate[] <- FALSE;
}
}
parametersNumber[2,1] <- parametersNumber[2,1] + length(armaParameters);
}
else{
arEstimate <- arRequired;
maEstimate <- maRequired;
armaParameters <- NULL;
}
}
else{
armaParameters <- NULL;
}
#### xreg preparation ####
# Check the regressors
if(!xregModel){
regressors[] <- "use";
formulaToUse <- NULL;
}
else{
if(regressors=="select"){
# If this has not happened by chance, then switch to optimisation
if(!is.null(initialXreg) && any(initialType==c("optimal","backcasting"))){
warning("Variables selection does not work with the provided initials for explantory variables. I will drop them.",
call.=FALSE);
initialXreg <- NULL;
initialXregEstimate <- TRUE;
}
if(!is.null(persistenceXreg) && any(persistenceXreg!=0)){
warning(paste0("I cannot do variables selection with the provided smoothing parameters ",
"for explantory variables. I will estimate them instead."),
call.=FALSE);
persistenceXreg <- NULL;
}
}
}
# Use alm() in order to fit the preliminary model for xreg
if(xregModel){
# List of initials. The first element is additive error, the second is the multiplicative one
xregModelInitials <- vector("list",2);
if(regressors!="select"){
#### Initial xreg are not provided ####
# If the initials are not provided, estimate them using ALM.
if(initialXregEstimate){
initialXregProvided <- FALSE;
# The function returns an ALM model
xregInitialiser <- function(Etype,distribution,formulaToUse,subset,responseName){
# Fix the default distribution for ALM
if(distribution=="default"){
distribution <- switch(Etype,
"A"="dnorm",
"M"="dlnorm");
}
else if(distribution=="dllaplace"){
distribution <- "dlaplace";
Etype <- "M";
}
else if(distribution=="dls"){
distribution <- "ds";
Etype <- "M";
}
else if(distribution=="dlgnorm"){
distribution <- "dgnorm";
Etype <- "M";
}
# This is needed to see if trend was asked explicitly. If not, we add it to get rid of bias
trendIncluded <- any(colnames(xregData)[-1]=="trend");
formulaIsAbsent <- is.null(formulaToUse);
formulaOriginal <- formulaToUse;
# If the formula is not provided, construct one
if(formulaIsAbsent){
if(Etype=="M" && any(distribution==c("dnorm","dlaplace","ds","dgnorm","dlogis","dt","dalaplace"))){
if(trendIncluded){
formulaOriginal <- formulaToUse <- as.formula(paste0("log(`",responseName,"`)~."));
}
else{
formulaOriginal <- as.formula(paste0("log(`",responseName,"`)~."));
formulaToUse <- as.formula(paste0("log(`",responseName,"`)~.+trend"));
}
}
else{
if(trendIncluded){
formulaOriginal <- formulaToUse <- as.formula(paste0("`",responseName,"`~."));
}
else{
formulaOriginal <- as.formula(paste0("`",responseName,"`~."));
formulaToUse <- as.formula(paste0("`",responseName,"`~.+trend"));
}
}
}
else{
# If formula only contains ".", then just change it
if(length(all.vars(formulaToUse))==2 && all.vars(formulaToUse)[2]=="."){
# Take logs if the model requires that
if((Etype=="M" && any(distribution==c("dnorm","dlaplace","ds","dgnorm","dlogis","dt","dalaplace")))){
if(trendIncluded){
formulaToUse <- as.formula(paste0("log(`",responseName,"`)~."));
}
else{
formulaToUse <- as.formula(paste0("log(`",responseName,"`)~.+trend"));
}
}
else{
if(trendIncluded){
formulaToUse <- as.formula(paste0(responseName,"~."));
}
else{
formulaToUse <- as.formula(paste0(responseName,"~.+trend"));
}
}
}
else{
trendIncluded <- any(all.vars(formulaToUse)[-1]=="trend");
# If formula contains only one element, or several, but no logs, then change response formula
if((length(formulaToUse[[2]])==1 ||
(length(formulaToUse[[2]])>1 & !any(as.character(formulaToUse[[2]])=="log"))) &&
(Etype=="M" && any(distribution==c("dnorm","dlaplace","ds","dgnorm","dlogis","dt","dalaplace")))){
if(trendIncluded){
formulaToUse <- update(formulaToUse,log(.)~.);
}
else{
formulaToUse <- update(formulaToUse,log(.)~.+trend);
}
}
else{
if(!trendIncluded){
formulaToUse <- update(formulaToUse,.~.+trend);
}
}
}
}
almModel <- do.call(alm,list(formula=formulaToUse,data=xregData,distribution=distribution,subset=which(subset)));
# Remove trend
if(!trendIncluded){
almModel$coefficients <- almModel$coefficients[names(almModel$coefficients)!="trend"];
almModel$data <- almModel$data[,colnames(almModel$data)!="trend",drop=FALSE];
if(formulaIsAbsent){
almModel$call$formula <- as.formula(paste0("`",responseName,"`~."));
}
else{
almModel$call$formula <- formulaOriginal;
}
# Reset terms, they are not needed for what comes, but confuse the formula() method
almModel$terms <- NULL;
}
return(almModel);
}
# Extract names of variables, fix the formula
if(!is.null(formulaToUse)){
formulaToUse <- as.formula(formulaToUse);
responseName <- all.vars(formulaToUse)[1];
# If there are spaces in names, give a warning
if(any(grepl("[^A-Za-z0-9,;._-]", all.vars(formulaToUse))) ||
# If the names only contain numbers
any(suppressWarnings(!is.na(as.numeric(all.vars(formulaToUse)))))){
warning("The names of your variables contain special characters ",
"(such as numbers, spaces, comas, brackets etc). adam() might not work properly. ",
"It is recommended to use `make.names()` function to fix the names of variables.",
call.=FALSE);
formulaToUse <- as.formula(paste0(gsub(paste0("`",all.vars(formulaToUse)[1],"`"),
make.names(all.vars(formulaToUse)[1]),
all.vars(formulaToUse)[1]),"~",
paste0(mapply(gsub, paste0("`",all.vars(formulaToUse)[-1],"`"),
make.names(all.vars(formulaToUse)[-1]),
labels(terms(formulaToUse))),
collapse="+")));
}
formulaProvided <- TRUE;
}
else{
formulaToUse <- reformulate(make.names(colnames(xregData)[-1]), response=responseName);
# Quotes are needed here for the insecure names of variables, such as "1", "2", "3" etc
# formulaToUse <- reformulate(setdiff(paste0("`",colnames(xregData),"`"),
# paste0("`",responseName,"`")),
# response=responseName);
formulaProvided <- FALSE;
}
# Robustify the names of variables
colnames(xregData) <- make.names(colnames(xregData));
# This allows to save the original data
xreg <- xregData;
obsXreg <- nrow(xregData);
# Form subset in order to use in-sample only
subset <- rep(FALSE, obsAll);
subset[1:obsInSample] <- TRUE;
# Exclude zeroes if this is an occurrence model
if(occurrenceModel){
subset[1:obsInSample][!otLogical] <- FALSE;
}
#### Pure regression ####
#### If this is just a regression ALM
if((!etsModel && !arimaModel) && regressors!="adapt"){
# Return the estimated model based on the provided xreg
if(is.null(formulaToUse)){
formulaToUse <- reformulate(setdiff(colnames(xregData), responseName), response=responseName);
# formulaToUse <- as.formula(paste0("`",responseName,"`~."));
formulaProvided <- FALSE;
}
else{
formulaProvided <- TRUE;
}
if(distribution=="default"){
distribution[] <- "dnorm";
}
# Redefine loss for ALM
loss <- switch(loss,
"MSEh"=,"TMSE"=,"GTMSE"=,"MSCE"="MSE",
"MAEh"=,"TMAE"=,"GTMAE"=,"MACE"="MAE",
"HAMh"=,"THAM"=,"GTHAM"=,"CHAM"="HAM",
loss);
lossOriginal <- loss;
if(loss=="custom"){
loss <- lossFunction;
}
# Either use or select the model via greybox functions
# Fisher Information
if(is.null(ellipsis$FI)){
FI <- FALSE;
}
else{
FI <- ellipsis$FI;
}
almModel <- do.call("alm", list(formula=formulaToUse, data=xregData,
distribution=distribution, loss=loss,
subset=which(subset),
occurrence=oesModel,FI=FI));
almModel$call$data <- as.name(yName);
return(almModel);
}
#### ETSX / ARIMAX ####
almModel <- NULL;
if(Etype!="Z"){
almModel <- xregInitialiser(Etype,distribution,formulaToUse,subset,responseName);
# If Intercept was not included, substitute with zero
if(!any(names(almModel$coefficients)=="(Intercept)")){
almIntercept <- 0;
}
else{
almIntercept <- almModel$coefficients["(Intercept)"];
}
if(Etype=="A"){
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregModelInitials[[1]]$initialXreg <- almModel$coefficients;
}
else{
xregModelInitials[[1]]$initialXreg <- almModel$coefficients[-1];
}
if(is.null(formulaToUse)){
xregModelInitials[[1]]$formula <- formulaToUse <- formula(almModel);
}
xregModelInitials[[1]]$other <- almModel$other;
}
else{
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregModelInitials[[2]]$initialXreg <- almModel$coefficients;
}
else{
xregModelInitials[[2]]$initialXreg <- almModel$coefficients[-1];
}
if(is.null(formulaToUse)){
xregModelInitials[[2]]$formula <- formulaToUse <- formula(almModel);
}
xregModelInitials[[2]]$other <- almModel$other;
}
}
# If we are selecting the appropriate error, produce two models: for "M" and for "A"
else{
# Additive model
almModel <- xregInitialiser("A",distribution,formulaToUse,subset,responseName);
# If Intercept was not included, substitute with zero
if(!any(names(almModel$coefficients)=="(Intercept)")){
almIntercept <- 0;
}
else{
almIntercept <- almModel$coefficients["(Intercept)"];
}
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregModelInitials[[1]]$initialXreg <- almModel$coefficients;
}
else{
xregModelInitials[[1]]$initialXreg <- almModel$coefficients[-1];
}
if(is.null(formulaToUse)){
xregModelInitials[[1]]$formula <- formula(almModel);
}
xregModelInitials[[1]]$other <- almModel$other;
# Multiplicative model
almModel[] <- xregInitialiser("M",distribution,formulaToUse,subset,responseName);
# If Intercept was not included, substitute with zero
if(!any(names(almModel$coefficients)=="(Intercept)")){
almIntercept <- 0;
}
else{
almIntercept <- almModel$coefficients["(Intercept)"];
}
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregModelInitials[[2]]$initialXreg <- almModel$coefficients;
}
else{
xregModelInitials[[2]]$initialXreg <- almModel$coefficients[-1];
}
if(is.null(formulaToUse)){
xregModelInitials[[2]]$formula <- formula(almModel);
}
xregModelInitials[[2]]$other <- almModel$other;
}
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregNumber <- ncol(almModel$data);
xregNames <- names(coef(almModel));
}
else{
# Write down the number and names of parameters
xregNumber <- ncol(almModel$data)-1;
xregNames <- names(coef(almModel))[-1];
}
# The original number of obs in xreg
obsXreg <- nrow(xreg);
#### Data manipulations for further use ####
# This formula is needed in order to expand the data
if(is.null(formulaToUse)){
formulaToUse <- as.formula(paste0("`",responseName,"`~",
paste0(colnames(xreg)[colnames(xreg)!=responseName],
collapse="+")));
}
# Remove variables without variability
noVariability <- setNames(vector("logical",ncol(xreg)),colnames(xreg));
noVariability[] <- apply((as.matrix(xreg[1:obsInSample,])==matrix(xreg[1,],obsInSample,ncol(xreg),byrow=TRUE)),2,all);
noVariabilityNames <- names(noVariability)[noVariability];
if(any(noVariability) && any(all.vars(formulaToUse) %in% names(noVariability))){
formulaToUse <- update.formula(formulaToUse,paste0(".~.-",paste0(noVariabilityNames,collapse="-")));
}
# Robustify the names of variables
colnames(xreg) <- make.names(colnames(xreg),unique=TRUE);
# The names of the original variables
xregNamesOriginal <- colnames(xregData)[-1];
if((is.matrix(xreg) && any(apply(xreg,2,is.character))) ||
(!is.matrix(xreg) && any(sapply(xreg,is.character)))){
warning("You have character variables in your data. ",
"I will treat them as factors, but it is advised to convert them to factors manually",
call.=FALSE);
}
# If xreg is data frame or formula is provided, do model.frame
additionalManipulations <- (is.data.frame(xreg) || formulaProvided);
if(additionalManipulations){
# Expand the variables. We cannot use alm, because it is based on obsInSample
xregData <- model.frame(formulaToUse,data=as.data.frame(xreg));
}
else{
xregData <- as.matrix(xreg)[,c(names(noVariability)[!noVariability]),drop=FALSE][,-1,drop=FALSE];
xregNumber <- ncol(xregData);
xregNames <- colnames(xregData);
}
# If the number of rows is different, this might be because of NAs
if(additionalManipulations && nrow(xregData)!=nrow(xreg)){
warning("Some variables contained NAs. This might cause issues in the estimation. ",
"I will substitute those values with the first non-NA values",
call.=FALSE);
# Get indices of NAs and nonNAs
xregNAs <- which(is.na(xreg),arr.ind=TRUE);
xregNonNAs <- which(!is.na(xreg),arr.ind=TRUE);
# Go through variables and substitute values
for(i in unique(xregNAs[,2])){
# This split on [row, column] is needed, because data.table is funny with cbind(row,column)
xreg[xregNAs[xregNAs[,2]==i,]] <- xreg[xregNonNAs[which(xregNonNAs[,2]==i)[1],1],
xregNonNAs[which(xregNonNAs[,2]==i)[1],2]];
}
xregData <- model.frame(formulaToUse,data=as.data.frame(xreg));
}
# Get the response variable, just in case it was transformed
if(additionalManipulations && length(formulaToUse[[2]])>1){
y <- xregData[,1];
yInSample <- matrix(y[1:obsInSample],ncol=1);
if(holdout){
yHoldout <- matrix(y[-c(1:obsInSample)],ncol=1);
}
}
#### Drop the variables with no variability and perfectly correlated
if(additionalManipulations){
xregExpanded <- colnames(model.matrix(xregData,data=xregData));
# Remove intercept
if(any(xregExpanded=="(Intercept)")){
xregExpanded <- xregExpanded[-1];
}
if(any(!(xregExpanded %in% xregNames))){
xregNamesRetained <- rep(TRUE,length(xregNamesOriginal));
for(i in 1:length(xregNamesOriginal)){
xregNamesRetained[i] <- any(grepl(xregNamesOriginal[i], xregNames));
}
# If the dropped variables are in the formula, update the formula
if(length(all.vars(formulaToUse))>2 &&
any(all.vars(formulaToUse) %in% xregNamesOriginal[!xregNamesRetained])){
formulaToUse <- update.formula(formulaToUse,
paste0(".~.-",
paste(xregNamesOriginal[!xregNamesRetained],
collapse="-")));
}
xregNamesOriginal <- c(responseName,xregNamesOriginal[xregNamesRetained]);
xregData <- model.frame(formulaToUse,data=as.data.frame(xreg[,xregNamesOriginal,drop=FALSE]));
# Remove response variable
xregNamesOriginal <- xregNamesOriginal[-1]
}
# Binary, flagging factors in the data
xregFactors <- (attr(terms(xregData),"dataClasses")=="factor")[-1];
# Expanded stuff with all levels for factors
if(any(xregFactors)){
# Levels for the factors
xregFactorsLevels <- lapply(xreg,levels);
xregFactorsLevels[[responseName]] <- NULL;
xregModelMatrix <- model.matrix(xregData,xregData,
contrasts.arg=lapply(xregData[attr(terms(xregData),"dataClasses")=="factor"],
contrasts, contrasts=FALSE));
xregNamesModified <- colnames(xregModelMatrix)[-1];
}
else{
xregModelMatrix <- model.matrix(xregData,data=xregData);
xregNamesModified <- xregNames;
}
# Drop the unused variables
xregData <- as.matrix(xregModelMatrix);
# Remove intercept
interceptIsPresent <- FALSE;
if(any(colnames(xregData)=="(Intercept)")){
interceptIsPresent[] <- TRUE;
xregData <- xregData[,-1,drop=FALSE];
}
xregNumber <- ncol(xregData);
# If there are more obs in xreg than the obsAll cut the thing
if(obsXreg>=obsAll){
xregData <- xregData[1:obsAll,,drop=FALSE]
}
# If there are less xreg observations than obsAll, use Naive
else{
newnRows <- obsAll-obsXreg;
xregData <- rbind(xregData,matrix(rep(tail(xregData,1),each=newnRows),newnRows,xregNumber));
}
#### Fix parameters for factors ####
# The indices of the original parameters
xregParametersMissing <- setNames(vector("numeric",xregNumber),xregNamesModified);
# # The indices of the original parameters
xregParametersIncluded <- setNames(vector("numeric",xregNumber),xregNamesModified);
# The vector, marking the same values of smoothing parameters
if(interceptIsPresent){
xregParametersPersistence <- setNames(attr(xregModelMatrix,"assign")[-1],xregNamesModified);
}
else{
xregParametersPersistence <- setNames(attr(xregModelMatrix,"assign"),xregNamesModified);
}
if(length(xregParametersPersistence)==0){
xregParametersPersistence <- 0;
}
# If there are factors not in the alm data, create additional initials
if(any(xregFactors) && any(!(xregNamesModified %in% xregNames))){
xregAbsent <- !(xregNamesModified %in% xregNames);
xregParametersNew <- setNames(rep(NA,xregNumber),xregNamesModified);
# If there is stuff for additive error model, fix parameters
if(!is.null(xregModelInitials[[1]])){
xregParametersNew[!xregAbsent] <- xregModelInitials[[1]]$initialXreg;
# Go through new names and find, where they came from. Then get the missing parameters
for(i in which(xregAbsent)){
# Find the name of the original variable
# Use only the last value... hoping that the names like x and x1 are not used.
xregNameFoundID <- sapply(xregNamesOriginal,grepl,xregNamesModified[i]);
xregNameFound <- tail(names(xregNameFoundID)[xregNameFoundID],1);
# Get the indices of all k-1 levels
xregParametersIncluded[xregNames[xregNames %in% paste0(xregNameFound,
xregFactorsLevels[[xregNameFound]])]] <- i;
# Get the index of the absent one
xregParametersMissing[i] <- i;
# Fill in the absent one, add intercept
xregParametersNew[i] <- almIntercept;
xregParametersNew[xregNamesModified[xregParametersIncluded==i]] <- almIntercept +
xregParametersNew[xregNamesModified[xregParametersIncluded==i]];
# normalise all of them
xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]] <-
xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]] -
mean(xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]]);
}
# Write down the new parameters
xregModelInitials[[1]]$initialXreg <- xregParametersNew;
}
# If there is stuff for multiplicative error model, fix parameters
if(!is.null(xregModelInitials[[2]])){
xregParametersNew[!xregAbsent] <- xregModelInitials[[2]]$initialXreg;
# Go through new names and find, where they came from. Then get the missing parameters
for(i in which(xregAbsent)){
# Find the name of the original variable
# Use only the last value... hoping that the names like x and x1 are not used.
xregNameFoundID <- sapply(xregNamesOriginal,grepl,xregNamesModified[i]);
xregNameFound <- tail(names(xregNameFoundID)[xregNameFoundID],1);
# Get the indices of all k-1 levels
xregParametersIncluded[xregNames[xregNames %in% paste0(xregNameFound,
xregFactorsLevels[[xregNameFound]])]] <- i;
# Get the index of the absent one
xregParametersMissing[i] <- i;
# Fill in the absent one, add intercept
xregParametersNew[i] <- almIntercept;
xregParametersNew[xregNamesModified[xregParametersIncluded==i]] <- almIntercept +
xregParametersNew[xregNamesModified[xregParametersIncluded==i]];
# normalise all of them
xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]] <-
xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]] -
mean(xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]]);
}
# Write down the new parameters
xregModelInitials[[2]]$initialXreg <- xregParametersNew;
}
xregNames <- xregNamesModified;
}
# The vector of parameters that should be estimated (numeric + original levels of factors)
xregParametersEstimated <- xregParametersIncluded
xregParametersEstimated[xregParametersEstimated!=0] <- 1;
xregParametersEstimated[xregParametersMissing==0 & xregParametersIncluded==0] <- 1;
}
else{
# If there are more obs in xreg than the obsAll cut the thing
if(obsXreg>=obsAll){
xregData <- xregData[1:obsAll,,drop=FALSE]
}
# If there are less xreg observations than obsAll, use Naive
else{
newnRows <- obsAll-obsXreg;
xregData <- rbind(xregData,matrix(rep(tail(xregData,1),each=newnRows),newnRows,xregNumber));
}
xregFactors <- FALSE;
xregParametersPersistence <- setNames(c(1:xregNumber),xregNames);
xregParametersEstimated <- setNames(rep(1,xregNumber),xregNames);
xregParametersMissing <- setNames(c(1:xregNumber),xregNames);
xregParametersIncluded <- setNames(c(1:xregNumber),xregNames);
}
# Remove xreg, just to preserve some memory
rm(xreg);
# Clean memory if have a big object
if(obsInSample>10000){
gc(verbose=FALSE);
