es: Exponential Smoothing in SSOE state space model

Description Usage Arguments Details Value Author(s) References See Also Examples

View source: R/es.R


Function constructs ETS model and returns forecast, fitted values, errors and matrix of states.


es(y, model = "ZZZ", persistence = NULL, phi = NULL,
  initial = c("optimal", "backcasting"), initialSeason = NULL,
  ic = c("AICc", "AIC", "BIC", "BICc"), loss = c("likelihood", "MSE",
  "MAE", "HAM", "MSEh", "TMSE", "GTMSE", "MSCE"), h = 10, holdout = FALSE,
  cumulative = FALSE, interval = c("none", "parametric", "likelihood",
  "semiparametric", "nonparametric"), level = 0.95, bounds = c("usual",
  "admissible", "none"), silent = c("all", "graph", "legend", "output",
  "none"), xreg = NULL, xregDo = c("use", "select"), initialX = NULL,



Vector or ts object, containing data needed to be forecasted.


The type of ETS model. The first letter stands for the type of the error term ("A" or "M"), the second (and sometimes the third as well) is for the trend ("N", "A", "Ad", "M" or "Md"), and the last one is for the type of seasonality ("N", "A" or "M"). So, the function accepts words with 3 or 4 characters: ANN, AAN, AAdN, AAA, AAdA, MAdM etc. ZZZ means that the model will be selected based on the chosen information criteria type. Models pool can be restricted with additive only components. This is done via model="XXX". For example, making selection between models with none / additive / damped additive trend component only (i.e. excluding multiplicative trend) can be done with model="ZXZ". Furthermore, selection between multiplicative models (excluding additive components) is regulated using model="YYY". This can be useful for positive data with low values (for example, slow moving products). Finally, if model="CCC", then all the models are estimated and combination of their forecasts using AIC weights is produced (Kolassa, 2011). This can also be regulated. For example, model="CCN" will combine forecasts of all non-seasonal models and model="CXY" will combine forecasts of all the models with non-multiplicative trend and non-additive seasonality with either additive or multiplicative error. Not sure why anyone would need this thing, but it is available.

The parameter model can also be a vector of names of models for a finer tuning (pool of models). For example, model=c("ANN","AAA") will estimate only two models and select the best of them.

Also model can accept a previously estimated ES or ETS (from forecast package) model and use all its parameters.

Keep in mind that model selection with "Z" components uses Branch and Bound algorithm and may skip some models that could have slightly smaller information criteria.


Persistence vector g, containing smoothing parameters. If NULL, then estimated.


Value of damping parameter. If NULL then it is estimated.


Can be either character or a vector of initial states. If it is character, then it can be "optimal", meaning that the initial states are optimised, or "backcasting", meaning that the initials are produced using backcasting procedure (advised for data with high frequency). If character, then initialSeason will be estimated in the way defined by initial.


Vector of initial values for seasonal components. If NULL, they are estimated during optimisation.


The information criterion used in the model selection procedure.


The type of Loss Function used in optimization. loss can be: likelihood (assuming Normal distribution of error term), MSE (Mean Squared Error), MAE (Mean Absolute Error), HAM (Half Absolute Moment), TMSE - Trace Mean Squared Error, GTMSE - Geometric Trace Mean Squared Error, MSEh - optimisation using only h-steps ahead error, MSCE - Mean Squared Cumulative Error. If loss!="MSE", then likelihood and model selection is done based on equivalent MSE. Model selection in this cases becomes not optimal.

There are also available analytical approximations for multistep functions: aMSEh, aTMSE and aGTMSE. These can be useful in cases of small samples.

Finally, just for fun the absolute and half analogues of multistep estimators are available: MAEh, TMAE, GTMAE, MACE, TMAE, HAMh, THAM, GTHAM, CHAM.


Length of forecasting horizon.


If TRUE, holdout sample of size h is taken from the end of the data.


If TRUE, then the cumulative forecast and prediction interval are produced instead of the normal ones. This is useful for inventory control systems.


Type of interval to construct. This can be:

  • "none", aka "n" - do not produce prediction interval.

  • "parametric", "p" - use state-space structure of ETS. In case of mixed models this is done using simulations, which may take longer time than for the pure additive and pure multiplicative models. This type of interval relies on unbiased estimate of in-sample error variance, which divides the sume of squared errors by T-k rather than just T.