}
}
#### Initial xreg are provided ####
else{
#### Pure regression ####
#### If this is just a regression, then this must be a reuse of alm.
if((!etsModel && !arimaModel) && regressors!="adapt"){
# Return the estimated model based on the provided xreg
if(is.null(formulaToUse)){
formulaToUse <- as.formula(paste0("`",responseName,"`~."));
}
if(distribution=="default"){
distribution[] <- "dnorm";
}
# Redefine loss for ALM
loss <- switch(loss,
"MSEh"=,"TMSE"=,"GTMSE"=,"MSCE"="MSE",
"MAEh"=,"TMAE"=,"GTMAE"=,"MACE"="MAE",
"HAMh"=,"THAM"=,"GTHAM"=,"CHAM"="HAM",
loss);
lossOriginal <- loss;
if(loss=="custom"){
loss <- lossFunction;
}
# Form subset in order to use in-sample only
subset <- rep(FALSE, obsAll);
subset[1:obsInSample] <- TRUE;
# Exclude zeroes if this is an occurrence model
if(occurrenceModel){
subset[1:obsInSample][!otLogical] <- FALSE;
}
# Fisher Information
if(is.null(ellipsis$FI)){
FI <- FALSE;
}
else{
FI <- ellipsis$FI;
}
if(is.null(ellipsis$B)){
B <- initialXreg;
}
else{
B <- ellipsis$B;
}
almModel <- do.call("alm", list(formula=formulaToUse, data=xregData,
distribution=distribution, loss=loss, subset=which(subset),
occurrence=occurrence,FI=FI,
parameters=B,fast=TRUE));
almModel$call$data <- as.name(yName);
return(almModel);
}
#### InitialXreg is provided ####
else{
# Save the original data
xreg <- xregData;
initialXregProvided <- TRUE;
# Robustify the names of variables
colnames(xreg) <- make.names(colnames(xreg),unique=TRUE);
# This formula is needed only to expand xreg
if(is.null(formulaToUse)){
responseName <- colnames(xreg)[1]
formulaToUse <- as.formula(paste0("`",responseName,"`~",
paste0(colnames(xreg)[colnames(xreg)!=responseName],
collapse="+")));
}
# Extract names and form a proper matrix for the regression
else{
responseName <- all.vars(formulaToUse)[1];
}
# Get the names of initials
xregNames <- names(initialXreg);
# The names of the original variables
xregNamesOriginal <- colnames(xregData)[-1];
# Expand the variables. We cannot use alm, because it is based on obsInSample
xregData <- model.frame(formulaToUse,data=as.data.frame(xreg));
# Get the response variable, just in case it was transformed
if(length(formulaToUse[[2]])>1){
y <- xregData[,1];
yInSample <- matrix(y[1:obsInSample],ncol=1);
if(holdout){
yHoldout <- matrix(y[-c(1:obsInSample)],ncol=1);
}
}
# Binary, flagging factors in the data
xregFactors <- (attr(terms(xregData),"dataClasses")=="factor")[-1];
# Expanded stuff with all levels for factors
if(any(xregFactors)){
# Levels for the factors
xregFactorsLevels <- lapply(xreg[,-1,drop=FALSE],levels);
xregModelMatrix <- model.matrix(xregData,xregData,
contrasts.arg=lapply(xregData[attr(terms(xregData),"dataClasses")=="factor"],
contrasts, contrasts=FALSE));
xregNamesModified <- colnames(xregModelMatrix)[-1];
}
else{
xregModelMatrix <- model.matrix(xregData,data=xregData);
xregNamesModified <- xregNames;
}
xregData <- as.matrix(xregModelMatrix);
# Remove intercept
interceptIsPresent <- FALSE;
if(any(colnames(xregData)=="(Intercept)")){
interceptIsPresent[] <- TRUE;
xregData <- xregData[,-1,drop=FALSE];
}
xregNumber <- ncol(xregData);
obsXreg <- nrow(xregData);
# If there are more obs in xreg than the obsAll cut the thing
if(obsXreg>=obsAll){
xregData <- xregData[1:obsAll,,drop=FALSE]
}
# If there are less xreg observations than obsAll, use Naive
else{
newnRows <- obsAll-obsXreg;
xregData <- rbind(xregData,matrix(rep(tail(xregData,1),each=newnRows),newnRows,xregNumber));
}
# This variable is needed in order to do model.matrix only, when required.
xregDataIsDataFrame <- is.data.frame(xregData);
if(xregDataIsDataFrame){
# Expand xregData if it is a data frame
xregData <- model.frame(formulaToUse,data=xregData);
}
xregNumber[] <- ncol(xregData);
#### Fix parameters for factors ####
# The indices of the original parameters
xregParametersMissing <- setNames(vector("numeric",xregNumber),xregNamesModified);
# # The indices of the original parameters
xregParametersIncluded <- setNames(vector("numeric",xregNumber),xregNamesModified);
# The vector, marking the same values of smoothing parameters
if(interceptIsPresent){
xregParametersPersistence <- setNames(attr(xregModelMatrix,"assign")[-1],xregNamesModified);
}
else{
xregParametersPersistence <- setNames(attr(xregModelMatrix,"assign"),xregNamesModified);
}
if(length(xregParametersPersistence)==0){
xregParametersPersistence <- 0;
}
# If there are factors and the number of initials is not the same as the number of parameters needed
# This stuff assumes that the provided xreg parameters are named.
if(any(xregFactors)){
# Expand the data again in order to find the missing elements
xregNames <- colnames(as.matrix(model.matrix(formulaToUse,
model.frame(formulaToUse,data=as.data.frame(xreg)))));
if(any(xregNames=="(Intercept)")){
xregNames <- xregNames[xregNames!="(Intercept)"];
}
if(length(xregNames)!=length(xregNamesModified)){
xregAbsent <- !(xregNamesModified %in% xregNames);
# Create the new vector of parameters
xregParametersNew <- setNames(rep(NA,xregNumber),xregNamesModified);
if(!is.null(names(initialXreg))){
xregParametersNew[xregNames] <- initialXreg[xregNames];
}
else{
# Insert values one by one
j <- 1;
for(i in 1:length(xregNamesModified)){
if(!xregAbsent[i]){
xregParametersNew[i] <- initialXreg[j];
j <- j+1;
}
}
}
# Go through new names and find, where they came from. Then get the missing parameters
for(i in which(xregAbsent)){
# Find the name of the original variable
# Use only the last value... hoping that the names like x and x1 are not used.
xregNameFound <- tail(names(unlist(sapply(xregNamesOriginal,grep,xregNamesModified[i]))),1);
# Get the indices of all k-1 levels
xregParametersIncluded[xregNames[xregNames %in% paste0(xregNameFound,
xregFactorsLevels[[xregNameFound]])]] <- i;
# Get the index of the absent one
xregParametersMissing[i] <- i;
# Fill in the absent one
xregParametersNew[i] <- -sum(xregParametersNew[xregParametersIncluded==i],
na.rm=TRUE);
}
# Write down the new parameters
xregModelInitials[[1]]$initialXreg <- xregParametersNew;
# Write down the new parameters to the second part
if(any(Etype==c("M","Z"))){
xregModelInitials[[2]]$initialXreg <- xregParametersNew;
}
xregNames <- xregNamesModified;
}
}
else{
xregModelInitials[[1]]$initialXreg <- initialXreg;
if(any(Etype==c("M","Z"))){
xregModelInitials[[2]]$initialXreg <- initialXreg;
}
}
# The vector of parameters that should be estimated (numeric + original levels of factors)
xregParametersEstimated <- xregParametersIncluded
xregParametersEstimated[xregParametersEstimated!=0] <- 1;
xregParametersEstimated[xregParametersMissing==0 & xregParametersIncluded==0] <- 1;
parametersNumber[2,2] <- parametersNumber[2,2] + xregNumber;
}
# Remove xreg, just to preserve some memory
rm(xreg);
# Clean memory if have a big object
if(obsInSample>10000){
gc(verbose=FALSE);
}
}
}
else{
#### Pure regression ####
#### If this is just a regression, use stepwise
if((!etsModel && !arimaModel) && regressors!="adapt"){
# Return the estimated model based on the provided xreg
if(is.null(formulaToUse)){
formulaToUse <- reformulate(setdiff(colnames(xregData), responseName), response=responseName);
# formulaToUse <- as.formula(paste0("`",responseName,"`~."));
formulaProvided <- FALSE;
}
else{
formulaProvided <- TRUE;
}
if(distribution=="default"){
distribution[] <- "dnorm";
}
# Form subset in order to use in-sample only
subset <- rep(FALSE, obsAll);
subset[1:obsInSample] <- TRUE;
# Exclude zeroes if this is an occurrence model
if(occurrenceModel){
subset[1:obsInSample][!otLogical] <- FALSE;
}
almModel <- do.call("stepwise", list(data=xregData, formula=formulaToUse, subset=subset,
distribution=distribution, occurrence=oesModel));
almModel$call$data <- as.name(yName);
return(almModel);
}
# Include only variables from the formula
if(is.null(formulaToUse)){
formulaToUse <- as.formula(paste0("`",responseName,"`~."));
}
else{
# Do model.frame manipulations
# We do it this way to avoid factors expansion into dummies at this stage
mf <- as.call(list(quote(stats::model.frame), formula=formulaToUse,
data=xregData, drop.unused.levels=FALSE));
if(!is.data.frame(xregData)){
mf$data <- as.data.frame(xregData);
}
# Evaluate data frame to do transformations of variables
xregData <- eval(mf, parent.frame());
# Remove variables that have "-x" in the formula
dataTerms <- terms(xregData);
xregData <- xregData[,colnames(attr(dataTerms,"factors"))];
}
xregNumber <- ncol(xregData);
xregNames <- colnames(xregData);
initialXregProvided <- FALSE;
xregFactors <- FALSE;
xregParametersPersistence <- setNames(c(1:xregNumber),xregNames);
xregParametersEstimated <- setNames(rep(1,xregNumber),xregNames);
xregParametersMissing <- setNames(c(1:xregNumber),xregNames);
xregParametersIncluded <- setNames(c(1:xregNumber),xregNames);
}
#### persistence for xreg ####
# Process the persistence for xreg
if(!is.null(persistenceXreg)){
if(length(persistenceXreg)!=xregNumber && length(persistenceXreg)!=1){
warning("The length of the provided persistence for the xreg variables is wrong. Reverting to the estimation.",
call.=FALSE);
persistenceXreg <- rep(0.5,xregNumber);
persistenceXregProvided <- FALSE;
persistenceXregEstimate <- TRUE;
}
else if(length(persistenceXreg)==1){
persistenceXreg <- rep(persistenceXreg,xregNumber);
persistenceXregProvided <- TRUE;
persistenceXregEstimate <- FALSE;
}
else{
persistenceXregProvided <- TRUE;
persistenceXregEstimate <- FALSE;
}
}
else{
# The persistenceXregEstimate is provided
if(regressors=="adapt" && !persistenceXregEstimate){
persistenceXregProvided <- TRUE;
persistenceXregEstimate <- FALSE;
}
else if(regressors=="adapt" && persistenceXregEstimate){
persistenceXreg <- rep(0.05,xregNumber);
persistenceXregProvided <- FALSE;
persistenceXregEstimate <- TRUE;
}
else{
persistenceXreg <- rep(0,xregNumber);
persistenceXregProvided <- FALSE;
persistenceXregEstimate <- FALSE;
}
}
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
lagsModelAll <- matrix(rep(1,xregNumber),ncol=1);
}
else{
lagsModelAll <- matrix(c(lagsModelAll,rep(1,xregNumber)),ncol=1);
}
# If there's only one explanatory variable, then there's nothing to select
if(xregNumber==1){
regressors[] <- "use";
}
# Fix the names of variables
colnames(xregData) <- make.names(colnames(xregData), unique=TRUE);
xregNames[] <- make.names(xregNames, unique=TRUE);
# If there are no variables after all of that, then xreg doesn't exist
if(xregNumber==0){
warning("It looks like there are no suitable explanatory variables. Check the xreg! I dropped them out.",
call.=FALSE);
xregModel[] <- FALSE;
xregData <- NULL;
}
}
else{
initialXregProvided <- FALSE;
initialXregEstimate <- FALSE;
persistenceXregProvided <- FALSE;
persistenceXregEstimate <- FALSE;
xregModelInitials <- NULL;
xregData <- NULL;
xregNumber <- 0;
xregNames <- NULL;
if(is.null(formulaToUse)){
formulaToUse <- as.formula(paste0("`",responseName,"`~."));
}
xregParametersMissing <- 0;
xregParametersIncluded <- 0;
xregParametersEstimated <- 0;
xregParametersPersistence <- 0;
}
# Fix the occurrenceModel for "provided"
if(occurrence=="provided"){
occurrenceModel <- FALSE;
}
# Redefine persitenceEstimate value
persistenceEstimate[] <- any(c(persistenceLevelEstimate,persistenceTrendEstimate,
persistenceSeasonalEstimate,persistenceXregEstimate));
#### Conclusions about the initials ####
# Make sure that only important elements are estimated.
if(!modelIsTrendy){
initialTrendEstimate <- FALSE;
initialTrend <- NULL;
}
if(!modelIsSeasonal){
initialSeasonalEstimate <- FALSE;
initialSeasonal <- NULL;
}
if(!arimaModel){
initialArimaEstimate <- FALSE;
initialArima <- NULL;
}
if(!xregModel){
initialXregEstimate <- FALSE;
initialXreg <- NULL;
}
# If we don't need to estimate anything, flag initialEstimate
if(!any(c(etsModel && initialLevelEstimate,
(etsModel && modelIsTrendy && initialTrendEstimate),
(etsModel & modelIsSeasonal & initialSeasonalEstimate),
(arimaModel && initialArimaEstimate),
(xregModel && initialXregEstimate)))){
initialEstimate[] <- FALSE;
}
else{
initialEstimate[] <- TRUE;
}
# If at least something is provided, flag it as "provided"
if((etsModel && !initialLevelEstimate) ||
(etsModel && modelIsTrendy && !initialTrendEstimate) ||
any(etsModel & modelIsSeasonal & !initialSeasonalEstimate) ||
(arimaModel && !initialArimaEstimate) ||
(xregModel && !initialXregEstimate)){
initialType[] <- "provided";
}
# Observations in the states matrix
# Define the number of cols that should be in the matvt
obsStates <- obsInSample + lagsModelMax;
if(any(yInSample<=0) && any(distribution==c("dinvgauss","dgamma","dlnorm","dllaplace","dls","dlgnorm")) &&
!occurrenceModel && (occurrence!="provided")){
warning(paste0("You have non-positive values in the data. ",
"The distribution ",distribution," does not support that. ",
"This might lead to problems in the estimation."),
call.=FALSE);
}
# Update the number of parameters
if(occurrenceModelProvided){
parametersNumber[2,3] <- nparam(oesModel);
pForecast <- c(forecast(oesModel, h=h, interval="none")$mean);
}
#### Information Criteria ####
ic <- match.arg(ic,c("AICc","AIC","BIC","BICc"));
icFunction <- switch(ic,
"AIC"=AIC,
"AICc"=AICc,
"BIC"=BIC,
"BICc"=BICc);
#### Bounds for the smoothing parameters ####
bounds <- match.arg(bounds,c("usual","admissible","none"));
#### Checks for the potential number of degrees of freedom ####
# This is needed in order to make the function work on small samples
# scale parameter, smoothing parameters and phi
nParamMax <- (1 +
# ETS model
etsModel*(persistenceLevelEstimate + modelIsTrendy*persistenceTrendEstimate +
modelIsSeasonal*sum(persistenceSeasonalEstimate) +
phiEstimate + (initialType=="optimal") *
(initialLevelEstimate + initialTrendEstimate + sum(initialSeasonalEstimate*lagsModelSeasonal))) +
# ARIMA components: initials + parameters
arimaModel*((initialType=="optimal")*initialArimaNumber +
arRequired*arEstimate*sum(arOrders) + maRequired*maEstimate*sum(maOrders)) +
# Xreg initials and smoothing parameters
xregModel*(xregNumber*(any(initialType==c("optimal","backcasting"))*initialXregEstimate+persistenceXregEstimate)));
# If the sample is smaller than the number of parameters
if(obsNonzero <= nParamMax){
# If there is both ETS and ARIMA, remove ARIMA
if(etsModel && arimaModel && !select){
warning("I don't have enough observations to fit ETS with ARIMA terms. I will construct the simple ETS.",
call.=FALSE);
lagsModelAll <- lagsModelAll[-c(componentsNumberETS+c(1:componentsNumberARIMA)),,drop=FALSE];
arRequired <- iRequired <- maRequired <- arimaModel <- FALSE;
arOrders <- iOrders <- maOrders <- NULL;
nonZeroARI <- nonZeroMA <- lagsModelARIMA <- NULL;
componentsNamesARIMA <- NULL;
initialArimaNumber <- componentsNumberARIMA <- 0;
lagsModelMax <- max(lagsModelAll);
}
else if(arimaModel && !etsModel && !select){
# If the backacasting helps, switch to it.
if(initialType=="optimal" && (obsNonzero > (nParamMax - (initialType=="optimal")*initialArimaNumber))){
warning(paste0("The number of parameter to estimate is ",nParamMax,
", while the number of observations is ",obsNonzero,
". Switching initial to 'backcasting' to save some degrees of freedom."), call.=FALSE);
initialType <- "complete";
}
else{
warning(paste0("The number of parameter to estimate is ",nParamMax,
", while the number of observations is ",obsNonzero,
". Switching to ARIMA(0,1,1) model."), call.=FALSE);
# Add ARIMA(0,1,1) lags
lagsModelAll <- matrix(1,1,1);
arRequired <- FALSE;
iRequired <- maRequired <- arimaModel <- TRUE;
arOrders <- 0;
iOrders <- maOrders <- 1;
nonZeroARI <- nonZeroMA <- matrix(c(2,1),1,2);
lagsModelARIMA <- matrix(1,1,1);
componentsNamesARIMA <- 1;
initialArimaNumber <- componentsNumberARIMA <- 1;
lagsModelMax <- max(lagsModelAll);
}
}
}
if(arimaModel && obsNonzero < (initialType=="optimal")*initialArimaNumber && !select){
warning(paste0("In-sample size is ",obsNonzero,", while number of ARIMA components is ",initialArimaNumber,
". Cannot fit the model."),call.=FALSE)
stop("Not enough observations for such a complicated model.",call.=FALSE);
}
# Recalculate the number of parameters
nParamMax[] <- (1 +
# ETS model
etsModel*(persistenceLevelEstimate + modelIsTrendy*persistenceTrendEstimate +
modelIsSeasonal*sum(persistenceSeasonalEstimate) +
phiEstimate + (initialType=="optimal") *
(initialLevelEstimate + initialTrendEstimate + sum(initialSeasonalEstimate*lagsModelSeasonal))) +
# ARIMA components: initials + parameters
arimaModel*((initialType=="optimal")*initialArimaNumber +
arRequired*arEstimate*sum(arOrders) + maRequired*maEstimate*sum(maOrders)) +
# Xreg initials and smoothing parameters
xregModel*(xregNumber*(any(initialType==c("optimal","backcasting"))*initialXregEstimate+persistenceXregEstimate)));
# If the sample is still smaller than the number of parameters (even after removing ARIMA)
if(etsModel){
if(obsNonzero <= nParamMax){
nParamExo <- xregNumber*(initialXregEstimate+persistenceXregEstimate);
if(!silent){
message(paste0("Number of non-zero observations is ",obsNonzero,
", while the maximum number of parameters to estimate is ", nParamMax,".\n",
"Updating pool of models."));
}
# If the number of observations is still enough for the model selection and the pool is not specified
if(obsNonzero > (3 + nParamExo) && is.null(modelsPool) && any(modelDo==c("select","combine"))){
# We have enough observations for local level model
modelsPool <- c("ANN");
if(allowMultiplicative){
modelsPool <- c(modelsPool,"MNN");
}
# We have enough observations for trend model
if(obsNonzero > (5 + nParamExo)){
if(any(Ttype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"AAN");
}
if(allowMultiplicative && any(Ttype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"AMN","MAN","MMN");
}
}
# We have enough observations for damped trend model
if(obsNonzero > (6 + nParamExo)){
if(any(Ttype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"AAdN");
}
if(allowMultiplicative && any(Ttype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"AMdN","MAdN","MMdN");
}
}
# We have enough observations for seasonal model
if((obsNonzero > (2*lagsModelMax)) && lagsModelMax!=1){
if(any(Stype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"ANA");
}
if(allowMultiplicative && any(Stype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"ANM","MNA","MNM");
}
}
# We have enough observations for seasonal model with trend
if((obsNonzero > (6 + lagsModelMax + nParamExo)) &&
(obsNonzero > 2*lagsModelMax) && lagsModelMax!=1){
if(any(Ttype==c("Z","X","A")) && any(Stype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"AAA");
}
if(allowMultiplicative && any(Ttype==c("Z","X","A")) && any(Stype==c("Z","Y","A"))){
modelsPool <- c(modelsPool,"MAA");
}
if(allowMultiplicative && any(Ttype==c("Z","X","A")) && any(Stype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"AAM","MAM");
}
if(allowMultiplicative && any(Ttype==c("Z","Y","M")) && any(Stype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"AMA","MMA");
}
if(allowMultiplicative && any(Ttype==c("Z","Y","M")) && any(Stype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"AMM","MMM");
}
}
warning("Not enough of non-zero observations for the fit of ETS(",model,")! Fitting what I can...",
call.=FALSE);
if(modelDo=="combine"){
model <- "CNN";
if(length(modelsPool)>2){
model <- "CCN";
}
if(length(modelsPool)>10){
model <- "CCC";
}
}
else{
modelDo <- "select"
model <- "ZZZ";
}
}
# If the pool is provided (so, select / combine), amend it
else if(obsNonzero > (3 + nParamExo) && !is.null(modelsPool)){
# We don't have enough observations for seasonal models with damped trend
if((obsNonzero <= (6 + lagsModelMax + 1 + nParamExo))){
modelsPool <- modelsPool[!(nchar(modelsPool)==4 &
substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="A")];
modelsPool <- modelsPool[!(nchar(modelsPool)==4 &
substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="M")];
}
# We don't have enough observations for seasonal models with trend
if((obsNonzero <= (5 + lagsModelMax + 1 + nParamExo))){
modelsPool <- modelsPool[!(substr(modelsPool,2,2)!="N" &
substr(modelsPool,nchar(modelsPool),nchar(modelsPool))!="N")];
}
# We don't have enough observations for seasonal models
if(obsNonzero <= 2*lagsModelMax){
modelsPool <- modelsPool[substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="N"];
}
# We don't have enough observations for damped trend
if(obsNonzero <= (6 + nParamExo)){
modelsPool <- modelsPool[nchar(modelsPool)!=4];
}
# We don't have enough observations for any trend
if(obsNonzero <= (5 + nParamExo)){
modelsPool <- modelsPool[substr(modelsPool,2,2)=="N"];
}
modelsPool <- unique(modelsPool);
warning("Not enough of non-zero observations for the fit of ETS(",model,")! Fitting what I can...",
call.=FALSE);
if(modelDo=="combine"){
model <- "CNN";
if(length(modelsPool)>2){
model <- "CCN";
}
if(length(modelsPool)>10){
model <- "CCC";
}
}
else{
modelDo <- "select"
model <- "ZZZ";
}
}
# If the model needs to be estimated / used, not selected
else if(obsNonzero > (3 + nParamExo) && any(modelDo==c("estimate","use"))){
# We don't have enough observations for seasonal models with damped trend
if((obsNonzero <= (6 + lagsModelMax + 1 + nParamExo))){
if(nchar(model)==4){
model <- paste0(substr(model,1,2),substr(model,4,4));
}
# model <- model[!(nchar(model)==4 &
# substr(model,nchar(model),nchar(model))=="A")];
# model <- model[!(nchar(model)==4 &
# substr(model,nchar(model),nchar(model))=="M")];
}
# We don't have enough observations for seasonal models with trend
if((obsNonzero <= (5 + lagsModelMax + 1 + nParamExo))){
model <- paste0(substr(model,1,1),"N",substr(model,3,3));
# model <- model[!(substr(model,2,2)!="N" &
# substr(model,nchar(model),nchar(model))!="N")];
}
# We don't have enough observations for seasonal models
if(obsNonzero <= 2*lagsModelMax){
model <- paste0(substr(model,1,2),"N");
# model <- model[substr(model,nchar(model),nchar(model))=="N"];
}
# We don't have enough observations for damped trend
if(obsNonzero <= (6 + nParamExo)){
if(nchar(model)==4){
model <- paste0(substr(model,1,2),substr(model,4,4));
}
# model <- model[nchar(model)!=4];
}
# We don't have enough observations for any trend
if(obsNonzero <= (5 + nParamExo)){
model <- paste0(substr(model,1,1),"N",substr(model,3,3));
# model <- model[substr(model,2,2)=="N"];
}
}
# Extreme cases of small samples
else if(obsNonzero==4){
if(any(Etype==c("A","M"))){
modelDo <- "estimate";
Ttype <- "N";
Stype <- "N";
}
else{
modelsPool <- c("ANN");
if(allowMultiplicative){
modelsPool <- c(modelsPool,"MNN");
}
modelDo <- "select";
model <- "ZZZ";
Etype <- "Z";
Ttype <- "N";
Stype <- "N";
warning("You have a very small sample. The only available model is level model.",
call.=FALSE);
}
phiEstimate <- FALSE;
}
# Even smaller sample
else if(obsNonzero==3){
if(any(Etype==c("A","M"))){
modelDo <- "estimate";
Ttype <- "N";
Stype <- "N";
model <- paste0(Etype,"NN");
}
else{
modelsPool <- c("ANN");
if(allowMultiplicative){
modelsPool <- c(modelsPool,"MNN");
}
modelDo <- "select";
model <- "ZNN";
Etype <- "Z";
Ttype <- "N";
Stype <- "N";
}
persistenceLevel <- 0;
persistenceEstimate <- persistenceLevelEstimate <- FALSE;
warning("I did not have enough of non-zero observations, so persistence value was set to zero.",
call.=FALSE);
phiEstimate <- FALSE;
}
# Can it be even smaller?
else if(obsNonzero==2){
modelsPool <- NULL;
persistenceLevel <- 0;
persistenceEstimate <- persistenceLevelEstimate <- FALSE;
initialLevel <- mean(yInSample);
initialType <- "provided";
initialEstimate <- initialLevelEstimate <- FALSE;
warning("I did not have enough of non-zero observations, so persistence value was set to zero and initial was preset.",
call.=FALSE);
modelDo <- "use";
model <- "ANN";
Etype <- "A";
Ttype <- "N";
Stype <- "N";
phiEstimate <- FALSE;
parametersNumber[1,1] <- 0;
parametersNumber[2,1] <- 2;
}
# And how about now?!
else if(obsNonzero==1){
modelsPool <- NULL;
persistenceLevel <- 0;
persistenceEstimate <- persistenceLevelEstimate <- FALSE;
initialLevel <- yInSample[yInSample!=0];
initialType <- "provided";
initialEstimate <- initialLevelEstimate <- FALSE;
warning("I did not have enough of non-zero observations, so I used Naive.",call.=FALSE);
modelDo <- "nothing"
model <- "ANN";
Etype <- "A";
Ttype <- "N";
Stype <- "N";
phiEstimate <- FALSE;
parametersNumber[1,1] <- 0;
parametersNumber[2,1] <- 2;
}
# Only zeroes in the data...
else if(obsNonzero==0 && obsInSample>1){
modelsPool <- NULL;
persistenceLevel <- 0;
persistenceEstimate <- persistenceLevelEstimate <- FALSE;
initialLevel <- 0;
initialType <- "provided";
initialEstimate <- initialLevelEstimate <- FALSE;
occurrenceModelProvided <- occurrenceModel <- FALSE;
occurrence <- "none";
warning("You have a sample with zeroes only. Your forecast will be zero.",call.=FALSE);
modelDo <- "nothing"
model <- "ANN";
Etype <- "A";
Ttype <- "N";
Stype <- "N";
phiEstimate <- FALSE;
parametersNumber[1,1] <- 0;
parametersNumber[2,1] <- 2;
}
# If you don't have observations, then fuck off!
else{
stop("Not enough observations... Even for fitting of ETS('ANN')!",call.=FALSE);
}
}
}
# Reset the maximum lag. This is in order to take potential changes into account
lagsModelMax[] <- max(lagsModelAll);
#### Process ellipsis ####
# Parameters for the optimiser
if(is.null(ellipsis$maxeval)){
maxeval <- NULL;
# Make a warning if this is a big computational task
if(any(lags>24) && arimaModel && initialType=="optimal"){
warning(paste0("The estimation of ARIMA model with initial='optimal' on high frequency data might ",
"take more time to converge to the optimum. Consider either setting maxeval parameter ",
"to a higher value (e.g. maxeval=10000, which will take ~25 times more than this) ",
"or using initial='backcasting'."),
call.=FALSE, immediate.=TRUE);
}
}
else{
maxeval <- ellipsis$maxeval;
}
if(is.null(ellipsis$maxtime)){
maxtime <- -1;
}
else{
maxtime <- ellipsis$maxtime;
}
if(is.null(ellipsis$xtol_rel)){
xtol_rel <- 1E-6;
}
else{
xtol_rel <- ellipsis$xtol_rel;
}
if(is.null(ellipsis$xtol_abs)){
xtol_abs <- 1E-8;
}
else{
xtol_abs <- ellipsis$xtol_abs;
}
if(is.null(ellipsis$ftol_rel)){
ftol_rel <- 1E-8;
}
else{
ftol_rel <- ellipsis$ftol_rel;
}
if(is.null(ellipsis$ftol_abs)){
ftol_abs <- 0;
}
else{
ftol_abs <- ellipsis$ftol_abs;
}
if(is.null(ellipsis$algorithm)){
algorithm <- "NLOPT_LN_SBPLX";
}
else{
algorithm <- ellipsis$algorithm;
}
if(is.null(ellipsis$print_level)){
print_level <- 0;
}
else{
print_level <- ellipsis$print_level;
}
# The following three arguments are used for the function itself, not the options
if(is.null(ellipsis$lb)){
lb <- NULL;
}
else{
lb <- ellipsis$lb;
}
if(is.null(ellipsis$ub)){
ub <- NULL;
}
else{
ub <- ellipsis$ub;
}
if(is.null(ellipsis$B)){
B <- NULL;
}
else{
B <- ellipsis$B;
}
# Initialise parameters
lambda <- other <- NULL;
otherParameterEstimate <- FALSE
# lambda for LASSO
if(any(loss==c("LASSO","RIDGE"))){
if(is.null(ellipsis$lambda)){
warning(paste0("You have not provided lambda parameter. I will set it to zero."), call.=FALSE);
lambda <- 0;
}
else{
lambda <- ellipsis$lambda;
}
}
# Parameters for distributions
if(distribution=="dalaplace"){
if(is.null(ellipsis$alpha)){
other <- 0.5
otherParameterEstimate <- TRUE;
}
else{
other <- ellipsis$alpha;
otherParameterEstimate <- FALSE;
}
names(other) <- "alpha";
}
else if(any(distribution==c("dgnorm","dlgnorm"))){
if(is.null(ellipsis$shape)){
other <- 2
otherParameterEstimate <- TRUE;
}
else{
other <- ellipsis$shape;
otherParameterEstimate <- FALSE;
}
names(other) <- "shape";
}
else if(distribution=="dt"){
if(is.null(ellipsis$nu)){
other <- 2
otherParameterEstimate <- TRUE;
}
else{
other <- ellipsis$nu;
otherParameterEstimate <- FALSE;
}
names(other) <- "nu";
}
# Fisher Information
if(is.null(ellipsis$FI)){
FI <- FALSE;
}
else{
FI <- ellipsis$FI;
}
# Step size for the hessian
if(is.null(ellipsis$stepSize)){
stepSize <- .Machine$double.eps^(1/4);
}
else{
stepSize <- ellipsis$stepSize;
}
# Add constant in the model
if(is.numeric(constant)){
constantRequired <- TRUE;
constantEstimate <- FALSE;
# This is just in case a vector was provided
constantValue <- constant[1];
if(any(iOrders!=0) || etsModel){
constantName <- "drift";
}
else{
constantName <- "constant";
}
}
else if(is.logical(constant)){
constantEstimate <- constantRequired <- constant;
constantValue <- NULL;
if(constantRequired){
if(any(iOrders!=0) || etsModel){
constantName <- "drift";
}
else{
constantName <- "constant";
}
}
else{
constantName <- NULL;
}
}
else{
warning("The parameter constant can only be TRUE or FALSE.",
"You have: ",constant,". Switching to FALSE",call.=FALSE);
constantEstimate <- constantRequired <- FALSE;
constantName <- NULL;
}
# If there is no model, return a constant level
if(!etsModel && !arimaModel && !xregModel){
modelsPool <- NULL;
if(!constantRequired){
constantEstimate <- constantRequired <- TRUE;
constantName <- "constant";
}
model <- "NNN";
if(is.null(B)){
modelDo <- "estimate";
}
Etype <- switch(distribution,
"default"=,"dnorm"=,"dlaplace"=,"ds"=,"dgnorm"=,"dlogis"=,"dt"=,"dalaplace"="A",
"dlnorm"=,"dllaplace"=,"dls"=,"dlgnorm"=,"dinvgauss"=,"dgamma"="M");
Ttype <- "N";
Stype <- "N";
phiEstimate <- FALSE;
parametersNumber[1,1] <- 0;
parametersNumber[2,1] <- 2;
}
# See if the estimation of the model is not needed (do we estimate anything?)
if(!any(c(etsModel & c(persistenceLevelEstimate, persistenceTrendEstimate,
persistenceSeasonalEstimate, phiEstimate,
(initialType!="complete") & c(initialLevelEstimate,
initialTrendEstimate,
initialSeasonalEstimate)),
arimaModel & c(arEstimate, maEstimate, (initialType!="complete") & initialArimaEstimate),
xregModel & c(persistenceXregEstimate, (initialType!="complete") & initialXregEstimate),
constantEstimate,
otherParameterEstimate))){
modelDo <- "use";
}
# Switch usual bounds to the admissible if there's no ETS - this speeds up ARIMA
if(!etsModel && bounds=="usual"){
bounds[] <- "admissible";
}
# If we do model selection / combination with non-standard losses, complain
if(any(modelDo==c("select","combine")) &&
((any(loss==c("MSE","MSEh","MSCE","GPL")) && all(distribution!=c("default","dnorm"))) ||
(any(loss==c("MAE","MAEh","MACE")) && all(distribution!=c("default","dlaplace"))) ||
(any(loss==c("HAM","HAMh","CHAM")) && all(distribution!=c("default","ds"))) ||
(any(loss==c("TMSE","GTMSE","TMAE","GTMAE","THAM","GTHAM"))))){
warning("The model selection only works in case of loss='likelihood'. I hope you know what you are doing.",
call.=FALSE);
}
# Just in case, give names to yHoldout and yInSample
colnames(yInSample) <- responseName;
if(holdout){
colnames(yHoldout) <- responseName;
}
#### Return the values to the previous environment ####
### Actuals
assign("y",y,ParentEnvironment);
assign("yHoldout",yHoldout,ParentEnvironment);
assign("yInSample",yInSample,ParentEnvironment);
assign("yNAValues",yNAValues,ParentEnvironment);
### Index and all related structure variables
assign("yClasses",yClasses,ParentEnvironment);
assign("yIndex",yIndex,ParentEnvironment);
assign("yInSampleIndex",yInSampleIndex,ParentEnvironment);
assign("yForecastIndex",yForecastIndex,ParentEnvironment);
assign("yIndexAll",yIndexAll,ParentEnvironment);
assign("yFrequency",yFrequency,ParentEnvironment);
assign("yStart",yStart,ParentEnvironment);
assign("yForecastStart",yForecastStart,ParentEnvironment);
# The rename of the variable is needed for the hessian to work
assign("horizon",h,ParentEnvironment);
assign("h",h,ParentEnvironment);
assign("holdout",holdout,ParentEnvironment);
### Number of observations and parameters
assign("obsInSample",obsInSample,ParentEnvironment);
assign("obsAll",obsAll,ParentEnvironment);
assign("obsStates",obsStates,ParentEnvironment);
assign("obsNonzero",obsNonzero,ParentEnvironment);
assign("obsZero",obsZero,ParentEnvironment);
assign("parametersNumber",parametersNumber,ParentEnvironment);
### Model type
assign("etsModel",etsModel,ParentEnvironment);
assign("model",model,ParentEnvironment);
assign("Etype",Etype,ParentEnvironment);
assign("Ttype",Ttype,ParentEnvironment);
assign("Stype",Stype,ParentEnvironment);
assign("modelIsTrendy",modelIsTrendy,ParentEnvironment);
assign("modelIsSeasonal",modelIsSeasonal,ParentEnvironment);
assign("modelsPool",modelsPool,ParentEnvironment);
assign("damped",damped,ParentEnvironment);
assign("modelDo",modelDo,ParentEnvironment);
assign("allowMultiplicative",allowMultiplicative,ParentEnvironment);
### Numbers and names of components
assign("componentsNumberETS",componentsNumberETS,ParentEnvironment);
assign("componentsNamesETS",componentsNamesETS,ParentEnvironment);
assign("componentsNumberETSNonSeasonal",componentsNumberETS-componentsNumberETSSeasonal,ParentEnvironment);
assign("componentsNumberETSSeasonal",componentsNumberETSSeasonal,ParentEnvironment);
# The number and names of ARIMA components
assign("componentsNumberARIMA",componentsNumberARIMA,ParentEnvironment);
assign("componentsNamesARIMA",componentsNamesARIMA,ParentEnvironment);
### Lags
# This is the original vector of lags, modified for the level components.
# This can be used in ARIMA
assign("lags",lags,ParentEnvironment);
# This is the vector of lags of ETS components
assign("lagsModel",lagsModel,ParentEnvironment);
# This is the vector of seasonal lags
assign("lagsModelSeasonal",lagsModelSeasonal,ParentEnvironment);
# This is the vector of lags for ARIMA components (not lags of ARIMA)
assign("lagsModelARIMA",lagsModelARIMA,ParentEnvironment);
# This is the vector of all the lags of model (ETS + ARIMA + X)
assign("lagsModelAll",lagsModelAll,ParentEnvironment);
# This is the maximum lag
assign("lagsModelMax",lagsModelMax,ParentEnvironment);
### Persistence
assign("persistence",persistence,ParentEnvironment);
assign("persistenceEstimate",persistenceEstimate,ParentEnvironment);
assign("persistenceLevel",persistenceLevel,ParentEnvironment);
assign("persistenceLevelEstimate",persistenceLevelEstimate,ParentEnvironment);
assign("persistenceTrend",persistenceTrend,ParentEnvironment);
assign("persistenceTrendEstimate",persistenceTrendEstimate,ParentEnvironment);
assign("persistenceSeasonal",persistenceSeasonal,ParentEnvironment);
assign("persistenceSeasonalEstimate",persistenceSeasonalEstimate,ParentEnvironment);
assign("persistenceXreg",persistenceXreg,ParentEnvironment);
assign("persistenceXregEstimate",persistenceXregEstimate,ParentEnvironment);
assign("persistenceXregProvided",persistenceXregProvided,ParentEnvironment);
### phi
assign("phi",phi,ParentEnvironment);
assign("phiEstimate",phiEstimate,ParentEnvironment);
### Initials
assign("initial",initial,ParentEnvironment);
assign("initialType",initialType,ParentEnvironment);
assign("initialEstimate",initialEstimate,ParentEnvironment);
assign("initialLevel",initialLevel,ParentEnvironment);
assign("initialLevelEstimate",initialLevelEstimate,ParentEnvironment);
assign("initialTrend",initialTrend,ParentEnvironment);
assign("initialTrendEstimate",initialTrendEstimate,ParentEnvironment);
assign("initialSeasonal",initialSeasonal,ParentEnvironment);
assign("initialSeasonalEstimate",initialSeasonalEstimate,ParentEnvironment);
assign("initialArima",initialArima,ParentEnvironment);
assign("initialArimaEstimate",initialArimaEstimate,ParentEnvironment);
# Number of initials that the ARIMA has (either provided or to estimate)
assign("initialArimaNumber",initialArimaNumber,ParentEnvironment);
assign("initialXregEstimate",initialXregEstimate,ParentEnvironment);
assign("initialXregProvided",initialXregProvided,ParentEnvironment);
### Occurrence model
assign("oesModel",oesModel,ParentEnvironment);
assign("occurrenceModel",occurrenceModel,ParentEnvironment);
assign("occurrenceModelProvided",occurrenceModelProvided,ParentEnvironment);
assign("occurrence",occurrence,ParentEnvironment);
assign("pFitted",pFitted,ParentEnvironment);
assign("pForecast",pForecast,ParentEnvironment);
assign("ot",ot,ParentEnvironment);
assign("otLogical",otLogical,ParentEnvironment);
### Outliers detection
assign("outliers",outliers,ParentEnvironment);
### Distribution, loss, bounds and IC
assign("distribution",distribution,ParentEnvironment);
assign("loss",loss,ParentEnvironment);
assign("lossFunction",lossFunction,ParentEnvironment);
assign("multisteps",multisteps,ParentEnvironment);
assign("ic",ic,ParentEnvironment);
assign("icFunction",icFunction,ParentEnvironment);
assign("bounds",bounds,ParentEnvironment);
### ARIMA components
assign("arimaModel",arimaModel,ParentEnvironment);
assign("arOrders",arOrders,ParentEnvironment);
assign("iOrders",iOrders,ParentEnvironment);
assign("maOrders",maOrders,ParentEnvironment);
assign("arRequired",arRequired,ParentEnvironment);
assign("iRequired",iRequired,ParentEnvironment);
assign("maRequired",maRequired,ParentEnvironment);
assign("arEstimate",arEstimate,ParentEnvironment);
assign("maEstimate",maEstimate,ParentEnvironment);
assign("armaParameters",armaParameters,ParentEnvironment);
assign("nonZeroARI",nonZeroARI,ParentEnvironment);
assign("nonZeroMA",nonZeroMA,ParentEnvironment);
assign("select",select,ParentEnvironment);
### Explanatory variables
assign("xregModel",xregModel,ParentEnvironment);
assign("regressors",regressors,ParentEnvironment);
assign("xregModelInitials",xregModelInitials,ParentEnvironment);
assign("xregData",xregData,ParentEnvironment);
assign("xregNumber",xregNumber,ParentEnvironment);
assign("xregNames",xregNames,ParentEnvironment);
assign("responseName",responseName,ParentEnvironment);
assign("formula",formulaToUse,ParentEnvironment);
assign("xregParametersMissing",xregParametersMissing,ParentEnvironment);
assign("xregParametersIncluded",xregParametersIncluded,ParentEnvironment);
assign("xregParametersEstimated",xregParametersEstimated,ParentEnvironment);
assign("xregParametersPersistence",xregParametersPersistence,ParentEnvironment);
### Constant
assign("constantRequired",constantRequired,ParentEnvironment);
assign("constantEstimate",constantEstimate,ParentEnvironment);
assign("constantValue",constantValue,ParentEnvironment);
assign("constantName",constantName,ParentEnvironment);
### Ellipsis thingies
# Optimisation related
assign("maxeval",maxeval,ParentEnvironment);
assign("maxtime",maxtime,ParentEnvironment);
assign("xtol_rel",xtol_rel,ParentEnvironment);
assign("xtol_abs",xtol_abs,ParentEnvironment);
assign("ftol_rel",ftol_rel,ParentEnvironment);
assign("ftol_abs",ftol_abs,ParentEnvironment);
assign("algorithm",algorithm,ParentEnvironment);
assign("print_level",print_level,ParentEnvironment);
assign("B",B,ParentEnvironment);
assign("lb",lb,ParentEnvironment);
assign("ub",ub,ParentEnvironment);
# Parameters for distributions
assign("other",other,ParentEnvironment);
assign("otherParameterEstimate",otherParameterEstimate,ParentEnvironment);
# LASSO / RIDGE
assign("lambda",lambda,ParentEnvironment);
# Fisher Information
assign("FI",FI,ParentEnvironment);
# Step size for the hessian
assign("stepSize",stepSize,ParentEnvironment);
return(list(select=FALSE));
}
config-i1/smooth documentation built on April 18, 2024, 6:25 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Defines functions parametersChecker
parametersChecker <- function(data, model, lags, formulaToUse, orders, constant=FALSE, arma,
outliers=c("ignore","use","select"), level=0.99,
persistence, phi, initial,
distribution=c("default","dnorm","dlaplace","dalaplace","ds","dgnorm",
"dlnorm","dinvgauss","dgamma"),
loss, h, holdout, occurrence,
ic=c("AICc","AIC","BIC","BICc"), bounds=c("traditional","usual","admissible","none"),
regressors, yName,
silent, modelDo, ParentEnvironment,
ellipsis, fast=FALSE){
# The function checks the provided parameters of adam and/or oes
##### data #####
# yName is the name of the object. It might differ from response if matrix is provided
responseName <- yName;
# If this is simulated, extract the actuals
if(is.adam.sim(data) || is.smooth.sim(data)){
data <- data$data;
}
# If this is Mdata, use all the available stuff
else if(inherits(data,"Mdata")){
h <- data$h;
holdout <- TRUE;
if(modelDo!="use"){
lags <- frequency(data$x);
}
data <- ts(c(data$x,data$xx),start=start(data$x),frequency=frequency(data$x));
}
# Extract index from the object in order to use it later
### tsibble has its own index function, so shit happens because of it...
if(inherits(data,"tbl_ts")){
yIndex <- data[[1]];
if(any(duplicated(yIndex))){
warning(paste0("You have duplicated time stamps in the variable ",yName,
". I will refactor this."),call.=FALSE);
yIndex <- yIndex[1] + c(1:length(data[[1]])) * diff(tail(yIndex,2));
}
}
else{
yIndex <- try(time(data),silent=TRUE);
# If we cannot extract time, do something
if(inherits(yIndex,"try-error")){
if(!is.data.frame(data) && !is.null(dim(data))){
yIndex <- as.POSIXct(rownames(data));
}
else if(is.data.frame(data)){
yIndex <- c(1:nrow(data));
}
else{
yIndex <- c(1:length(data));
}
}
}
yClasses <- class(data);
# If this is something like a matrix
if(!is.null(ncol(data)) && ncol(data)>1){
xregData <- data;
# Get rid of the bloody tibble class. Gives me headaches!
if(inherits(data,"tbl_df") || inherits(data,"tbl")){
data <- as.data.frame(data);
}
if(!is.null(formulaToUse)){
responseName <- all.vars(formulaToUse)[1];
y <- data[,responseName];
}
else{
responseName <- colnames(xregData)[1];