  • "likelihood", "l" - these are the same as "p", but relies on the biased estimate of variance from the likelihood (division by T, not by T-k).

  • "semiparametric", "sp" - interval based on covariance matrix of 1 to h steps ahead errors and assumption of normal / log-normal distribution (depending on error type).

  • "nonparametric", "np" - interval based on values from a quantile regression on error matrix (see Taylor and Bunn, 1999). The model used in this process is e[j] = a j^b, where j=1,..,h.

The parameter also accepts TRUE and FALSE. The former means that parametric interval are constructed, while the latter is equivalent to none. If the forecasts of the models were combined, then the interval are combined quantile-wise (Lichtendahl et al., 2013).


Confidence level. Defines width of prediction interval.


What type of bounds to use in the model estimation. The first letter can be used instead of the whole word.


If silent="none", then nothing is silent, everything is printed out and drawn. silent="all" means that nothing is produced or drawn (except for warnings). In case of silent="graph", no graph is produced. If silent="legend", then legend of the graph is skipped. And finally silent="output" means that nothing is printed out in the console, but the graph is produced. silent also accepts TRUE and FALSE. In this case silent=TRUE is equivalent to silent="all", while silent=FALSE is equivalent to silent="none". The parameter also accepts first letter of words ("n", "a", "g", "l", "o").


The vector (either numeric or time series) or the matrix (or data.frame) of exogenous variables that should be included in the model. If matrix included than columns should contain variables and rows - observations. Note that xreg should have number of observations equal either to in-sample or to the whole series. If the number of observations in xreg is equal to in-sample, then values for the holdout sample are produced using es function.


The variable defines what to do with the provided xreg: "use" means that all of the data should be used, while "select" means that a selection using ic should be done. "combine" will be available at some point in future...


The vector of initial parameters for exogenous variables. Ignored if xreg is NULL.


Other non-documented parameters. For example FI=TRUE will make the function also produce Fisher Information matrix, which then can be used to calculated variances of smoothing parameters and initial states of the model. Parameters B, lb and ub can be passed via ellipsis as well. In this case they will be used for optimisation. B sets the initial values before the optimisation, lb and ub define lower and upper bounds for the search inside of the specified bounds. These values should have length equal to the number of parameters to estimate. You can also pass two parameters to the optimiser: 1. maxeval - maximum number of evaluations to carry on; 2. xtol_rel - the precision of the optimiser. The default values used in es() are maxeval=500 and xtol_rel=1e-8. You can read more about these parameters in the documentation of nloptr function.


Function estimates ETS in a form of the Single Source of Error state space model of the following type:

y_{t} = o_t (w(v_{t-l}) + h(x_t, a_{t-1}) + r(v_{t-l}) ε_{t})

v_{t} = f(v_{t-l}) + g(v_{t-l}) ε_{t}

a_{t} = F_{X} a_{t-1} + g_{X} ε_{t} / x_{t}

Where o_{t} is the Bernoulli distributed random variable (in case of normal data it equals to 1 for all observations), v_{t} is the state vector and l is the vector of lags, x_t is the vector of exogenous variables. w(.) is the measurement function, r(.) is the error function, f(.) is the transition function, g(.) is the persistence function and h(.) is the explanatory variables function. a_t is the vector of parameters for exogenous variables, F_{X} is the transitionX matrix and g_{X} is the persistenceX matrix. Finally, ε_{t} is the error term.

For the details see Hyndman et al.(2008).

For some more information about the model and its implementation, see the vignette: vignette("es","smooth").

Also, there are posts about the functions of the package smooth on the website of Ivan Svetunkov: - they explain the underlying models and how to use the functions.


Object of class "smooth" is returned. It contains the list of the following values for classical ETS models:

If combination of forecasts is produced (using model="CCC"), then a shorter list of values is returned:


Ivan Svetunkov,


See Also

adam, forecast, ts,


# See how holdout and trace parameters influence the forecast

# Model selection example

# Model selection. Compare AICc of these two models:

# Model selection, excluding multiplicative trend

# Combination example

# Model selection using a specified pool of models
ourModel <- es(Mcomp::M3$N1587$x,model=c("ANN","AAM","AMdA"),h=18)

# Redo previous model and produce prediction interval

# Semiparametric interval example

# This will be the same model as in previous line but estimated on new portion of data

config-i1/smooth documentation built on June 16, 2021, 2:13 p.m.