# If we deal with data.table / tibble / data.frame, the syntax is different.
# We don't want to import specific classes, so just use inherits()
if(inherits(data,"tbl_ts")){
# With tsibble we cannot extract explanatory variables easily...
y <- data$value;
}
else if(inherits(data,"data.table") || inherits(data,"data.frame")){
y <- data[[1]];
}
else if(inherits(data,"zoo")){
if(ncol(data)>1){
xregData <- as.data.frame(data);
}
y <- zoo(data[,1],order.by=time(data));
}
else{
y <- data[,1];
}
}
# Give the indeces another try
yIndex <- try(time(y),silent=TRUE);
# If we cannot extract time, do something
if(inherits(yIndex,"try-error")){
if(!is.null(dim(data))){
yIndex <- try(as.POSIXct(rownames(data)),silent=TRUE);
if(inherits(yIndex,"try-error")){
yIndex <- c(1:nrow(data));
}
}
else{
yIndex <- c(1:length(y));
}
}
else{
yClasses <- class(y);
}
}
else{
xregData <- NULL;
if(!is.null(ncol(data)) && !is.null(colnames(data)[1])){
responseName <- colnames(data)[1];
y <- data[,1];
}
else{
y <- data;
}
}
# Make the response a secure name
responseName <- make.names(responseName);
# Define obs, the number of observations of in-sample
obsAll <- length(y) + (1 - holdout)*h;
obsInSample <- length(y) - holdout*h;
if(obsInSample<=0){
stop("The number of in-sample observations is not positive. Cannot do anything.",
call.=FALSE);
}
# Interpolate NAs using fourier + polynomials
yNAValues <- is.na(y);
if(any(yNAValues)){
warning("Data contains NAs. The values will be ignored during the model construction.",call.=FALSE);
X <- cbind(1,poly(c(1:obsAll),degree=min(max(trunc(obsAll/10),1),5)),
sinpi(matrix(c(1:obsAll)*rep(c(1:max(max(lags),10)),each=obsAll)/max(max(lags),10), ncol=max(max(lags),10))));
# If we deal with purely positive data, take logarithms to deal with multiplicative seasonality
if(any(y[!yNAValues]<=0)){
lmFit <- .lm.fit(X[!yNAValues,,drop=FALSE], matrix(y[!yNAValues],ncol=1));
y[yNAValues] <- (X %*% coef(lmFit))[yNAValues];
}
else{
lmFit <- .lm.fit(X[!yNAValues,,drop=FALSE], matrix(log(y[!yNAValues]),ncol=1));
y[yNAValues] <- exp(X %*% coef(lmFit))[yNAValues];
}
if(!is.null(xregData)){
xregData[yNAValues,responseName] <- y[yNAValues];
}
rm(X);
# Clean memory if have a big object
if(obsInSample>10000){
gc(verbose=FALSE);
}
}
# If this is just a numeric variable, use ts class
if(all(yClasses=="integer") || all(yClasses=="numeric") ||
all(yClasses=="data.frame") || all(yClasses=="matrix")){
if(any(class(yIndex) %in% c("POSIXct","Date"))){
yClasses <- "zoo";
}
else{
yClasses <- "ts";
}
}
yFrequency <- frequency(y);
yStart <- yIndex[1];
yInSample <- matrix(y[1:obsInSample],ncol=1);
if(holdout){
yForecastStart <- yIndex[obsInSample+1];
yHoldout <- matrix(y[-c(1:obsInSample)],ncol=1);
yForecastIndex <- yIndex[-c(1:obsInSample)];
yInSampleIndex <- yIndex[c(1:obsInSample)];
yIndexAll <- yIndex;
}
else{
yInSampleIndex <- yIndex;
yIndexDiff <- diff(tail(yIndex,2));
yForecastStart <- yIndex[obsInSample]+yIndexDiff;
if(any(yClasses=="ts")){
yForecastIndex <- yIndex[obsInSample]+as.numeric(yIndexDiff)*c(1:max(h,1));
}
else{
yForecastIndex <- yIndex[obsInSample]+yIndexDiff*c(1:max(h,1));
}
yHoldout <- NULL;
yIndexAll <- c(yIndex,yForecastIndex);
}
if(!is.numeric(yInSample)){
stop("The provided data is not numeric! Can't construct any model!", call.=FALSE);
}
# If the user asked for trend, but it's not in the data, add it
if(!is.null(formulaToUse) &&
any(all.vars(formulaToUse)=="trend") && all(colnames(xregData)!="trend")){
if(!is.null(xregData)){
xregData <- cbind(xregData,trend=c(1:obsAll));
}
else{
xregData <- cbind(y=y,trend=c(1:obsAll));
}
}
# Number of parameters to estimate / provided
parametersNumber <- matrix(0,2,5,
dimnames=list(c("Estimated","Provided"),
c("nParamInternal","nParamXreg","nParamOccurrence","nParamScale","nParamAll")));
#### Check what is used for the model ####
if(!is.character(model)){
stop(paste0("Something strange is provided instead of character object in model: ",
paste0(model,collapse=",")),call.=FALSE);
}
# Predefine models pool for a model selection
modelsPool <- NULL;
if(!fast){
# Deal with the list of models. Check what has been provided. Stop if there is a mistake.
if(length(model)>1){
if(any(nchar(model)>4) || any(nchar(model)<3)){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[nchar(model)>4 | nchar(model)<3],collapse=",")),call.=FALSE);
}
else if(any(substr(model,1,1)!="A" & substr(model,1,1)!="M" & substr(model,1,1)!="C")){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[substr(model,1,1)!="A" & substr(model,1,1)!="M"],collapse=",")),
call.=FALSE);
}
else if(any(substr(model,2,2)!="N" & substr(model,2,2)!="A" &
substr(model,2,2)!="M" & substr(model,2,2)!="C")){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[substr(model,2,2)!="N" & substr(model,2,2)!="A" &
substr(model,2,2)!="M"],collapse=",")),call.=FALSE);
}
else if(any(substr(model,3,3)!="N" & substr(model,3,3)!="A" &
substr(model,3,3)!="M" & substr(model,3,3)!="d" & substr(model,3,3)!="C")){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[substr(model,3,3)!="N" & substr(model,3,3)!="A" &
substr(model,3,3)!="M" & substr(model,3,3)!="d"],collapse=",")),
call.=FALSE);
}
else if(any(nchar(model)==4 & substr(model,4,4)!="N" & substr(model,4,4)!="A" &
substr(model,4,4)!="M" & substr(model,4,4)!="C")){
stop(paste0("You have defined strange model(s) in the pool: ",
paste0(model[nchar(model)==4 & substr(model,4,4)!="N" &
substr(model,4,4)!="A" & substr(model,4,4)!="M"],collapse=",")),
call.=FALSE);
}
else{
modelsPoolCombiner <- (substr(model,1,1)=="C" | substr(model,2,2)=="C" |
substr(model,3,3)=="C" | substr(model,4,4)=="C");
modelsPool <- model[!modelsPoolCombiner];
modelsPool <- unique(modelsPool);
if(any(modelsPoolCombiner)){
if(any(substr(model,nchar(model),nchar(model))!="N")){
model <- "CCC";
}
else{
model <- "CCN";
}
}
else{
model <- c("Z","Z","Z");
if(all(substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="N")){
model[3] <- "N";
}
if(all(substr(modelsPool,2,2)=="N")){
model[2] <- "N";
}
model <- paste0(model,collapse="");
}
}
}
}
# If chosen model is "AAdN" or anything like that, we are taking the appropriate values
if(nchar(model)==4){
Etype <- substr(model,1,1);
Ttype <- substr(model,2,2);
Stype <- substr(model,4,4);
damped <- TRUE;
if(substr(model,3,3)!="d"){
message(paste0("You have defined a strange model: ", model,
". Switching to ", paste0(Etype,Ttype,"d",Stype)));
model <- paste0(Etype,Ttype,"d",Stype);
}
}
else if(nchar(model)==3){
Etype <- substr(model,1,1);
Ttype <- substr(model,2,2);
Stype <- substr(model,3,3);
if(any(Ttype==c("Z","X","Y"))){
damped <- TRUE;
}
else{
damped <- FALSE;
}
}
else{
message(paste0("You have defined a strange model: ",model));
message("Switching to 'ZZZ'");
model <- "ZZZ";
Etype <- "Z";
Ttype <- "Z";
Stype <- "Z";
damped <- TRUE;
}
# Define if we want to select or combine models... or do none of the above.
if(is.null(modelsPool)){
if(any(unlist(strsplit(model,""))=="C")){
modelDo <- "combine";
if(Etype=="C"){
Etype <- "Z";
}
if(Ttype=="C"){
Ttype <- "Z";
}
if(Stype=="C"){
Stype <- "Z";
}
}
else if(any(unlist(strsplit(model,""))=="Z") ||
any(unlist(strsplit(model,""))=="X") ||
any(unlist(strsplit(model,""))=="Y") ||
any(unlist(strsplit(model,""))=="F") ||
any(unlist(strsplit(model,""))=="P")){
modelDo <- "select";
# The full test, sidestepping branch and bound
if(any(unlist(strsplit(model,""))=="F")){
modelsPool <- c("ANN","AAN","AAdN","AMN","AMdN",
"ANA","AAA","AAdA","AMA","AMdA",
"ANM","AAM","AAdM","AMM","AMdM",
"MNN","MAN","MAdN","MMN","MMdN",
"MNA","MAA","MAdA","MMA","MMdA",
"MNM","MAM","MAdM","MMM","MMdM");
# Remove models from pool if specific elements are provided
if(Etype!="F"){
modelsPool <- modelsPool[substr(modelsPool,1,1)==Etype];
}
else{
Etype[] <- "Z"
}
if(Ttype!="F"){
modelsPool <- modelsPool[substr(modelsPool,2,2)==Ttype];
}
else{
Ttype[] <- "Z"
}
if(Stype!="F"){
modelsPool <- modelsPool[substr(modelsPool,nchar(modelsPool),nchar(modelsPool))==Stype];
}
else{
Stype[] <- "Z"
}
model <- "FFF";
}
# The test for pure models only
if(any(unlist(strsplit(model,""))=="P")){
modelsPool <- c("ANN","AAN","AAdN","ANA","AAA","AAdA",
"MNN","MMN","MMdN","MNM","MMM","MMdM");
# Remove models from pool if specific elements are provided
if(Etype!="P"){
modelsPool <- modelsPool[substr(modelsPool,1,1)==Etype];
}
else{
Etype[] <- "Z"
}
if(Ttype!="P"){
modelsPool <- modelsPool[substr(modelsPool,2,2)==Ttype];
}
else{
Ttype[] <- "Z"
}
if(Stype!="P"){
modelsPool <- modelsPool[substr(modelsPool,nchar(modelsPool),nchar(modelsPool))==Stype];
}
else{
Stype[] <- "Z"
}
model <- "PPP";
}
}
else{
modelDo <- "estimate";
}
if(Etype=="X"){
Etype <- "A";
}
else if(Etype=="Y"){
Etype <- "M";
}
}
else{
if(any(unlist(strsplit(model,""))=="C")){
modelDo <- "combine";
}
else{
modelDo <- "select";
}
}
#### Check the components of model ####
### Check error type
if(all(Etype!=c("Z","X","Y","A","M","C","N"))){
warning(paste0("Wrong error type: ",Etype,". Should be 'Z', 'X', 'Y', 'C', 'A', 'M' or 'N'. ",
"Changing to 'Z'"),call.=FALSE);
Etype <- "Z";
modelDo <- "select";
}
# If the error is "N", then switch off the ETS model
if(Etype=="N"){
componentsNamesETS <- NULL;
componentsNumberETS <- 0;
Ttype[] <- "N";
damped[] <- FALSE;
Stype[] <- "N";
modelDo[] <- "estimate";
modelsPool[] <- NULL;
etsModel <- FALSE;
}
else{
componentsNamesETS <- "level";
componentsNumberETS <- 1;
etsModel <- TRUE;
}
if(etsModel){
### Check trend type
if(all(Ttype!=c("Z","X","Y","N","A","M","C"))){
warning(paste0("Wrong trend type: ",Ttype,". Should be 'Z', 'X', 'Y', 'C' 'N', 'A', or 'M'. ",
"Changing to 'Z'"),call.=FALSE);
Ttype <- "Z";
modelDo <- "select";
}
modelIsTrendy <- (Ttype!="N");
if(modelIsTrendy){
componentsNamesETS <- c(componentsNamesETS,"trend");
componentsNumberETS[] <- componentsNumberETS+1;
}
}
else{
modelIsTrendy <- FALSE;
}
#### Check the lags vector ####
if(any(c(lags)<0)){
stop("Right! Why don't you try complex lags then, mister smart guy?",call.=FALSE);
}
# If there are zero lags, drop them
if(any(lags==0)){
lags <- lags[lags!=0];
}
# Form the lags based on the provided stuff. Get rid of ones and leave unique seasonals
# Add one for the level
lags <- c(1,unique(lags[lags>1]));
# Warning if the lags length is higher than the sample size
if(max(lags) >= obsInSample){
warning("The maximum lags value is ", max(lags),
", while the sample size is ", obsInSample,
". I cannot guarantee that I'll be able to fit the model.",
call.=FALSE);
}
#### ARIMA term ####
# This should be available for pure models only
if(is.list(orders)){
arOrders <- orders$ar;
iOrders <- orders$i;
maOrders <- orders$ma;
select <- orders$select;
if(is.null(select)){
select <- FALSE;
}
}
else if(is.vector(orders)){
arOrders <- orders[1];
iOrders <- orders[2];
maOrders <- orders[3];
select <- FALSE;
}
else{
select <- FALSE;
}
# If there is arima, prepare orders
if(sum(c(arOrders,iOrders,maOrders))>0){
arimaModel <- TRUE;
# See if AR is needed
arRequired <- FALSE;
if(sum(arOrders)>0){
arRequired[] <- TRUE;
}
# See if I is needed
iRequired <- FALSE;
if(sum(iOrders)>0){
iRequired[] <- TRUE;
}
# See if I is needed
maRequired <- FALSE;
if(sum(maOrders)>0){
maRequired[] <- TRUE;
}
# Define maxOrder and make all the values look similar (for the polynomials)
maxOrder <- max(length(arOrders),length(iOrders),length(maOrders),length(lags));
if(length(arOrders)!=maxOrder){
arOrders <- c(arOrders,rep(0,maxOrder-length(arOrders)));
}
if(length(iOrders)!=maxOrder){
iOrders <- c(iOrders,rep(0,maxOrder-length(iOrders)));
}
if(length(maOrders)!=maxOrder){
maOrders <- c(maOrders,rep(0,maxOrder-length(maOrders)));
}
if(length(lags)!=maxOrder){
lagsNew <- c(lags,rep(0,maxOrder-length(lags)));
arOrders <- arOrders[lagsNew!=0];
iOrders <- iOrders[lagsNew!=0];
maOrders <- maOrders[lagsNew!=0];
}
# Define the non-zero values. This is done via the calculation of orders of polynomials
ariValues <- list(NA);
maValues <- list(NA);
for(i in 1:length(lags)){
ariValues[[i]] <- c(0,min(1,arOrders[i]):arOrders[i])
if(iOrders[i]!=0){
ariValues[[i]] <- c(ariValues[[i]],1:iOrders[i]+arOrders[i]);
}
ariValues[[i]] <- unique(ariValues[[i]] * lags[i]);
maValues[[i]] <- unique(c(0,min(1,maOrders[i]):maOrders[i]) * lags[i]);
}
# Produce ARI polynomials
ariLengths <- unlist(lapply(ariValues,length));
ariPolynomial <- array(0,ariLengths);
for(i in 1:length(ariValues)){
if(i==1){
ariPolynomial <- ariPolynomial + array(ariValues[[i]], ariLengths);
}
else{
ariPolynomial <- ariPolynomial + array(rep(ariValues[[i]],each=prod(ariLengths[1:(i-1)])),
ariLengths);
}
}
# Produce MA polynomials
maLengths <- unlist(lapply(maValues,length));
maPolynomial <- array(0,maLengths);
for(i in 1:length(maValues)){
if(i==1){
maPolynomial <- maPolynomial + array(maValues[[i]], maLengths);
}
else{
maPolynomial <- maPolynomial + array(rep(maValues[[i]],each=prod(maLengths[1:(i-1)])),
maLengths);
}
}
# What are the non-zero ARI and MA polynomials?
### What are their positions in transition matrix?
nonZeroARI <- unique(matrix(c(ariPolynomial)[-1],ncol=1));
nonZeroMA <- unique(matrix(c(maPolynomial)[-1],ncol=1));
# Lags for the ARIMA components
lagsModelARIMA <- matrix(sort(unique(c(nonZeroARI,nonZeroMA))),ncol=1);
nonZeroARI <- cbind(nonZeroARI+1,which(lagsModelARIMA %in% nonZeroARI));
nonZeroMA <- cbind(nonZeroMA+1,which(lagsModelARIMA %in% nonZeroMA));
# Number of components
componentsNumberARIMA <- length(lagsModelARIMA);
# Their names
componentsNamesARIMA <- paste0("ARIMAState",c(1:componentsNumberARIMA));
# If all orders are zero, drop ARIMA part
if(all(c(arOrders,iOrders,maOrders)==0)){
arOrders <- NULL;
iOrders <- NULL;
maOrders <- NULL;
arimaModel <- FALSE;
arRequired <- arEstimate <- FALSE;
iRequired <- FALSE;
maRequired <- maEstimate <- FALSE;
lagsModelARIMA <- initialArimaNumber <- 0;
componentsNumberARIMA <- 0;
componentsNamesARIMA <- NULL;
nonZeroARI <- NULL;
nonZeroMA <- NULL;
}
else{
# Number of initials needed. This is based on the longest one. The others are just its transformations
initialArimaNumber <- max(lagsModelARIMA);
}
}
else{
arOrders <- NULL;
iOrders <- NULL;
maOrders <- NULL;
arimaModel <- FALSE;
arRequired <- arEstimate <- FALSE;
iRequired <- FALSE;
maRequired <- maEstimate <- FALSE;
lagsModelARIMA <- initialArimaNumber <- 0;
componentsNumberARIMA <- 0;
componentsNamesARIMA <- NULL;
nonZeroARI <- NULL;
nonZeroMA <- NULL;
select <- FALSE;
}
if(etsModel){
modelIsSeasonal <- Stype!="N";
}
else{
modelIsSeasonal <- FALSE;
}
# Lags of the model used inside the functions
lagsModel <- matrix(lags,ncol=1);
# If we have a trend add one more lag
if(modelIsTrendy){
lagsModel <- rbind(1,lagsModel);
}
# If we don't have seasonality, remove seasonal lag
if(!modelIsSeasonal && any(lagsModel>1)){
lagsModel <- lagsModel[lagsModel==1,,drop=FALSE];
}
# If this is non-seasonal model and there are no seasonal ARIMA lags, trim the original lags
if(!modelIsSeasonal && all(c(arOrders[lags>1],iOrders[lags>1],maOrders[lags>1])==0) && any(lags>1)){
arOrders <- arOrders[lags==1];
iOrders <- iOrders[lags==1];
maOrders <- maOrders[lags==1];
lags <- lags[lags==1];
}
# Lags of the model
lagsModelSeasonal <- lagsModel[lagsModel>1];
lagsModelMax <- max(lagsModel);
lagsLength <- length(lagsModel);
if(etsModel){
#### Check the seasonal model vs lags ####
if(all(Stype!=c("Z","X","Y","N","A","M","C"))){
warning(paste0("Wrong seasonality type: ",Stype,". Should be 'Z', 'X', 'Y', 'C', 'N', 'A' or 'M'. ",
"Setting to 'Z'."),call.=FALSE);
if(lagsModelMax==1){
Stype <- "N";
modelIsSeasonal <- FALSE;
}
else{
Stype <- "Z";
modelDo <- "select";
}
}
if(all(modelIsSeasonal,lagsModelMax==1)){
if(all(Stype!=c("Z","X","Y"))){
warning(paste0("Cannot build the seasonal model on data with the unity lags.\n",
"Switching to non-seasonal model: ETS(",substr(model,1,nchar(model)-1),"N)"));
}
Stype <- "N";
modelIsSeasonal <- FALSE;
substr(model,nchar(model),nchar(model)) <- "N";
}
# Check the pool of models to combine if it was decided that the data is not seasonal
if(!modelIsSeasonal && !is.null(modelsPool)){
modelsPool <- modelsPool[substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="N"];
}
# Check the type of seasonal
if(Stype!="N"){
componentsNamesETS <- c(componentsNamesETS,"seasonal");
componentsNumberETS[] <- componentsNumberETS+1;
componentsNumberETSSeasonal <- 1;
}
else{
componentsNumberETSSeasonal <- 0;
}
# Check, whether the number of lags and the number of components are the same
if(lagsLength>componentsNumberETS){
if(Stype!="N"){
componentsNamesETS <- c(componentsNamesETS[-length(componentsNamesETS)],
paste0("seasonal",c(1:(lagsLength-componentsNumberETS-1))));
componentsNumberETSSeasonal[] <- lagsLength-componentsNumberETS+1;
componentsNumberETS[] <- lagsLength;
}
else{
lagsModel <- lagsModel[1:componentsNumberETS,,drop=FALSE];
lagsModelMax <- max(lagsModel);
lagsLength <- length(lagsModel);
}
}
else if(lagsLength<componentsNumberETS){
stop("The number of components of the model is smaller than the number of provided lags", call.=FALSE);
}
}
else{
componentsNumberETSSeasonal <- 0;
}
outliers <- match.arg(outliers);
if(outliers!="ignore"){
select <- TRUE;
}
if(!fast){
#### Distribution selected ####
distribution <- match.arg(distribution[1], c("default","dnorm","dlaplace","dalaplace","ds","dgnorm",
"dlnorm","dinvgauss","dgamma"));
}
if(select){
assign("distribution",distribution,ParentEnvironment);
assign("outliers",outliers,ParentEnvironment);
# This stuff is needed for switch to auto.adam.
return(list(select=select));
}
#### Loss function type ####
if(is.function(loss)){
lossFunction <- loss;
loss <- "custom";
multisteps <- FALSE;
}
else{
loss <- match.arg(loss[1],c("likelihood","MSE","MAE","HAM","LASSO","RIDGE",
"MSEh","TMSE","GTMSE","MSCE",
"MAEh","TMAE","GTMAE","MACE",
"HAMh","THAM","GTHAM","CHAM","GPL",
"aMSEh","aTMSE","aGTMSE","aMSCE","aGPL"));
if(any(loss==c("MSEh","TMSE","GTMSE","MSCE","MAEh","TMAE","GTMAE","MACE",
"HAMh","THAM","GTHAM","CHAM","GPL",
"aMSEh","aTMSE","aGTMSE","aMSCE","aGPL"))){
if(!is.null(h) && h>0){
multisteps <- TRUE;
}
else{
stop("The horizon \"h\" needs to be specified and be positive in order for the multistep loss to work.",
call.=FALSE);
multisteps <- FALSE;
}
}
else{
multisteps <- FALSE;
}
lossFunction <- NULL;
}
#### Explanatory variables: xregModel and regressors ####
regressors <- match.arg(regressors,c("use","select","adapt"));
xregModel <- !is.null(xregData);
#### Persistence provided ####
# Vectors for persistence of different components
persistenceLevel <- NULL;
persistenceTrend <- NULL;
persistenceSeasonal <- NULL;
persistenceXreg <- NULL;
# InitialEstimate vectors, defining what needs to be estimated
persistenceEstimate <- persistenceLevelEstimate <- persistenceTrendEstimate <-
persistenceXregEstimate <- TRUE;
# persistence of seasonal is a vector, not a scalar, because we can have several lags
persistenceSeasonalEstimate <- rep(TRUE,componentsNumberETSSeasonal);
if(!is.null(persistence)){
if(all(modelDo!=c("estimate","use"))){
warning(paste0("Predefined persistence can only be used with ",
"preselected ETS model.\n",
"Changing to estimation of persistence values."),call.=FALSE);
persistence <- NULL;
persistenceEstimate <- TRUE;
}
else{
# If it is a list
if(is.list(persistence)){
# If this is a named list, then extract stuff using names
if(!is.null(names(persistence))){
if(!is.null(persistence$level)){
persistenceLevel <- persistence$level;
}
else if(!is.null(persistence$alpha)){
persistenceLevel <- persistence$alpha;
}
if(!is.null(persistence$trend)){
persistenceTrend <- persistence$trend;
}
else if(!is.null(persistence$beta)){
persistenceTrend <- persistence$beta;
}
if(!is.null(persistence$seasonal)){
persistenceSeasonal <- persistence$seasonal;
}
else if(!is.null(persistence$gamma)){
persistenceSeasonal <- persistence$gamma;
}
if(!is.null(persistence$xreg)){
persistenceXreg <- persistence$xreg;
}
else if(!is.null(persistence$delta)){
persistenceXreg <- persistence$delta;
}
}
else{
if(!is.null(persistence[[1]])){
persistenceLevel <- persistence[[1]];
}
if(!is.null(persistence[[2]])){
persistenceTrend <- persistence[[2]];
}
if(!is.null(persistence[[3]])){
persistenceSeasonal <- persistence[[3]];
}
if(!is.null(persistence[[4]])){
persistenceXreg <- persistence[[4]];
}
}
# Define estimate variables
if(!is.null(persistenceLevel)){
persistenceLevelEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
if(!is.null(persistenceTrend)){
persistenceTrendEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
if(!is.null(persistenceSeasonal)){
if(is.list(persistenceSeasonal)){
persistenceSeasonalEstimate[] <- length(persistenceSeasonal)==length(lagsModelSeasonal);
}
else{
persistenceSeasonalEstimate[] <- FALSE;
}
parametersNumber[2,1] <- parametersNumber[2,1] + length(unlist(persistenceSeasonal));
}
if(!is.null(persistenceXreg)){
persistenceXregEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + length(persistenceXreg);
}
}
else if(is.numeric(persistence)){
# If it is smaller... We don't know the length of xreg yet at this stage
if(length(persistence)<lagsLength){
warning(paste0("Length of persistence vector is wrong! ",
"Changing to estimation of persistence vector values."),
call.=FALSE);
persistence <- NULL;
persistenceEstimate <- TRUE;
}
else{
# If there ARIMA elements, remove them
if(any(substr(names(persistence),1,3)=="psi")){
persistence <- persistence[substr(names(persistence),1,3)!="psi"];
}
j <- 0;
if(etsModel){
j <- j+1;
persistenceLevel <- as.vector(persistence)[1];
names(persistenceLevel) <- "alpha";
if(modelIsTrendy && length(persistence)>j){
j <- j+1;
persistenceTrend <- as.vector(persistence)[j];
names(persistenceTrend) <- "beta";
}
if(Stype!="N" && length(persistence)>j){
persistenceSeasonal <- as.vector(persistence)[j+c(1:length(lagsModelSeasonal))];
names(persistenceSeasonal) <- paste0("gamma",c(1:length(persistenceSeasonal)));
j <- j+length(lagsModelSeasonal);
}
}
if(xregModel && length(persistence)>j){
if(j>0){
persistenceXreg <- persistence[-c(1:j)];
}
else{
persistenceXreg <- persistence;
}
# If there are names, make sure that only deltas are used
if(!is.null(names(persistenceXreg))){
persistenceXreg <- persistenceXreg[substr(names(persistenceXreg),1,5)=="delta"];
}
else{
names(persistenceXreg) <- paste0("delta",c(1:length(persistenceXreg)));
}
}
persistenceEstimate[] <- persistenceLevelEstimate[] <- persistenceTrendEstimate[] <-
persistenceXregEstimate[] <- persistenceSeasonalEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + length(persistence);
# bounds <- "none";
}
}
else{
warning(paste0("Persistence is not a numeric vector!\n",
"Changing to estimation of persistence vector values."),call.=FALSE);
persistence <- NULL;
persistenceEstimate <- TRUE;
}
}
}
else{
persistenceEstimate <- TRUE;
}
# Make sure that only important elements are estimated.
if(!etsModel){
persistenceLevelEstimate[] <- FALSE;
}
if(!etsModel || Ttype=="N"){
persistenceTrendEstimate[] <- FALSE;
persistenceTrend <- NULL;
}
if(!etsModel || Stype=="N"){
persistenceSeasonalEstimate[] <- FALSE;
persistenceSeasonal <- NULL;
}
if(!xregModel){
persistenceXregEstimate[] <- FALSE;
persistenceXreg <- NULL;
}
#### Phi ####
if(etsModel){
if(!is.null(phi)){
if(all(modelDo!=c("estimate","use"))){
warning(paste0("Predefined phi can only be used with preselected ETS model.\n",
"Changing to estimation."),call.=FALSE);
phi <- 0.95;
phiEstimate <- TRUE;
}
else{
if(!is.numeric(phi) & (damped)){
warning(paste0("Provided value of phi is meaningless. phi will be estimated."),
call.=FALSE);
phi <- 0.95;
phiEstimate <- TRUE;
}
else if(is.numeric(phi) & (phi<0 | phi>2)){
warning(paste0("Damping parameter should lie in (0, 2) region. ",
"Changing to the estimation of phi."),call.=FALSE);
phi[] <- 0.95;
phiEstimate <- TRUE;
}
else{
phiEstimate <- FALSE;
if(damped){
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
}
}
}
else{
if(damped){
phiEstimate <- TRUE;
phi <- 0.95;
}
else{
phiEstimate <- FALSE;
phi <- 1;
}
}
}
else{
phi <- 1;
phiEstimate <- FALSE;
}
#### Lags for ARIMA ####
if(arimaModel){
lagsModelAll <- rbind(lagsModel,lagsModelARIMA);
lagsModelMax <- max(lagsModel);
}
else{
lagsModelAll <- lagsModel;
}
#### This needs to be amended after developing the first prototype! ####
# If we have the zoo class and weird lags, amend profiles
# Weird lags (dst and fractional): 24, 24*2 (half hour), 52, 24*4 (15 minutes), 7*24, 7*48, 365, 24*52, 24*365
# if(any(yClasses=="zoo") && any(lags %in% c(24, 48, 52, 96, 168, 336, 365, 1248, 8760))){
# For hourly, half-hourly and quarter hour data, just amend the profiles for DST.
# For daily, repeat profile of 28th on 29th February.
# For weekly, repeat the last week, when we have 53 instead of 52.
# }
#### Occurrence variable ####
if(is.occurrence(occurrence)){
oesModel <- occurrence;
occurrence <- oesModel$occurrence;
if(occurrence=="provided"){
occurrenceModelProvided <- FALSE;
}
else{
occurrenceModelProvided <- TRUE;
}
pFitted <- matrix(fitted(oesModel), obsInSample, 1);
}
else{
occurrenceModelProvided <- FALSE;
oesModel <- NULL;
pFitted <- matrix(1, obsInSample, 1);
}
pForecast <- rep(NA,h);
# If it is logical, convert to numeric
if(is.logical(occurrence)){
occurrence <- occurrence*1;
}
if(is.numeric(occurrence)){
# If it is data, then it should correspond to the in-sample.
if(all(occurrence==1)){
occurrence <- "none";
}
else{
if(any(occurrence<0,occurrence>1)){
warning(paste0("Parameter 'occurrence' should contain values between zero and one.\n",
"Converting to appropriate vector."),call.=FALSE);
occurrence[] <- (occurrence!=0)*1;
}
# "provided", meaning that we have been provided the values of p
pFitted[] <- occurrence[1:obsInSample];
# Create forecasted values for occurrence
if(h>0){
if(length(occurrence)>obsInSample){
pForecast <- occurrence[-c(1:obsInSample)];
}
else{
pForecast <- rep(tail(occurrence,1),h);
}
if(length(pForecast)>h){
pForecast <- pForecast[1:h];
}
else if(length(pForecast)<h){
pForecast <- c(pForecast,rep(tail(pForecast,1),h-length(pForecast)));
}
}
else{
pForecast <- NA;
}
occurrence <- "provided";
oesModel <- list(fitted=pFitted,forecast=pForecast,occurrence="provided");
}
}
occurrence <- match.arg(occurrence[1],c("none","auto","fixed","general","odds-ratio",
"inverse-odds-ratio","direct","provided"));
otLogical <- yInSample!=0;
# If the data is not occurrence, let's assume that the parameter was switched unintentionally.
if(all(otLogical) & all(occurrence!=c("none","provided"))){
occurrence <- "none";
occurrenceModelProvided <- FALSE;
}
# If there were NAs and the occurrence was not specified, do something with it
# In all the other cases, NAs will be sorted out by the model
if(any(yNAValues) && (occurrence=="none")){
otLogical <- (!yNAValues)[1:obsInSample];
occurrence[] <- "provided";
pFitted[] <- otLogical*1;
pForecast[] <- 1;
occurrenceModel <- FALSE;
oesModel <- structure(list(y=matrix((otLogical)*1,ncol=1),fitted=pFitted,forecast=pForecast,
occurrence="provided"),class="occurrence");
}
else{
if(occurrence=="none"){
occurrenceModel <- FALSE;
otLogical <- rep(TRUE,obsInSample);
}
else if(occurrence=="provided"){
occurrenceModel <- TRUE;
oesModel$y <- matrix(otLogical*1,ncol=1);
}
else{
occurrenceModel <- TRUE;
}
# Include NAs in the zeroes, so that we avoid weird cases
if(any(yNAValues)){
otLogical <- !(!otLogical | yNAValues[1:obsInSample]);
}
}
if(any(yClasses=="ts")){
ot <- ts(matrix(otLogical*1,ncol=1), start=yStart, frequency=yFrequency);
}
else{
ot <- ts(matrix(otLogical*1,ncol=1), start=c(0,0), frequency=lagsModelMax);
}
obsNonzero <- sum(ot);
obsZero <- obsInSample - obsNonzero;
# If occurrence is provided, use it as is
if(occurrence=="provided"){
ot[] <- pFitted;
}
# Check if multiplicative models can be fitted
allowMultiplicative <- !((any(yInSample<=0) && !occurrenceModel) || (occurrenceModel && any(yInSample<0)));
if(etsModel){
# Clean the pool of models if only additive are allowed
if(!allowMultiplicative && !is.null(modelsPool)){
modelsPoolMultiplicative <- ((substr(modelsPool,1,1)=="M") |
substr(modelsPool,2,2)=="M" |
substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="M");
if(any(modelsPoolMultiplicative)){
modelsPool <- modelsPool[!modelsPoolMultiplicative];
# This is needed, because PPP and FFF use pool, not Branch and bound
if(!any(c(any(unlist(strsplit(model,""))=="P"),any(unlist(strsplit(model,""))=="F")))){
warning("Only additive models are allowed for your data. Amending the pool.",
call.=FALSE);
}
}
}
if((any(model==c("PPP","FFF","YYY")) || any(unlist(strsplit(model,""))=="Z")) && !allowMultiplicative){
model <- "XXX";
Etype[] <- "A";
Ttype[] <- switch(Ttype,"Y"=,"Z"=,"P"=,"F"="X",Ttype);
Stype[] <- switch(Stype,"Y"=,"Z"=,"P"=,"F"="X",Stype);
modelsPool <- NULL;
warning("Only additive models are allowed for your data. Changing the selection mechanism.",
call.=FALSE);
}
else if(!allowMultiplicative && any(c(Etype,Ttype,Stype)=="M")){
warning("Your data contains non-positive values, so the ETS(",model,") might break down.",
call.=FALSE);
}
else if(any(model==c("PPP","FFF")) && allowMultiplicative){
model <- "ZZZ";
}
}
#### Initial values ####
# Vectors for initials of different components
initialLevel <- NULL;
initialTrend <- NULL;
initialSeasonal <- NULL;
initialArima <- NULL;
initialXreg <- NULL;
# InitialEstimate vectors, defining what needs to be estimated
# NOTE: that initial==c("optimal","complete") means initialEstimate==TRUE!
initialEstimate <- initialLevelEstimate <- initialTrendEstimate <-
initialArimaEstimate <- initialXregEstimate <- TRUE;
# initials of seasonal is a vector, not a scalar, because we can have several lags
initialSeasonalEstimate <- rep(TRUE,componentsNumberETSSeasonal);
# This is an initialisation of the variable
initialType <- "optimal"
# initial type can be: "o" - optimal, "b" - backcasting, "p" - provided.
if(any(is.character(initial))){
initialType[] <- match.arg(initial, c("optimal","backcasting","complete"));
}
else if(is.null(initial)){
if(!silent){
message("Initial value is not selected. Switching to optimal.");
}
initialType[] <- "optimal";
}
else if(!is.null(initial)){
if(all(modelDo!=c("estimate","use"))){
warning(paste0("Predefined initials vector can only be used with preselected ETS model.\n",
"Changing to estimation of initials."),call.=FALSE);
initialType[] <- "optimal";
initialEstimate[] <- initialLevelEstimate[] <- initialTrendEstimate[] <-
initialSeasonalEstimate[] <- initialArimaEstimate[] <- initialXregEstimate[] <- TRUE;
}
else{
# If the list is provided, then check what this is.
# This should be: level, trend, seasonal[[1]], seasonal[[2]], ..., ARIMA, xreg
if(is.list(initial)){
# If this is a named list, then extract stuff using names
if(!is.null(names(initial))){
if(!is.null(initial$level)){
initialLevel <- initial$level;
}
if(!is.null(initial$trend)){
initialTrend <- initial$trend;
}
if(!is.null(initial$seasonal)){
initialSeasonal <- initial$seasonal;
}
if(!is.null(initial$arima)){
initialArima <- initial$arima;
}
if(!is.null(initial$xreg)){
initialXreg <- initial$xreg;
}
}
else{
if(!is.null(initial[[1]])){
initialLevel <- initial[[1]];
}
if(!is.null(initial[[2]])){
initialTrend <- initial[[2]];
}
if(!is.null(initial[[3]])){
initialSeasonal <- initial[[3]];
}
if(!is.null(initial[[4]])){
initialArima <- initial[[4]];
}
if(!is.null(initial[[5]])){
initialXreg <- initial[[5]];
}
}
}
else{
if(!is.numeric(initial)){
warning(paste0("Initial vector is not numeric!\n",
"Values of initial vector will be estimated."),call.=FALSE);
initialType[] <- "optimal";
}
else{
# If this is a vector, then it should contain values in the order:
# level, trend, seasonal1, seasonal2, ..., ARIMA, xreg
# if(length(initial)<(sum(lagsModelAll))){
# warning(paste0("The vector of initials contains only values for several components. ",
# "We will use what we can."),call.=FALSE);
# }
# else{
j <- 0;
if(etsModel){
initialLevel <- initial[1];
initialLevelEstimate[] <- FALSE;
if(modelIsTrendy){
j <- 2;
# If there is something in the vector, use it
if(all(!is.na(initial[j]))){
initialTrend <- initial[j];
initialTrendEstimate[] <- FALSE;
}
}
if(Stype!="N"){
# If there is something in the vector, use it
if(length(initial[-c(1:j)])>0){
initialSeasonal <- vector("list",componentsNumberETSSeasonal);
m <- 0;
for(i in 1:componentsNumberETSSeasonal){
if(all(!is.na(initial[j+m+1:lagsModelSeasonal[i]]))){
initialSeasonal[[i]] <- initial[j+m+1:lagsModelSeasonal[i]];
m <- m + lagsModelSeasonal[i];
}
else{
break;
}
}
j <- j+m;
initialSeasonalEstimate[] <- FALSE;
}
}
}
if(arimaModel){
# If there is something else left, this must be ARIMA
if(all(!is.na(initial[j+c(1:initialArimaNumber)]))){
initialArima <- initial[j+c(1:initialArimaNumber)];
j <- j+max(lagsModelARIMA);
initialArimaEstimate[] <- FALSE;
}
}
if(xregModel){
# Something else? xreg for sure!
if(length(initial[-c(1:j)])>0){
initialXreg <- initial[-c(1:j)];
initialXregEstimate[] <- FALSE
}
}
parametersNumber[2,1] <- parametersNumber[2,1] + j;
}
# }
}
}
}
#### Check the provided initials and define initialEstimate variables ####
if(etsModel){
# Level
if(!is.null(initialLevel)){
if(length(initialLevel)>1){
warning("Initial level contains more than one value! Using the first one.",
call.=FALSE);
initialLevel <- initialLevel[1];
}
initialLevelEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
# Trend
if(!is.null(initialTrend)){
if(length(initialTrend)>1){
warning("Initial trend contains more than one value! Using the first one.",
call.=FALSE);
initialTrend <- initialTrend[1];
}
initialTrendEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + 1;
}
# Seasonal
if(!is.null(initialSeasonal)){
# The list means several seasonal lags
if(is.list(initialSeasonal)){
# Is the number of seasonal initials correct? If it is bigger, then remove redundant
if(length(initialSeasonal)>componentsNumberETSSeasonal){
warning("Initial seasonals contained more elements than needed! Removing redundant ones.",
call.=FALSE);
initialSeasonal <- initialSeasonal[1:componentsNumberETSSeasonal];
}
# Is the number of initials in each season correct? Use the correct ones only
if(any(!(sapply(initialSeasonal,length) %in% lagsModelSeasonal))){
warning(paste0("Some of initial seasonals have a wrong length, ",
"not corresponding to the provided lags. I will estimate them."),
call.=FALSE);
initialSeasonalToUse <- sapply(initialSeasonal,length) %in% lagsModelSeasonal;
initialSeasonal <- initialSeasonal[initialSeasonalToUse];
}
initialSeasonalEstimate[] <- !(lagsModelSeasonal %in% sapply(initialSeasonal,length));
# If there are some gaps in what to estimate, reform initialSeason to make sense in the future creator function
if(!all(initialSeasonalEstimate) && !all(!initialSeasonalEstimate)){
initialSeasonalCorrect <- vector("list",componentsNumberETSSeasonal);
initialSeasonalCorrect[which(!initialSeasonalEstimate)] <- initialSeasonal;
initialSeasonal <- initialSeasonalCorrect;
}
}
# The vector implies only one seasonal
else{
if(all(length(initialSeasonal)!=lagsModelSeasonal)){
warning(paste0("Wrong length of seasonal initial: ",length(initialSeasonal),
"Instead of ",lagsModelSeasonal,". Switching to estimation."),
call.=FALSE)
initialSeasonalEstimate[] <- TRUE;
}
else{
initialSeasonalEstimate[] <- FALSE;
}
# Create a list from the vector for consistency purposes
initialSeasonal <- list(initialSeasonal);
}
parametersNumber[2,1] <- parametersNumber[2,1] + length(unlist(initialSeasonal));
}
}
else{
initialLevel <- initialTrend <- initialSeasonal <- NULL;
initialLevelEstimate <- initialTrendEstimate <- initialSeasonalEstimate <- FALSE
# If ETS is switched off, set error to whatever, based on the used distribution
Etype[] <- switch(distribution,
"dinvgauss"=,"dlnorm"=,"dllaplace"=,"dls"=,"dlgnorm"=,"dgamma"="M",
"A");
}
# ARIMA
if(!is.null(initialArima)){
if(length(initialArima)!=initialArimaNumber){
warning(paste0("The length of ARIMA initials is ",length(initialArima),
" instead of ",initialArimaNumber,". Estimating initials instead!"),
call.=FALSE);
initialArimaEstimate[] <- TRUE;
}
else{
initialArimaEstimate[] <- FALSE;
parametersNumber[2,1] <- parametersNumber[2,1] + length(initialArima);
}
}
# xreg
if(!is.null(initialXreg)){
initialXregEstimate[] <- FALSE;
}
#### Check ARIMA parameters, if they are provided ####
if(arimaModel){
# Check the provided parameters for AR and MA
if(!is.null(arma)){
arEstimate <- arRequired;
maEstimate <- maRequired;
# If this is a proper list, extract parameters separately
if(is.list(arma)){
armaParameters <- vector("numeric",sum(sapply(arma,length)));
j <- arIndex <- maIndex <- 0;
# The named list (proper thing)
if(!is.null(names(arma))){
# Check if the length of the provided parameters is correct
if(arRequired && !is.null(arma$ar) && length(arma$ar)!=sum(arOrders)){
warning(paste0("The number of provided AR parameters is ",length(arma$ar),
"while I need ",sum(arOrders),". ",
"Switching to estimation."),call.=FALSE);
arEstimate[] <- TRUE;
}
if(maRequired && !is.null(arma$ma) && length(arma$ma)!=sum(maOrders)){
warning(paste0("The number of provided MA parameters is ",length(arma$ma),
"while I need ",sum(maOrders),". ",
"Switching to estimation."),call.=FALSE);
maEstimate[] <- TRUE;
}
for(i in 1:length(lags)){
if(arRequired && !is.null(arma$ar) && arOrders[i]>0){
armaParameters[j+c(1:arOrders[i])] <- arma$ar[arIndex+c(1:arOrders[i])]
names(armaParameters)[j+c(1:arOrders[i])] <- paste0("phi",1:arOrders[i],"[",lags[i],"]");
j[] <- j+arOrders[i];
arIndex[] <- arIndex+arOrders[i];
arEstimate[] <- FALSE;
}
if(maRequired && !is.null(arma$ma) && maOrders[i]>0){
armaParameters[j+c(1:maOrders[i])] <- arma$ma[maIndex+c(1:maOrders[i])]
names(armaParameters)[j+c(1:maOrders[i])] <- paste0("theta",1:maOrders[i],"[",lags[i],"]");
j[] <- j+maOrders[i];
maIndex[] <- maIndex+maOrders[i];
maEstimate[] <- FALSE;
}
}
}
# Just a list. Assume that the first element is ar, and the second is ma
else{
for(i in 1:length(lags)){
k <- 1;
if(arRequired && !is.null(arma[[1]]) && arOrders[i]>0){
armaParameters[j+c(1:arOrders[i])] <- arma[[1]][arIndex+c(1:arOrders[i])]
names(armaParameters)[j+c(1:arOrders[i])] <- paste0("phi",1:arOrders[i],"[",lags[i],"]");
j[] <- j+arOrders[i];
arIndex[] <- arIndex+arOrders[i];
arEstimate[] <- FALSE;
k[] <- 2;
}
if(maRequired && !is.null(arma[[k]]) && maOrders[i]>0){
armaParameters[j+c(1:maOrders[i])] <- arma[[k]][maIndex+c(1:maOrders[i])]
names(armaParameters)[j+c(1:maOrders[i])] <- paste0("theta",1:maOrders[i],"[",lags[i],"]");
j[] <- j+maOrders[i];
maIndex[] <- maIndex+maOrders[i];
maEstimate[] <- FALSE;
}
}
# Check if the length of the provided parameters is correct
if(length(armaParameters)!=sum(arOrders)+sum(maOrders)){
warning(paste0("The number of provided ARMA parameters is ",length(armaParameters),
"while I need ",sum(arOrders)+sum(maOrders),". ",
"Switching to estimation."),call.=FALSE);
maEstimate <- arEstimate[] <- TRUE;
armaParameters <- NULL;
}
}
}
else if(is.vector(arma)){
# Check the length of the vector
if(length(arma)!=sum(arOrders)+sum(maOrders)){
warning(paste0("The number of provided ARMA parameters is ",length(arma),
"while I need ",sum(arOrders)+sum(maOrders),". ",
"Switching to estimation."),call.=FALSE);
maEstimate <- arEstimate[] <- TRUE;
armaParameters <- NULL;
}
else{
armaParameters <- arma;
j <- 0;
for(i in 1:length(lags)){
if(arOrders[i]>0){
names(armaParameters)[j+1:arOrders[i]] <- paste0("phi",1:arOrders[i],"[",lags[i],"]");
j <- j + arOrders[i];
}
if(maOrders[i]>0){
names(armaParameters)[j+1:maOrders[i]] <- paste0("theta",1:maOrders[i],"[",lags[i],"]");
j <- j + maOrders[i];
}
}
maEstimate <- arEstimate[] <- FALSE;
}
}
parametersNumber[2,1] <- parametersNumber[2,1] + length(armaParameters);
}
else{
arEstimate <- arRequired;
maEstimate <- maRequired;
armaParameters <- NULL;
}
}
else{
armaParameters <- NULL;
}
#### xreg preparation ####
# Check the regressors
if(!xregModel){
regressors[] <- "use";
formulaToUse <- NULL;
}
else{
if(regressors=="select"){
# If this has not happened by chance, then switch to optimisation
if(!is.null(initialXreg) && any(initialType==c("optimal","backcasting"))){
warning("Variables selection does not work with the provided initials for explantory variables. I will drop them.",
call.=FALSE);
initialXreg <- NULL;
initialXregEstimate <- TRUE;
}
if(!is.null(persistenceXreg) && any(persistenceXreg!=0)){
warning(paste0("I cannot do variables selection with the provided smoothing parameters ",
"for explantory variables. I will estimate them instead."),
call.=FALSE);
persistenceXreg <- NULL;
}
}
}
# Use alm() in order to fit the preliminary model for xreg
if(xregModel){
# List of initials. The first element is additive error, the second is the multiplicative one
xregModelInitials <- vector("list",2);
if(regressors!="select"){
#### Initial xreg are not provided ####
# If the initials are not provided, estimate them using ALM.
if(initialXregEstimate){
initialXregProvided <- FALSE;
# The function returns an ALM model
xregInitialiser <- function(Etype,distribution,formulaToUse,subset,responseName){
# Fix the default distribution for ALM
if(distribution=="default"){
distribution <- switch(Etype,
"A"="dnorm",
"M"="dlnorm");
}
else if(distribution=="dllaplace"){
distribution <- "dlaplace";
Etype <- "M";
}
else if(distribution=="dls"){
distribution <- "ds";
Etype <- "M";
}
else if(distribution=="dlgnorm"){
distribution <- "dgnorm";
Etype <- "M";
}
# This is needed to see if trend was asked explicitly. If not, we add it to get rid of bias
trendIncluded <- any(colnames(xregData)[-1]=="trend");
formulaIsAbsent <- is.null(formulaToUse);
formulaOriginal <- formulaToUse;
# If the formula is not provided, construct one
if(formulaIsAbsent){
if(Etype=="M" && any(distribution==c("dnorm","dlaplace","ds","dgnorm","dlogis","dt","dalaplace"))){
if(trendIncluded){
formulaOriginal <- formulaToUse <- as.formula(paste0("log(`",responseName,"`)~."));
}
else{
formulaOriginal <- as.formula(paste0("log(`",responseName,"`)~."));
formulaToUse <- as.formula(paste0("log(`",responseName,"`)~.+trend"));
}
}
else{
if(trendIncluded){
formulaOriginal <- formulaToUse <- as.formula(paste0("`",responseName,"`~."));
}
else{
formulaOriginal <- as.formula(paste0("`",responseName,"`~."));
formulaToUse <- as.formula(paste0("`",responseName,"`~.+trend"));
}
}
}
else{
# If formula only contains ".", then just change it
if(length(all.vars(formulaToUse))==2 && all.vars(formulaToUse)[2]=="."){
# Take logs if the model requires that
if((Etype=="M" && any(distribution==c("dnorm","dlaplace","ds","dgnorm","dlogis","dt","dalaplace")))){
if(trendIncluded){
formulaToUse <- as.formula(paste0("log(`",responseName,"`)~."));
}
else{
formulaToUse <- as.formula(paste0("log(`",responseName,"`)~.+trend"));
}
}
else{
if(trendIncluded){
formulaToUse <- as.formula(paste0(responseName,"~."));
}
else{
formulaToUse <- as.formula(paste0(responseName,"~.+trend"));
}
}
}
else{
trendIncluded <- any(all.vars(formulaToUse)[-1]=="trend");
# If formula contains only one element, or several, but no logs, then change response formula
if((length(formulaToUse[[2]])==1 ||
(length(formulaToUse[[2]])>1 & !any(as.character(formulaToUse[[2]])=="log"))) &&
(Etype=="M" && any(distribution==c("dnorm","dlaplace","ds","dgnorm","dlogis","dt","dalaplace")))){
if(trendIncluded){
formulaToUse <- update(formulaToUse,log(.)~.);
}
else{
formulaToUse <- update(formulaToUse,log(.)~.+trend);
}
}
else{
if(!trendIncluded){
formulaToUse <- update(formulaToUse,.~.+trend);
}
}
}
}
almModel <- do.call(alm,list(formula=formulaToUse,data=xregData,distribution=distribution,subset=which(subset)));
# Remove trend
if(!trendIncluded){
almModel$coefficients <- almModel$coefficients[names(almModel$coefficients)!="trend"];
almModel$data <- almModel$data[,colnames(almModel$data)!="trend",drop=FALSE];
if(formulaIsAbsent){
almModel$call$formula <- as.formula(paste0("`",responseName,"`~."));
}
else{
almModel$call$formula <- formulaOriginal;
}
# Reset terms, they are not needed for what comes, but confuse the formula() method
almModel$terms <- NULL;
}
return(almModel);
}
# Extract names of variables, fix the formula
if(!is.null(formulaToUse)){
formulaToUse <- as.formula(formulaToUse);
responseName <- all.vars(formulaToUse)[1];
# If there are spaces in names, give a warning
if(any(grepl("[^A-Za-z0-9,;._-]", all.vars(formulaToUse))) ||
# If the names only contain numbers
any(suppressWarnings(!is.na(as.numeric(all.vars(formulaToUse)))))){
warning("The names of your variables contain special characters ",
"(such as numbers, spaces, comas, brackets etc). adam() might not work properly. ",
"It is recommended to use `make.names()` function to fix the names of variables.",
call.=FALSE);
formulaToUse <- as.formula(paste0(gsub(paste0("`",all.vars(formulaToUse)[1],"`"),
make.names(all.vars(formulaToUse)[1]),
all.vars(formulaToUse)[1]),"~",
paste0(mapply(gsub, paste0("`",all.vars(formulaToUse)[-1],"`"),
make.names(all.vars(formulaToUse)[-1]),
labels(terms(formulaToUse))),
collapse="+")));
}
formulaProvided <- TRUE;
}
else{
formulaToUse <- reformulate(make.names(colnames(xregData)[-1]), response=responseName);
# Quotes are needed here for the insecure names of variables, such as "1", "2", "3" etc
# formulaToUse <- reformulate(setdiff(paste0("`",colnames(xregData),"`"),
# paste0("`",responseName,"`")),
# response=responseName);
formulaProvided <- FALSE;
}
# Robustify the names of variables
colnames(xregData) <- make.names(colnames(xregData));
# This allows to save the original data
xreg <- xregData;
obsXreg <- nrow(xregData);
# Form subset in order to use in-sample only
subset <- rep(FALSE, obsAll);
subset[1:obsInSample] <- TRUE;
# Exclude zeroes if this is an occurrence model
if(occurrenceModel){
subset[1:obsInSample][!otLogical] <- FALSE;
}
#### Pure regression ####
#### If this is just a regression ALM
if((!etsModel && !arimaModel) && regressors!="adapt"){
# Return the estimated model based on the provided xreg
if(is.null(formulaToUse)){
formulaToUse <- reformulate(setdiff(colnames(xregData), responseName), response=responseName);
# formulaToUse <- as.formula(paste0("`",responseName,"`~."));
formulaProvided <- FALSE;
}
else{
formulaProvided <- TRUE;
}
if(distribution=="default"){
distribution[] <- "dnorm";
}
# Redefine loss for ALM
loss <- switch(loss,
"MSEh"=,"TMSE"=,"GTMSE"=,"MSCE"="MSE",
"MAEh"=,"TMAE"=,"GTMAE"=,"MACE"="MAE",
"HAMh"=,"THAM"=,"GTHAM"=,"CHAM"="HAM",
loss);
lossOriginal <- loss;
if(loss=="custom"){
loss <- lossFunction;
}
# Either use or select the model via greybox functions
# Fisher Information
if(is.null(ellipsis$FI)){
FI <- FALSE;
}
else{
FI <- ellipsis$FI;
}
almModel <- do.call("alm", list(formula=formulaToUse, data=xregData,
distribution=distribution, loss=loss,
subset=which(subset),
occurrence=oesModel,FI=FI));
almModel$call$data <- as.name(yName);
return(almModel);
}
#### ETSX / ARIMAX ####
almModel <- NULL;
if(Etype!="Z"){
almModel <- xregInitialiser(Etype,distribution,formulaToUse,subset,responseName);
# If Intercept was not included, substitute with zero
if(!any(names(almModel$coefficients)=="(Intercept)")){
almIntercept <- 0;
}
else{
almIntercept <- almModel$coefficients["(Intercept)"];
}
if(Etype=="A"){
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregModelInitials[[1]]$initialXreg <- almModel$coefficients;
}
else{
xregModelInitials[[1]]$initialXreg <- almModel$coefficients[-1];
}
if(is.null(formulaToUse)){
xregModelInitials[[1]]$formula <- formulaToUse <- formula(almModel);
}
xregModelInitials[[1]]$other <- almModel$other;
}
else{
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregModelInitials[[2]]$initialXreg <- almModel$coefficients;
}
else{
xregModelInitials[[2]]$initialXreg <- almModel$coefficients[-1];
}
if(is.null(formulaToUse)){
xregModelInitials[[2]]$formula <- formulaToUse <- formula(almModel);
}
xregModelInitials[[2]]$other <- almModel$other;
}
}
# If we are selecting the appropriate error, produce two models: for "M" and for "A"
else{
# Additive model
almModel <- xregInitialiser("A",distribution,formulaToUse,subset,responseName);
# If Intercept was not included, substitute with zero
if(!any(names(almModel$coefficients)=="(Intercept)")){
almIntercept <- 0;
}
else{
almIntercept <- almModel$coefficients["(Intercept)"];
}
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregModelInitials[[1]]$initialXreg <- almModel$coefficients;
}
else{
xregModelInitials[[1]]$initialXreg <- almModel$coefficients[-1];
}
if(is.null(formulaToUse)){
xregModelInitials[[1]]$formula <- formula(almModel);
}
xregModelInitials[[1]]$other <- almModel$other;
# Multiplicative model
almModel[] <- xregInitialiser("M",distribution,formulaToUse,subset,responseName);
# If Intercept was not included, substitute with zero
if(!any(names(almModel$coefficients)=="(Intercept)")){
almIntercept <- 0;
}
else{
almIntercept <- almModel$coefficients["(Intercept)"];
}
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregModelInitials[[2]]$initialXreg <- almModel$coefficients;
}
else{
xregModelInitials[[2]]$initialXreg <- almModel$coefficients[-1];
}
if(is.null(formulaToUse)){
xregModelInitials[[2]]$formula <- formula(almModel);
}
xregModelInitials[[2]]$other <- almModel$other;
}
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
xregNumber <- ncol(almModel$data);
xregNames <- names(coef(almModel));
}
else{
# Write down the number and names of parameters
xregNumber <- ncol(almModel$data)-1;
xregNames <- names(coef(almModel))[-1];
}
# The original number of obs in xreg
obsXreg <- nrow(xreg);
#### Data manipulations for further use ####
# This formula is needed in order to expand the data
if(is.null(formulaToUse)){
formulaToUse <- as.formula(paste0("`",responseName,"`~",
paste0(colnames(xreg)[colnames(xreg)!=responseName],
collapse="+")));
}
# Remove variables without variability
noVariability <- setNames(vector("logical",ncol(xreg)),colnames(xreg));
noVariability[] <- apply((as.matrix(xreg[1:obsInSample,])==matrix(xreg[1,],obsInSample,ncol(xreg),byrow=TRUE)),2,all);
noVariabilityNames <- names(noVariability)[noVariability];
if(any(noVariability) && any(all.vars(formulaToUse) %in% names(noVariability))){
formulaToUse <- update.formula(formulaToUse,paste0(".~.-",paste0(noVariabilityNames,collapse="-")));
}
# Robustify the names of variables
colnames(xreg) <- make.names(colnames(xreg),unique=TRUE);
# The names of the original variables
xregNamesOriginal <- colnames(xregData)[-1];
if((is.matrix(xreg) && any(apply(xreg,2,is.character))) ||
(!is.matrix(xreg) && any(sapply(xreg,is.character)))){
warning("You have character variables in your data. ",
"I will treat them as factors, but it is advised to convert them to factors manually",
call.=FALSE);
}
# If xreg is data frame or formula is provided, do model.frame
additionalManipulations <- (is.data.frame(xreg) || formulaProvided);
if(additionalManipulations){
# Expand the variables. We cannot use alm, because it is based on obsInSample
xregData <- model.frame(formulaToUse,data=as.data.frame(xreg));
}
else{
xregData <- as.matrix(xreg)[,c(names(noVariability)[!noVariability]),drop=FALSE][,-1,drop=FALSE];
xregNumber <- ncol(xregData);
xregNames <- colnames(xregData);
}
# If the number of rows is different, this might be because of NAs
if(additionalManipulations && nrow(xregData)!=nrow(xreg)){
warning("Some variables contained NAs. This might cause issues in the estimation. ",
"I will substitute those values with the first non-NA values",
call.=FALSE);
# Get indices of NAs and nonNAs
xregNAs <- which(is.na(xreg),arr.ind=TRUE);
xregNonNAs <- which(!is.na(xreg),arr.ind=TRUE);
# Go through variables and substitute values
for(i in unique(xregNAs[,2])){
# This split on [row, column] is needed, because data.table is funny with cbind(row,column)
xreg[xregNAs[xregNAs[,2]==i,]] <- xreg[xregNonNAs[which(xregNonNAs[,2]==i)[1],1],
xregNonNAs[which(xregNonNAs[,2]==i)[1],2]];
}
xregData <- model.frame(formulaToUse,data=as.data.frame(xreg));
}
# Get the response variable, just in case it was transformed
if(additionalManipulations && length(formulaToUse[[2]])>1){
y <- xregData[,1];
yInSample <- matrix(y[1:obsInSample],ncol=1);
if(holdout){
yHoldout <- matrix(y[-c(1:obsInSample)],ncol=1);
}
}
#### Drop the variables with no variability and perfectly correlated
if(additionalManipulations){
xregExpanded <- colnames(model.matrix(xregData,data=xregData));
# Remove intercept
if(any(xregExpanded=="(Intercept)")){
xregExpanded <- xregExpanded[-1];
}
if(any(!(xregExpanded %in% xregNames))){
xregNamesRetained <- rep(TRUE,length(xregNamesOriginal));
for(i in 1:length(xregNamesOriginal)){
xregNamesRetained[i] <- any(grepl(xregNamesOriginal[i], xregNames));
}
# If the dropped variables are in the formula, update the formula
if(length(all.vars(formulaToUse))>2 &&
any(all.vars(formulaToUse) %in% xregNamesOriginal[!xregNamesRetained])){
formulaToUse <- update.formula(formulaToUse,
paste0(".~.-",
paste(xregNamesOriginal[!xregNamesRetained],
collapse="-")));
}
xregNamesOriginal <- c(responseName,xregNamesOriginal[xregNamesRetained]);
xregData <- model.frame(formulaToUse,data=as.data.frame(xreg[,xregNamesOriginal,drop=FALSE]));
# Remove response variable
xregNamesOriginal <- xregNamesOriginal[-1]
}
# Binary, flagging factors in the data
xregFactors <- (attr(terms(xregData),"dataClasses")=="factor")[-1];
# Expanded stuff with all levels for factors
if(any(xregFactors)){
# Levels for the factors
xregFactorsLevels <- lapply(xreg,levels);
xregFactorsLevels[[responseName]] <- NULL;
xregModelMatrix <- model.matrix(xregData,xregData,
contrasts.arg=lapply(xregData[attr(terms(xregData),"dataClasses")=="factor"],
contrasts, contrasts=FALSE));
xregNamesModified <- colnames(xregModelMatrix)[-1];
}
else{
xregModelMatrix <- model.matrix(xregData,data=xregData);
xregNamesModified <- xregNames;
}
# Drop the unused variables
xregData <- as.matrix(xregModelMatrix);
# Remove intercept
interceptIsPresent <- FALSE;
if(any(colnames(xregData)=="(Intercept)")){
interceptIsPresent[] <- TRUE;
xregData <- xregData[,-1,drop=FALSE];
}
xregNumber <- ncol(xregData);
# If there are more obs in xreg than the obsAll cut the thing
if(obsXreg>=obsAll){
xregData <- xregData[1:obsAll,,drop=FALSE]
}
# If there are less xreg observations than obsAll, use Naive
else{
newnRows <- obsAll-obsXreg;
xregData <- rbind(xregData,matrix(rep(tail(xregData,1),each=newnRows),newnRows,xregNumber));
}
#### Fix parameters for factors ####
# The indices of the original parameters
xregParametersMissing <- setNames(vector("numeric",xregNumber),xregNamesModified);
# # The indices of the original parameters
xregParametersIncluded <- setNames(vector("numeric",xregNumber),xregNamesModified);
# The vector, marking the same values of smoothing parameters
if(interceptIsPresent){
xregParametersPersistence <- setNames(attr(xregModelMatrix,"assign")[-1],xregNamesModified);
}
else{
xregParametersPersistence <- setNames(attr(xregModelMatrix,"assign"),xregNamesModified);
}
if(length(xregParametersPersistence)==0){
xregParametersPersistence <- 0;
}
# If there are factors not in the alm data, create additional initials
if(any(xregFactors) && any(!(xregNamesModified %in% xregNames))){
xregAbsent <- !(xregNamesModified %in% xregNames);
xregParametersNew <- setNames(rep(NA,xregNumber),xregNamesModified);
# If there is stuff for additive error model, fix parameters
if(!is.null(xregModelInitials[[1]])){
xregParametersNew[!xregAbsent] <- xregModelInitials[[1]]$initialXreg;
# Go through new names and find, where they came from. Then get the missing parameters
for(i in which(xregAbsent)){
# Find the name of the original variable
# Use only the last value... hoping that the names like x and x1 are not used.
xregNameFoundID <- sapply(xregNamesOriginal,grepl,xregNamesModified[i]);
xregNameFound <- tail(names(xregNameFoundID)[xregNameFoundID],1);
# Get the indices of all k-1 levels
xregParametersIncluded[xregNames[xregNames %in% paste0(xregNameFound,
xregFactorsLevels[[xregNameFound]])]] <- i;
# Get the index of the absent one
xregParametersMissing[i] <- i;
# Fill in the absent one, add intercept
xregParametersNew[i] <- almIntercept;
xregParametersNew[xregNamesModified[xregParametersIncluded==i]] <- almIntercept +
xregParametersNew[xregNamesModified[xregParametersIncluded==i]];
# normalise all of them
xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]] <-
xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]] -
mean(xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]]);
}
# Write down the new parameters
xregModelInitials[[1]]$initialXreg <- xregParametersNew;
}
# If there is stuff for multiplicative error model, fix parameters
if(!is.null(xregModelInitials[[2]])){
xregParametersNew[!xregAbsent] <- xregModelInitials[[2]]$initialXreg;
# Go through new names and find, where they came from. Then get the missing parameters
for(i in which(xregAbsent)){
# Find the name of the original variable
# Use only the last value... hoping that the names like x and x1 are not used.
xregNameFoundID <- sapply(xregNamesOriginal,grepl,xregNamesModified[i]);
xregNameFound <- tail(names(xregNameFoundID)[xregNameFoundID],1);
# Get the indices of all k-1 levels
xregParametersIncluded[xregNames[xregNames %in% paste0(xregNameFound,
xregFactorsLevels[[xregNameFound]])]] <- i;
# Get the index of the absent one
xregParametersMissing[i] <- i;
# Fill in the absent one, add intercept
xregParametersNew[i] <- almIntercept;
xregParametersNew[xregNamesModified[xregParametersIncluded==i]] <- almIntercept +
xregParametersNew[xregNamesModified[xregParametersIncluded==i]];
# normalise all of them
xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]] <-
xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]] -
mean(xregParametersNew[xregNamesModified[c(which(xregParametersIncluded==i),i)]]);
}
# Write down the new parameters
xregModelInitials[[2]]$initialXreg <- xregParametersNew;
}
xregNames <- xregNamesModified;
}
# The vector of parameters that should be estimated (numeric + original levels of factors)
xregParametersEstimated <- xregParametersIncluded
xregParametersEstimated[xregParametersEstimated!=0] <- 1;
xregParametersEstimated[xregParametersMissing==0 & xregParametersIncluded==0] <- 1;
}
else{
# If there are more obs in xreg than the obsAll cut the thing
if(obsXreg>=obsAll){
xregData <- xregData[1:obsAll,,drop=FALSE]
}
# If there are less xreg observations than obsAll, use Naive
else{
newnRows <- obsAll-obsXreg;
xregData <- rbind(xregData,matrix(rep(tail(xregData,1),each=newnRows),newnRows,xregNumber));
}
xregFactors <- FALSE;
xregParametersPersistence <- setNames(c(1:xregNumber),xregNames);
xregParametersEstimated <- setNames(rep(1,xregNumber),xregNames);
xregParametersMissing <- setNames(c(1:xregNumber),xregNames);
xregParametersIncluded <- setNames(c(1:xregNumber),xregNames);
}
# Remove xreg, just to preserve some memory
rm(xreg);
# Clean memory if have a big object
if(obsInSample>10000){
gc(verbose=FALSE);
}
}
#### Initial xreg are provided ####
else{
#### Pure regression ####
#### If this is just a regression, then this must be a reuse of alm.
if((!etsModel && !arimaModel) && regressors!="adapt"){
# Return the estimated model based on the provided xreg
if(is.null(formulaToUse)){
formulaToUse <- as.formula(paste0("`",responseName,"`~."));
}
if(distribution=="default"){
distribution[] <- "dnorm";
}
# Redefine loss for ALM
loss <- switch(loss,
"MSEh"=,"TMSE"=,"GTMSE"=,"MSCE"="MSE",
"MAEh"=,"TMAE"=,"GTMAE"=,"MACE"="MAE",
"HAMh"=,"THAM"=,"GTHAM"=,"CHAM"="HAM",
loss);
lossOriginal <- loss;
if(loss=="custom"){
loss <- lossFunction;
}
# Form subset in order to use in-sample only
subset <- rep(FALSE, obsAll);
subset[1:obsInSample] <- TRUE;
# Exclude zeroes if this is an occurrence model
if(occurrenceModel){
subset[1:obsInSample][!otLogical] <- FALSE;
}
# Fisher Information
if(is.null(ellipsis$FI)){
FI <- FALSE;
}
else{
FI <- ellipsis$FI;
}
if(is.null(ellipsis$B)){
B <- initialXreg;
}
else{
B <- ellipsis$B;
}
almModel <- do.call("alm", list(formula=formulaToUse, data=xregData,
distribution=distribution, loss=loss, subset=which(subset),
occurrence=occurrence,FI=FI,
parameters=B,fast=TRUE));
almModel$call$data <- as.name(yName);
return(almModel);
}
#### InitialXreg is provided ####
else{
# Save the original data
xreg <- xregData;
initialXregProvided <- TRUE;
# Robustify the names of variables
colnames(xreg) <- make.names(colnames(xreg),unique=TRUE);
# This formula is needed only to expand xreg
if(is.null(formulaToUse)){
responseName <- colnames(xreg)[1]
formulaToUse <- as.formula(paste0("`",responseName,"`~",
paste0(colnames(xreg)[colnames(xreg)!=responseName],
collapse="+")));
}
# Extract names and form a proper matrix for the regression
else{
responseName <- all.vars(formulaToUse)[1];
}
# Get the names of initials
xregNames <- names(initialXreg);
# The names of the original variables
xregNamesOriginal <- colnames(xregData)[-1];
# Expand the variables. We cannot use alm, because it is based on obsInSample
xregData <- model.frame(formulaToUse,data=as.data.frame(xreg));
# Get the response variable, just in case it was transformed
if(length(formulaToUse[[2]])>1){
y <- xregData[,1];
yInSample <- matrix(y[1:obsInSample],ncol=1);
if(holdout){
yHoldout <- matrix(y[-c(1:obsInSample)],ncol=1);
}
}
# Binary, flagging factors in the data
xregFactors <- (attr(terms(xregData),"dataClasses")=="factor")[-1];
# Expanded stuff with all levels for factors
if(any(xregFactors)){
# Levels for the factors
xregFactorsLevels <- lapply(xreg[,-1,drop=FALSE],levels);
xregModelMatrix <- model.matrix(xregData,xregData,
contrasts.arg=lapply(xregData[attr(terms(xregData),"dataClasses")=="factor"],
contrasts, contrasts=FALSE));
xregNamesModified <- colnames(xregModelMatrix)[-1];
}
else{
xregModelMatrix <- model.matrix(xregData,data=xregData);
xregNamesModified <- xregNames;
}
xregData <- as.matrix(xregModelMatrix);
# Remove intercept
interceptIsPresent <- FALSE;
if(any(colnames(xregData)=="(Intercept)")){
interceptIsPresent[] <- TRUE;
xregData <- xregData[,-1,drop=FALSE];
}
xregNumber <- ncol(xregData);
obsXreg <- nrow(xregData);
# If there are more obs in xreg than the obsAll cut the thing
if(obsXreg>=obsAll){
xregData <- xregData[1:obsAll,,drop=FALSE]
}
# If there are less xreg observations than obsAll, use Naive
else{
newnRows <- obsAll-obsXreg;
xregData <- rbind(xregData,matrix(rep(tail(xregData,1),each=newnRows),newnRows,xregNumber));
}
# This variable is needed in order to do model.matrix only, when required.
xregDataIsDataFrame <- is.data.frame(xregData);
if(xregDataIsDataFrame){
# Expand xregData if it is a data frame
xregData <- model.frame(formulaToUse,data=xregData);
}
xregNumber[] <- ncol(xregData);
#### Fix parameters for factors ####
# The indices of the original parameters
xregParametersMissing <- setNames(vector("numeric",xregNumber),xregNamesModified);
# # The indices of the original parameters
xregParametersIncluded <- setNames(vector("numeric",xregNumber),xregNamesModified);
# The vector, marking the same values of smoothing parameters
if(interceptIsPresent){
xregParametersPersistence <- setNames(attr(xregModelMatrix,"assign")[-1],xregNamesModified);
}
else{
xregParametersPersistence <- setNames(attr(xregModelMatrix,"assign"),xregNamesModified);
}
if(length(xregParametersPersistence)==0){
xregParametersPersistence <- 0;
}
# If there are factors and the number of initials is not the same as the number of parameters needed
# This stuff assumes that the provided xreg parameters are named.
if(any(xregFactors)){
# Expand the data again in order to find the missing elements
xregNames <- colnames(as.matrix(model.matrix(formulaToUse,
model.frame(formulaToUse,data=as.data.frame(xreg)))));
if(any(xregNames=="(Intercept)")){
xregNames <- xregNames[xregNames!="(Intercept)"];
}
if(length(xregNames)!=length(xregNamesModified)){
xregAbsent <- !(xregNamesModified %in% xregNames);
# Create the new vector of parameters
xregParametersNew <- setNames(rep(NA,xregNumber),xregNamesModified);
if(!is.null(names(initialXreg))){
xregParametersNew[xregNames] <- initialXreg[xregNames];
}
else{
# Insert values one by one
j <- 1;
for(i in 1:length(xregNamesModified)){
if(!xregAbsent[i]){
xregParametersNew[i] <- initialXreg[j];
j <- j+1;
}
}
}
# Go through new names and find, where they came from. Then get the missing parameters
for(i in which(xregAbsent)){
# Find the name of the original variable
# Use only the last value... hoping that the names like x and x1 are not used.
xregNameFound <- tail(names(unlist(sapply(xregNamesOriginal,grep,xregNamesModified[i]))),1);
# Get the indices of all k-1 levels
xregParametersIncluded[xregNames[xregNames %in% paste0(xregNameFound,
xregFactorsLevels[[xregNameFound]])]] <- i;
# Get the index of the absent one
xregParametersMissing[i] <- i;
# Fill in the absent one
xregParametersNew[i] <- -sum(xregParametersNew[xregParametersIncluded==i],
na.rm=TRUE);
}
# Write down the new parameters
xregModelInitials[[1]]$initialXreg <- xregParametersNew;
# Write down the new parameters to the second part
if(any(Etype==c("M","Z"))){
xregModelInitials[[2]]$initialXreg <- xregParametersNew;
}
xregNames <- xregNamesModified;
}
}
else{
xregModelInitials[[1]]$initialXreg <- initialXreg;
if(any(Etype==c("M","Z"))){
xregModelInitials[[2]]$initialXreg <- initialXreg;
}
}
# The vector of parameters that should be estimated (numeric + original levels of factors)
xregParametersEstimated <- xregParametersIncluded
xregParametersEstimated[xregParametersEstimated!=0] <- 1;
xregParametersEstimated[xregParametersMissing==0 & xregParametersIncluded==0] <- 1;
parametersNumber[2,2] <- parametersNumber[2,2] + xregNumber;
}
# Remove xreg, just to preserve some memory
rm(xreg);
# Clean memory if have a big object
if(obsInSample>10000){
gc(verbose=FALSE);
}
}
}
else{
#### Pure regression ####
#### If this is just a regression, use stepwise
if((!etsModel && !arimaModel) && regressors!="adapt"){
# Return the estimated model based on the provided xreg
if(is.null(formulaToUse)){
formulaToUse <- reformulate(setdiff(colnames(xregData), responseName), response=responseName);
# formulaToUse <- as.formula(paste0("`",responseName,"`~."));
formulaProvided <- FALSE;
}
else{
formulaProvided <- TRUE;
}
if(distribution=="default"){
distribution[] <- "dnorm";
}
# Form subset in order to use in-sample only
subset <- rep(FALSE, obsAll);
subset[1:obsInSample] <- TRUE;
# Exclude zeroes if this is an occurrence model
if(occurrenceModel){
subset[1:obsInSample][!otLogical] <- FALSE;
}
almModel <- do.call("stepwise", list(data=xregData, formula=formulaToUse, subset=subset,
distribution=distribution, occurrence=oesModel));
almModel$call$data <- as.name(yName);
return(almModel);
}
# Include only variables from the formula
if(is.null(formulaToUse)){
formulaToUse <- as.formula(paste0("`",responseName,"`~."));
}
else{
# Do model.frame manipulations
# We do it this way to avoid factors expansion into dummies at this stage
mf <- as.call(list(quote(stats::model.frame), formula=formulaToUse,
data=xregData, drop.unused.levels=FALSE));
if(!is.data.frame(xregData)){
mf$data <- as.data.frame(xregData);
}
# Evaluate data frame to do transformations of variables
xregData <- eval(mf, parent.frame());
# Remove variables that have "-x" in the formula
dataTerms <- terms(xregData);
xregData <- xregData[,colnames(attr(dataTerms,"factors"))];
}
xregNumber <- ncol(xregData);
xregNames <- colnames(xregData);
initialXregProvided <- FALSE;
xregFactors <- FALSE;
xregParametersPersistence <- setNames(c(1:xregNumber),xregNames);
xregParametersEstimated <- setNames(rep(1,xregNumber),xregNames);
xregParametersMissing <- setNames(c(1:xregNumber),xregNames);
xregParametersIncluded <- setNames(c(1:xregNumber),xregNames);
}
#### persistence for xreg ####
# Process the persistence for xreg
if(!is.null(persistenceXreg)){
if(length(persistenceXreg)!=xregNumber && length(persistenceXreg)!=1){
warning("The length of the provided persistence for the xreg variables is wrong. Reverting to the estimation.",
call.=FALSE);
persistenceXreg <- rep(0.5,xregNumber);
persistenceXregProvided <- FALSE;
persistenceXregEstimate <- TRUE;
}
else if(length(persistenceXreg)==1){
persistenceXreg <- rep(persistenceXreg,xregNumber);
persistenceXregProvided <- TRUE;
persistenceXregEstimate <- FALSE;
}
else{
persistenceXregProvided <- TRUE;
persistenceXregEstimate <- FALSE;
}
}
else{
# The persistenceXregEstimate is provided
if(regressors=="adapt" && !persistenceXregEstimate){
persistenceXregProvided <- TRUE;
persistenceXregEstimate <- FALSE;
}
else if(regressors=="adapt" && persistenceXregEstimate){
persistenceXreg <- rep(0.05,xregNumber);
persistenceXregProvided <- FALSE;
persistenceXregEstimate <- TRUE;
}
else{
persistenceXreg <- rep(0,xregNumber);
persistenceXregProvided <- FALSE;
persistenceXregEstimate <- FALSE;
}
}
# If this is just a regression, include intercept
if(!etsModel && !arimaModel){
lagsModelAll <- matrix(rep(1,xregNumber),ncol=1);
}
else{
lagsModelAll <- matrix(c(lagsModelAll,rep(1,xregNumber)),ncol=1);
}
# If there's only one explanatory variable, then there's nothing to select
if(xregNumber==1){
regressors[] <- "use";
}
# Fix the names of variables
colnames(xregData) <- make.names(colnames(xregData), unique=TRUE);
xregNames[] <- make.names(xregNames, unique=TRUE);
# If there are no variables after all of that, then xreg doesn't exist
if(xregNumber==0){
warning("It looks like there are no suitable explanatory variables. Check the xreg! I dropped them out.",
call.=FALSE);
xregModel[] <- FALSE;
xregData <- NULL;
}
}
else{
initialXregProvided <- FALSE;
initialXregEstimate <- FALSE;
persistenceXregProvided <- FALSE;
persistenceXregEstimate <- FALSE;
xregModelInitials <- NULL;
xregData <- NULL;
xregNumber <- 0;
xregNames <- NULL;
if(is.null(formulaToUse)){
formulaToUse <- as.formula(paste0("`",responseName,"`~."));
}
xregParametersMissing <- 0;
xregParametersIncluded <- 0;
xregParametersEstimated <- 0;
xregParametersPersistence <- 0;
}
# Fix the occurrenceModel for "provided"
if(occurrence=="provided"){
occurrenceModel <- FALSE;
}
# Redefine persitenceEstimate value
persistenceEstimate[] <- any(c(persistenceLevelEstimate,persistenceTrendEstimate,
persistenceSeasonalEstimate,persistenceXregEstimate));
#### Conclusions about the initials ####
# Make sure that only important elements are estimated.
if(!modelIsTrendy){
initialTrendEstimate <- FALSE;
initialTrend <- NULL;
}
if(!modelIsSeasonal){
initialSeasonalEstimate <- FALSE;
initialSeasonal <- NULL;
}
if(!arimaModel){
initialArimaEstimate <- FALSE;
initialArima <- NULL;
}
if(!xregModel){
initialXregEstimate <- FALSE;
initialXreg <- NULL;
}
# If we don't need to estimate anything, flag initialEstimate
if(!any(c(etsModel && initialLevelEstimate,
(etsModel && modelIsTrendy && initialTrendEstimate),
(etsModel & modelIsSeasonal & initialSeasonalEstimate),
(arimaModel && initialArimaEstimate),
(xregModel && initialXregEstimate)))){
initialEstimate[] <- FALSE;
}
else{
initialEstimate[] <- TRUE;
}
# If at least something is provided, flag it as "provided"
if((etsModel && !initialLevelEstimate) ||
(etsModel && modelIsTrendy && !initialTrendEstimate) ||
any(etsModel & modelIsSeasonal & !initialSeasonalEstimate) ||
(arimaModel && !initialArimaEstimate) ||
(xregModel && !initialXregEstimate)){
initialType[] <- "provided";
}
# Observations in the states matrix
# Define the number of cols that should be in the matvt
obsStates <- obsInSample + lagsModelMax;
if(any(yInSample<=0) && any(distribution==c("dinvgauss","dgamma","dlnorm","dllaplace","dls","dlgnorm")) &&
!occurrenceModel && (occurrence!="provided")){
warning(paste0("You have non-positive values in the data. ",
"The distribution ",distribution," does not support that. ",
"This might lead to problems in the estimation."),
call.=FALSE);
}
# Update the number of parameters
if(occurrenceModelProvided){
parametersNumber[2,3] <- nparam(oesModel);
pForecast <- c(forecast(oesModel, h=h, interval="none")$mean);
}
#### Information Criteria ####
ic <- match.arg(ic,c("AICc","AIC","BIC","BICc"));
icFunction <- switch(ic,
"AIC"=AIC,
"AICc"=AICc,
"BIC"=BIC,
"BICc"=BICc);
#### Bounds for the smoothing parameters ####
bounds <- match.arg(bounds,c("usual","admissible","none"));
#### Checks for the potential number of degrees of freedom ####
# This is needed in order to make the function work on small samples
# scale parameter, smoothing parameters and phi
nParamMax <- (1 +
# ETS model
etsModel*(persistenceLevelEstimate + modelIsTrendy*persistenceTrendEstimate +
modelIsSeasonal*sum(persistenceSeasonalEstimate) +
phiEstimate + (initialType=="optimal") *
(initialLevelEstimate + initialTrendEstimate + sum(initialSeasonalEstimate*lagsModelSeasonal))) +
# ARIMA components: initials + parameters
arimaModel*((initialType=="optimal")*initialArimaNumber +
arRequired*arEstimate*sum(arOrders) + maRequired*maEstimate*sum(maOrders)) +
# Xreg initials and smoothing parameters
xregModel*(xregNumber*(any(initialType==c("optimal","backcasting"))*initialXregEstimate+persistenceXregEstimate)));
# If the sample is smaller than the number of parameters
if(obsNonzero <= nParamMax){
# If there is both ETS and ARIMA, remove ARIMA
if(etsModel && arimaModel && !select){
warning("I don't have enough observations to fit ETS with ARIMA terms. I will construct the simple ETS.",
call.=FALSE);
lagsModelAll <- lagsModelAll[-c(componentsNumberETS+c(1:componentsNumberARIMA)),,drop=FALSE];
arRequired <- iRequired <- maRequired <- arimaModel <- FALSE;
arOrders <- iOrders <- maOrders <- NULL;
nonZeroARI <- nonZeroMA <- lagsModelARIMA <- NULL;
componentsNamesARIMA <- NULL;
initialArimaNumber <- componentsNumberARIMA <- 0;
lagsModelMax <- max(lagsModelAll);
}
else if(arimaModel && !etsModel && !select){
# If the backacasting helps, switch to it.
if(initialType=="optimal" && (obsNonzero > (nParamMax - (initialType=="optimal")*initialArimaNumber))){
warning(paste0("The number of parameter to estimate is ",nParamMax,
", while the number of observations is ",obsNonzero,
". Switching initial to 'backcasting' to save some degrees of freedom."), call.=FALSE);
initialType <- "complete";
}
else{
warning(paste0("The number of parameter to estimate is ",nParamMax,
", while the number of observations is ",obsNonzero,
". Switching to ARIMA(0,1,1) model."), call.=FALSE);
# Add ARIMA(0,1,1) lags
lagsModelAll <- matrix(1,1,1);
arRequired <- FALSE;
iRequired <- maRequired <- arimaModel <- TRUE;
arOrders <- 0;
iOrders <- maOrders <- 1;
nonZeroARI <- nonZeroMA <- matrix(c(2,1),1,2);
lagsModelARIMA <- matrix(1,1,1);
componentsNamesARIMA <- 1;
initialArimaNumber <- componentsNumberARIMA <- 1;
lagsModelMax <- max(lagsModelAll);
}
}
}
if(arimaModel && obsNonzero < (initialType=="optimal")*initialArimaNumber && !select){
warning(paste0("In-sample size is ",obsNonzero,", while number of ARIMA components is ",initialArimaNumber,
". Cannot fit the model."),call.=FALSE)
stop("Not enough observations for such a complicated model.",call.=FALSE);
}
# Recalculate the number of parameters
nParamMax[] <- (1 +
# ETS model
etsModel*(persistenceLevelEstimate + modelIsTrendy*persistenceTrendEstimate +
modelIsSeasonal*sum(persistenceSeasonalEstimate) +
phiEstimate + (initialType=="optimal") *
(initialLevelEstimate + initialTrendEstimate + sum(initialSeasonalEstimate*lagsModelSeasonal))) +
# ARIMA components: initials + parameters
arimaModel*((initialType=="optimal")*initialArimaNumber +
arRequired*arEstimate*sum(arOrders) + maRequired*maEstimate*sum(maOrders)) +
# Xreg initials and smoothing parameters
xregModel*(xregNumber*(any(initialType==c("optimal","backcasting"))*initialXregEstimate+persistenceXregEstimate)));
# If the sample is still smaller than the number of parameters (even after removing ARIMA)
if(etsModel){
if(obsNonzero <= nParamMax){
nParamExo <- xregNumber*(initialXregEstimate+persistenceXregEstimate);
if(!silent){
message(paste0("Number of non-zero observations is ",obsNonzero,
", while the maximum number of parameters to estimate is ", nParamMax,".\n",
"Updating pool of models."));
}
# If the number of observations is still enough for the model selection and the pool is not specified
if(obsNonzero > (3 + nParamExo) && is.null(modelsPool) && any(modelDo==c("select","combine"))){
# We have enough observations for local level model
modelsPool <- c("ANN");
if(allowMultiplicative){
modelsPool <- c(modelsPool,"MNN");
}
# We have enough observations for trend model
if(obsNonzero > (5 + nParamExo)){
if(any(Ttype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"AAN");
}
if(allowMultiplicative && any(Ttype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"AMN","MAN","MMN");
}
}
# We have enough observations for damped trend model
if(obsNonzero > (6 + nParamExo)){
if(any(Ttype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"AAdN");
}
if(allowMultiplicative && any(Ttype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"AMdN","MAdN","MMdN");
}
}
# We have enough observations for seasonal model
if((obsNonzero > (2*lagsModelMax)) && lagsModelMax!=1){
if(any(Stype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"ANA");
}
if(allowMultiplicative && any(Stype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"ANM","MNA","MNM");
}
}
# We have enough observations for seasonal model with trend
if((obsNonzero > (6 + lagsModelMax + nParamExo)) &&
(obsNonzero > 2*lagsModelMax) && lagsModelMax!=1){
if(any(Ttype==c("Z","X","A")) && any(Stype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"AAA");
}
if(allowMultiplicative && any(Ttype==c("Z","X","A")) && any(Stype==c("Z","Y","A"))){
modelsPool <- c(modelsPool,"MAA");
}
if(allowMultiplicative && any(Ttype==c("Z","X","A")) && any(Stype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"AAM","MAM");
}
if(allowMultiplicative && any(Ttype==c("Z","Y","M")) && any(Stype==c("Z","X","A"))){
modelsPool <- c(modelsPool,"AMA","MMA");
}
if(allowMultiplicative && any(Ttype==c("Z","Y","M")) && any(Stype==c("Z","Y","M"))){
modelsPool <- c(modelsPool,"AMM","MMM");
}
}
warning("Not enough of non-zero observations for the fit of ETS(",model,")! Fitting what I can...",
call.=FALSE);
if(modelDo=="combine"){
model <- "CNN";
if(length(modelsPool)>2){
model <- "CCN";
}
if(length(modelsPool)>10){
model <- "CCC";
}
}
else{
modelDo <- "select"
model <- "ZZZ";
}
}
# If the pool is provided (so, select / combine), amend it
else if(obsNonzero > (3 + nParamExo) && !is.null(modelsPool)){
# We don't have enough observations for seasonal models with damped trend
if((obsNonzero <= (6 + lagsModelMax + 1 + nParamExo))){
modelsPool <- modelsPool[!(nchar(modelsPool)==4 &
substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="A")];
modelsPool <- modelsPool[!(nchar(modelsPool)==4 &
substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="M")];
}
# We don't have enough observations for seasonal models with trend
if((obsNonzero <= (5 + lagsModelMax + 1 + nParamExo))){
modelsPool <- modelsPool[!(substr(modelsPool,2,2)!="N" &
substr(modelsPool,nchar(modelsPool),nchar(modelsPool))!="N")];
}
# We don't have enough observations for seasonal models
if(obsNonzero <= 2*lagsModelMax){
modelsPool <- modelsPool[substr(modelsPool,nchar(modelsPool),nchar(modelsPool))=="N"];
}
# We don't have enough observations for damped trend
if(obsNonzero <= (6 + nParamExo)){
modelsPool <- modelsPool[nchar(modelsPool)!=4];
}
# We don't have enough observations for any trend
if(obsNonzero <= (5 + nParamExo)){
modelsPool <- modelsPool[substr(modelsPool,2,2)=="N"];
}
modelsPool <- unique(modelsPool);
warning("Not enough of non-zero observations for the fit of ETS(",model,")! Fitting what I can...",
call.=FALSE);
if(modelDo=="combine"){
model <- "CNN";
if(length(modelsPool)>2){
model <- "CCN";
}
if(length(modelsPool)>10){
model <- "CCC";
}
}
else{
modelDo <- "select"
model <- "ZZZ";
}
}
# If the model needs to be estimated / used, not selected
else if(obsNonzero > (3 + nParamExo) && any(modelDo==c("estimate","use"))){
# We don't have enough observations for seasonal models with damped trend
if((obsNonzero <= (6 + lagsModelMax + 1 + nParamExo))){
if(nchar(model)==4){
model <- paste0(substr(model,1,2),substr(model,4,4));
}
# model <- model[!(nchar(model)==4 &
# substr(model,nchar(model),nchar(model))=="A")];
# model <- model[!(nchar(model)==4 &
# substr(model,nchar(model),nchar(model))=="M")];
}
# We don't have enough observations for seasonal models with trend
if((obsNonzero <= (5 + lagsModelMax + 1 + nParamExo))){
model <- paste0(substr(model,1,1),"N",substr(model,3,3));
# model <- model[!(substr(model,2,2)!="N" &
# substr(model,nchar(model),nchar(model))!="N")];
}
# We don't have enough observations for seasonal models
if(obsNonzero <= 2*lagsModelMax){
model <- paste0(substr(model,1,2),"N");
# model <- model[substr(model,nchar(model),nchar(model))=="N"];
}
# We don't have enough observations for damped trend
if(obsNonzero <= (6 + nParamExo)){
if(nchar(model)==4){
model <- paste0(substr(model,1,2),substr(model,4,4));
}
# model <- model[nchar(model)!=4];
}
# We don't have enough observations for any trend
if(obsNonzero <= (5 + nParamExo)){
model <- paste0(substr(model,1,1),"N",substr(model,3,3));
# model <- model[substr(model,2,2)=="N"];
}
}
# Extreme cases of small samples
else if(obsNonzero==4){
if(any(Etype==c("A","M"))){
modelDo <- "estimate";
Ttype <- "N";
Stype <- "N";
}
else{
modelsPool <- c("ANN");
if(allowMultiplicative){
modelsPool <- c(modelsPool,"MNN");
}
modelDo <- "select";
model <- "ZZZ";
Etype <- "Z";
Ttype <- "N";
Stype <- "N";
warning("You have a very small sample. The only available model is level model.",
call.=FALSE);
}
phiEstimate <- FALSE;
}
# Even smaller sample
else if(obsNonzero==3){
if(any(Etype==c("A","M"))){
modelDo <- "estimate";
Ttype <- "N";
Stype <- "N";
model <- paste0(Etype,"NN");
}
else{
modelsPool <- c("ANN");
if(allowMultiplicative){
modelsPool <- c(modelsPool,"MNN");
}
modelDo <- "select";
model <- "ZNN";
Etype <- "Z";
Ttype <- "N";
Stype <- "N";
}
persistenceLevel <- 0;
persistenceEstimate <- persistenceLevelEstimate <- FALSE;
warning("I did not have enough of non-zero observations, so persistence value was set to zero.",
call.=FALSE);
phiEstimate <- FALSE;
}
# Can it be even smaller?
else if(obsNonzero==2){
modelsPool <- NULL;
persistenceLevel <- 0;
persistenceEstimate <- persistenceLevelEstimate <- FALSE;
initialLevel <- mean(yInSample);
initialType <- "provided";
initialEstimate <- initialLevelEstimate <- FALSE;
warning("I did not have enough of non-zero observations, so persistence value was set to zero and initial was preset.",
call.=FALSE);
modelDo <- "use";
model <- "ANN";
Etype <- "A";
Ttype <- "N";
Stype <- "N";
phiEstimate <- FALSE;
parametersNumber[1,1] <- 0;
parametersNumber[2,1] <- 2;
}
# And how about now?!
else if(obsNonzero==1){
modelsPool <- NULL;
persistenceLevel <- 0;
persistenceEstimate <- persistenceLevelEstimate <- FALSE;
initialLevel <- yInSample[yInSample!=0];
initialType <- "provided";
initialEstimate <- initialLevelEstimate <- FALSE;
warning("I did not have enough of non-zero observations, so I used Naive.",call.=FALSE);
modelDo <- "nothing"
model <- "ANN";
Etype <- "A";
Ttype <- "N";
Stype <- "N";
phiEstimate <- FALSE;
parametersNumber[1,1] <- 0;
parametersNumber[2,1] <- 2;
}
# Only zeroes in the data...
else if(obsNonzero==0 && obsInSample>1){
modelsPool <- NULL;
persistenceLevel <- 0;
persistenceEstimate <- persistenceLevelEstimate <- FALSE;
initialLevel <- 0;
initialType <- "provided";
initialEstimate <- initialLevelEstimate <- FALSE;
occurrenceModelProvided <- occurrenceModel <- FALSE;
occurrence <- "none";
warning("You have a sample with zeroes only. Your forecast will be zero.",call.=FALSE);
modelDo <- "nothing"
model <- "ANN";
Etype <- "A";
Ttype <- "N";
Stype <- "N";
phiEstimate <- FALSE;
parametersNumber[1,1] <- 0;
parametersNumber[2,1] <- 2;
}
# If you don't have observations, then fuck off!
else{
stop("Not enough observations... Even for fitting of ETS('ANN')!",call.=FALSE);
}
}
}
# Reset the maximum lag. This is in order to take potential changes into account
lagsModelMax[] <- max(lagsModelAll);
#### Process ellipsis ####
# Parameters for the optimiser
if(is.null(ellipsis$maxeval)){
maxeval <- NULL;
# Make a warning if this is a big computational task
if(any(lags>24) && arimaModel && initialType=="optimal"){
warning(paste0("The estimation of ARIMA model with initial='optimal' on high frequency data might ",
"take more time to converge to the optimum. Consider either setting maxeval parameter ",
"to a higher value (e.g. maxeval=10000, which will take ~25 times more than this) ",
"or using initial='backcasting'."),
call.=FALSE, immediate.=TRUE);
}
}
else{
maxeval <- ellipsis$maxeval;
}
if(is.null(ellipsis$maxtime)){
maxtime <- -1;
}
else{
maxtime <- ellipsis$maxtime;
}
if(is.null(ellipsis$xtol_rel)){
xtol_rel <- 1E-6;
}
else{
xtol_rel <- ellipsis$xtol_rel;
}
if(is.null(ellipsis$xtol_abs)){
xtol_abs <- 1E-8;
}
else{
xtol_abs <- ellipsis$xtol_abs;
}
if(is.null(ellipsis$ftol_rel)){
ftol_rel <- 1E-8;
}
else{
ftol_rel <- ellipsis$ftol_rel;
}
if(is.null(ellipsis$ftol_abs)){
ftol_abs <- 0;
}
else{
ftol_abs <- ellipsis$ftol_abs;
}
if(is.null(ellipsis$algorithm)){
algorithm <- "NLOPT_LN_SBPLX";
}
else{
algorithm <- ellipsis$algorithm;
}
if(is.null(ellipsis$print_level)){
print_level <- 0;
}
else{
print_level <- ellipsis$print_level;
}
# The following three arguments are used for the function itself, not the options
if(is.null(ellipsis$lb)){
lb <- NULL;
}
else{
lb <- ellipsis$lb;
}
if(is.null(ellipsis$ub)){
ub <- NULL;
}
else{
ub <- ellipsis$ub;
}
if(is.null(ellipsis$B)){
B <- NULL;
}
else{
B <- ellipsis$B;
}
# Initialise parameters
lambda <- other <- NULL;
otherParameterEstimate <- FALSE
# lambda for LASSO
if(any(loss==c("LASSO","RIDGE"))){
if(is.null(ellipsis$lambda)){
warning(paste0("You have not provided lambda parameter. I will set it to zero."), call.=FALSE);
lambda <- 0;
}
else{
lambda <- ellipsis$lambda;
}
}
# Parameters for distributions
if(distribution=="dalaplace"){
if(is.null(ellipsis$alpha)){
other <- 0.5
otherParameterEstimate <- TRUE;
}
else{
other <- ellipsis$alpha;
otherParameterEstimate <- FALSE;
}
names(other) <- "alpha";
}
else if(any(distribution==c("dgnorm","dlgnorm"))){
if(is.null(ellipsis$shape)){
other <- 2
otherParameterEstimate <- TRUE;
}
else{
other <- ellipsis$shape;
otherParameterEstimate <- FALSE;
}
names(other) <- "shape";
}
else if(distribution=="dt"){
if(is.null(ellipsis$nu)){
other <- 2
otherParameterEstimate <- TRUE;
}
else{
other <- ellipsis$nu;
otherParameterEstimate <- FALSE;
}
names(other) <- "nu";
}
# Fisher Information
if(is.null(ellipsis$FI)){
FI <- FALSE;
}
else{
FI <- ellipsis$FI;
}
# Step size for the hessian
if(is.null(ellipsis$stepSize)){
stepSize <- .Machine$double.eps^(1/4);
}
else{
stepSize <- ellipsis$stepSize;
}
# Add constant in the model
if(is.numeric(constant)){
constantRequired <- TRUE;
constantEstimate <- FALSE;
# This is just in case a vector was provided
constantValue <- constant[1];
if(any(iOrders!=0) || etsModel){
constantName <- "drift";
}
else{
constantName <- "constant";
}
}
else if(is.logical(constant)){
constantEstimate <- constantRequired <- constant;
constantValue <- NULL;
if(constantRequired){
if(any(iOrders!=0) || etsModel){
constantName <- "drift";
}
else{
constantName <- "constant";
}
}
else{
constantName <- NULL;
}
}
else{
warning("The parameter constant can only be TRUE or FALSE.",
"You have: ",constant,". Switching to FALSE",call.=FALSE);
constantEstimate <- constantRequired <- FALSE;
constantName <- NULL;
}
# If there is no model, return a constant level
if(!etsModel && !arimaModel && !xregModel){
modelsPool <- NULL;
if(!constantRequired){
constantEstimate <- constantRequired <- TRUE;
constantName <- "constant";
}
model <- "NNN";
if(is.null(B)){
modelDo <- "estimate";
}
Etype <- switch(distribution,
"default"=,"dnorm"=,"dlaplace"=,"ds"=,"dgnorm"=,"dlogis"=,"dt"=,"dalaplace"="A",
"dlnorm"=,"dllaplace"=,"dls"=,"dlgnorm"=,"dinvgauss"=,"dgamma"="M");
Ttype <- "N";
Stype <- "N";
phiEstimate <- FALSE;
parametersNumber[1,1] <- 0;
parametersNumber[2,1] <- 2;
}
# See if the estimation of the model is not needed (do we estimate anything?)
if(!any(c(etsModel & c(persistenceLevelEstimate, persistenceTrendEstimate,
persistenceSeasonalEstimate, phiEstimate,
(initialType!="complete") & c(initialLevelEstimate,
initialTrendEstimate,
initialSeasonalEstimate)),
arimaModel & c(arEstimate, maEstimate, (initialType!="complete") & initialArimaEstimate),
xregModel & c(persistenceXregEstimate, (initialType!="complete") & initialXregEstimate),
constantEstimate,
otherParameterEstimate))){
modelDo <- "use";
}
# Switch usual bounds to the admissible if there's no ETS - this speeds up ARIMA
if(!etsModel && bounds=="usual"){
bounds[] <- "admissible";
}
# If we do model selection / combination with non-standard losses, complain
if(any(modelDo==c("select","combine")) &&
((any(loss==c("MSE","MSEh","MSCE","GPL")) && all(distribution!=c("default","dnorm"))) ||
(any(loss==c("MAE","MAEh","MACE")) && all(distribution!=c("default","dlaplace"))) ||
(any(loss==c("HAM","HAMh","CHAM")) && all(distribution!=c("default","ds"))) ||
(any(loss==c("TMSE","GTMSE","TMAE","GTMAE","THAM","GTHAM"))))){
warning("The model selection only works in case of loss='likelihood'. I hope you know what you are doing.",
call.=FALSE);
}
# Just in case, give names to yHoldout and yInSample
colnames(yInSample) <- responseName;
if(holdout){
colnames(yHoldout) <- responseName;
}
#### Return the values to the previous environment ####
### Actuals
assign("y",y,ParentEnvironment);
assign("yHoldout",yHoldout,ParentEnvironment);
assign("yInSample",yInSample,ParentEnvironment);
assign("yNAValues",yNAValues,ParentEnvironment);
### Index and all related structure variables
assign("yClasses",yClasses,ParentEnvironment);
assign("yIndex",yIndex,ParentEnvironment);
assign("yInSampleIndex",yInSampleIndex,ParentEnvironment);
assign("yForecastIndex",yForecastIndex,ParentEnvironment);
assign("yIndexAll",yIndexAll,ParentEnvironment);
assign("yFrequency",yFrequency,ParentEnvironment);
assign("yStart",yStart,ParentEnvironment);
assign("yForecastStart",yForecastStart,ParentEnvironment);
# The rename of the variable is needed for the hessian to work
assign("horizon",h,ParentEnvironment);
assign("h",h,ParentEnvironment);
assign("holdout",holdout,ParentEnvironment);
### Number of observations and parameters
assign("obsInSample",obsInSample,ParentEnvironment);
assign("obsAll",obsAll,ParentEnvironment);
assign("obsStates",obsStates,ParentEnvironment);
assign("obsNonzero",obsNonzero,ParentEnvironment);
assign("obsZero",obsZero,ParentEnvironment);
assign("parametersNumber",parametersNumber,ParentEnvironment);
### Model type
assign("etsModel",etsModel,ParentEnvironment);
assign("model",model,ParentEnvironment);
assign("Etype",Etype,ParentEnvironment);
assign("Ttype",Ttype,ParentEnvironment);
assign("Stype",Stype,ParentEnvironment);
assign("modelIsTrendy",modelIsTrendy,ParentEnvironment);
assign("modelIsSeasonal",modelIsSeasonal,ParentEnvironment);
assign("modelsPool",modelsPool,ParentEnvironment);
assign("damped",damped,ParentEnvironment);
assign("modelDo",modelDo,ParentEnvironment);
assign("allowMultiplicative",allowMultiplicative,ParentEnvironment);
### Numbers and names of components
assign("componentsNumberETS",componentsNumberETS,ParentEnvironment);
assign("componentsNamesETS",componentsNamesETS,ParentEnvironment);
assign("componentsNumberETSNonSeasonal",componentsNumberETS-componentsNumberETSSeasonal,ParentEnvironment);
assign("componentsNumberETSSeasonal",componentsNumberETSSeasonal,ParentEnvironment);
# The number and names of ARIMA components
assign("componentsNumberARIMA",componentsNumberARIMA,ParentEnvironment);
assign("componentsNamesARIMA",componentsNamesARIMA,ParentEnvironment);
### Lags
# This is the original vector of lags, modified for the level components.
# This can be used in ARIMA
assign("lags",lags,ParentEnvironment);
# This is the vector of lags of ETS components
assign("lagsModel",lagsModel,ParentEnvironment);
# This is the vector of seasonal lags
assign("lagsModelSeasonal",lagsModelSeasonal,ParentEnvironment);
# This is the vector of lags for ARIMA components (not lags of ARIMA)
assign("lagsModelARIMA",lagsModelARIMA,ParentEnvironment);
# This is the vector of all the lags of model (ETS + ARIMA + X)
assign("lagsModelAll",lagsModelAll,ParentEnvironment);
# This is the maximum lag
assign("lagsModelMax",lagsModelMax,ParentEnvironment);
### Persistence
assign("persistence",persistence,ParentEnvironment);
assign("persistenceEstimate",persistenceEstimate,ParentEnvironment);
assign("persistenceLevel",persistenceLevel,ParentEnvironment);
assign("persistenceLevelEstimate",persistenceLevelEstimate,ParentEnvironment);
assign("persistenceTrend",persistenceTrend,ParentEnvironment);
assign("persistenceTrendEstimate",persistenceTrendEstimate,ParentEnvironment);
assign("persistenceSeasonal",persistenceSeasonal,ParentEnvironment);
assign("persistenceSeasonalEstimate",persistenceSeasonalEstimate,ParentEnvironment);
assign("persistenceXreg",persistenceXreg,ParentEnvironment);
assign("persistenceXregEstimate",persistenceXregEstimate,ParentEnvironment);
assign("persistenceXregProvided",persistenceXregProvided,ParentEnvironment);
### phi
assign("phi",phi,ParentEnvironment);
assign("phiEstimate",phiEstimate,ParentEnvironment);
### Initials
assign("initial",initial,ParentEnvironment);
assign("initialType",initialType,ParentEnvironment);
assign("initialEstimate",initialEstimate,ParentEnvironment);
assign("initialLevel",initialLevel,ParentEnvironment);
assign("initialLevelEstimate",initialLevelEstimate,ParentEnvironment);
assign("initialTrend",initialTrend,ParentEnvironment);
assign("initialTrendEstimate",initialTrendEstimate,ParentEnvironment);
assign("initialSeasonal",initialSeasonal,ParentEnvironment);
assign("initialSeasonalEstimate",initialSeasonalEstimate,ParentEnvironment);
assign("initialArima",initialArima,ParentEnvironment);
assign("initialArimaEstimate",initialArimaEstimate,ParentEnvironment);
# Number of initials that the ARIMA has (either provided or to estimate)
assign("initialArimaNumber",initialArimaNumber,ParentEnvironment);
assign("initialXregEstimate",initialXregEstimate,ParentEnvironment);
assign("initialXregProvided",initialXregProvided,ParentEnvironment);
### Occurrence model
assign("oesModel",oesModel,ParentEnvironment);
assign("occurrenceModel",occurrenceModel,ParentEnvironment);
assign("occurrenceModelProvided",occurrenceModelProvided,ParentEnvironment);
assign("occurrence",occurrence,ParentEnvironment);
assign("pFitted",pFitted,ParentEnvironment);
assign("pForecast",pForecast,ParentEnvironment);
assign("ot",ot,ParentEnvironment);
assign("otLogical",otLogical,ParentEnvironment);
### Outliers detection
assign("outliers",outliers,ParentEnvironment);
### Distribution, loss, bounds and IC
assign("distribution",distribution,ParentEnvironment);
assign("loss",loss,ParentEnvironment);
assign("lossFunction",lossFunction,ParentEnvironment);
assign("multisteps",multisteps,ParentEnvironment);
assign("ic",ic,ParentEnvironment);
assign("icFunction",icFunction,ParentEnvironment);
assign("bounds",bounds,ParentEnvironment);
### ARIMA components
assign("arimaModel",arimaModel,ParentEnvironment);
assign("arOrders",arOrders,ParentEnvironment);
assign("iOrders",iOrders,ParentEnvironment);
assign("maOrders",maOrders,ParentEnvironment);
assign("arRequired",arRequired,ParentEnvironment);
assign("iRequired",iRequired,ParentEnvironment);
assign("maRequired",maRequired,ParentEnvironment);
assign("arEstimate",arEstimate,ParentEnvironment);
assign("maEstimate",maEstimate,ParentEnvironment);
assign("armaParameters",armaParameters,ParentEnvironment);
assign("nonZeroARI",nonZeroARI,ParentEnvironment);
assign("nonZeroMA",nonZeroMA,ParentEnvironment);
assign("select",select,ParentEnvironment);
### Explanatory variables
assign("xregModel",xregModel,ParentEnvironment);
assign("regressors",regressors,ParentEnvironment);
assign("xregModelInitials",xregModelInitials,ParentEnvironment);
assign("xregData",xregData,ParentEnvironment);
assign("xregNumber",xregNumber,ParentEnvironment);
assign("xregNames",xregNames,ParentEnvironment);
assign("responseName",responseName,ParentEnvironment);
assign("formula",formulaToUse,ParentEnvironment);
assign("xregParametersMissing",xregParametersMissing,ParentEnvironment);
assign("xregParametersIncluded",xregParametersIncluded,ParentEnvironment);
assign("xregParametersEstimated",xregParametersEstimated,ParentEnvironment);
assign("xregParametersPersistence",xregParametersPersistence,ParentEnvironment);
### Constant
assign("constantRequired",constantRequired,ParentEnvironment);
assign("constantEstimate",constantEstimate,ParentEnvironment);
assign("constantValue",constantValue,ParentEnvironment);
assign("constantName",constantName,ParentEnvironment);
### Ellipsis thingies
# Optimisation related
assign("maxeval",maxeval,ParentEnvironment);
assign("maxtime",maxtime,ParentEnvironment);
assign("xtol_rel",xtol_rel,ParentEnvironment);
assign("xtol_abs",xtol_abs,ParentEnvironment);
assign("ftol_rel",ftol_rel,ParentEnvironment);
assign("ftol_abs",ftol_abs,ParentEnvironment);
assign("algorithm",algorithm,ParentEnvironment);
assign("print_level",print_level,ParentEnvironment);
assign("B",B,ParentEnvironment);
assign("lb",lb,ParentEnvironment);
assign("ub",ub,ParentEnvironment);
# Parameters for distributions
assign("other",other,ParentEnvironment);
assign("otherParameterEstimate",otherParameterEstimate,ParentEnvironment);
# LASSO / RIDGE
assign("lambda",lambda,ParentEnvironment);
# Fisher Information
assign("FI",FI,ParentEnvironment);
# Step size for the hessian
assign("stepSize",stepSize,ParentEnvironment);
return(list(select=FALSE));
}
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.