FcDecompositionModelByRegression: Forecasting with Automatic Decomposition Model By Regression

In Mthrun/TSAT: Time Series Analysis Tools (TSAT)

Forecasting with Automatic Decomposition Model By Regression

Description

Automatic approach for additive of multiplicative decomposition model generation. Dataset is divided into training and test data by SplitDataAt.

Usage

FcDecompositionModelByRegression(DataFrame, TimeColumnName = "Time", 

FeatureName ="Absatz", SplitDataAt, Frequency = "day",

                 ForecastPeriods = 10, Holidays = NULL, PlotIt = TRUE,
                 
                 xlab = "Time", ylab = "Feature", EquiDist = TRUE,
                 
                 MinLowerBound = NULL, MaxUpperBound = NULL, ...)

Arguments

DataFrame	data frame of at least two columns of Time and a Feauture, see example
TimeColumnName	name of time column in data frame, see example
FeatureName	Feature name of variable which should be forecasted, see example
SplitDataAt	index of row where the DataFrame is divided into test and train data, see example
Frequency	'day', 'week', 'month', 'quarter', 'year', 1(1 sec), 60(1 minute) or 3600(1 hour)
ForecastPeriods	how many days, weeks, etc,, should the model be foracested on?
Holidays	data frame of time and dates, see example
PlotIt	Plots data model and prediction with black dots: data dark green line: model based on training data red dots: prediction blue range: confidence intervall of model
xlab	see ggplot2
ylab	see ggplot2
EquiDist	TRUE: time difference between each two points is equidisant, if not set FALSE
MinLowerBound	due to bug from prophet does not work currently
MaxUpperBound	due to bug from prophet does not work currently
...	model paramters, see example

Details

training data is used for model generation, test data is used for evaluating the quality by a comparision to given data.

Value

Forecast	Forecast, see example
TestData	TestData, see example
TrainingData	TrainingData, see example
Model	model paramters, see example
ggObject	ggplot object to change plot with black dots: data dark green line: model based on training data red dots: prediction blue range: confidence intervall of model

Note

Usually, a model should also have a third part of data only used once and for manual verification.

Author(s)

Michael Thrun for the easy-to-use wrapper, algorithm of prophet is used (see url)

References

About the package behind it https://facebook.github.io/prophet/

Examplary published approach of an (manual) compund model:

Aubert, A. H., Thrun, M. C., Breuer, L., Ultsch, A.: Knowledge discovery from high-frequency stream nitrate

concentrations: hydrology and biology contributions, Scientific Reports, Nature, Vol. 6(31536), pp. 1-8, 2016.

Examples

dates=c("2015-01-01","2015-01-02")
names=c("New Year's Day","New Year's weekend")
holidays=data.frame(ds=as.Date(dates),holiday=names
)
data('WeeklySalesUniformRandom')


##Forcasting 10 weeks of 166 weeks data of sales

weeklyres=FcDecompositionModelByRegression(DataFrame = WeeklySalesUniformRandom,TimeColumnName ='Time' ,

FeatureName = 'Sales',SplitDataAt = 156,Frequency = 'week',

ForecastPeriods = 10,PlotIt = T,

Holidays = holidays)


#Plot components
prophet::plot_forecast_component(weeklyres$Forecast,name = c('seasonal'))
prophet::plot_forecast_component(weeklyres$Forecast,name = c('trend'))
#further compoents in names(weeklyres$Forecast)



###Forcasting thirdy days of 1093 days sales data frame
##Grid Search fo Best Parameters

#Generate all combinations of relevant paramters
prophetGrid <- expand.grid(changepoint_prior_scale = c(0.05, 0.5, 0.001),
                           seasonality_prior_scale = c(150,100,50, 10, 1,0.1),
                           holidays_prior_scale = c(150,100,50, 10, 1,0.1),
                           daily_seasonality=c(TRUE,FALSE)
)

results <- matrix(NaN, ncol=2,nrow = nrow(prophetGrid))
#calculate all
for (i in seq_len(nrow(prophetGrid))) {
  parameters <- prophetGrid[i, ]
  temp=FcDecompositionModelByRegression(
    DataFrame = DailySales,TimeColumnName ='Time' ,
    FeatureName = 'Sales',SplitDataAt = 1093,Frequency = 'day',
    ForecastPeriods = 30,PlotIt = F,Holidays = holidays2,
    seasonality.prior.scale = parameters$seasonality_prior_scale, 
    changepoint.prior.scale = parameters$changepoint_prior_scale,
    holidays.prior.scale = parameters$holidays_prior_scale,
    daily.seasonality=parameters$daily_seasonality
  )
  results[i,]=temp$Accuracy[1,2:3]
}
# choose rmse or mae
plot(results)
#rmse chosen
prophetGrid <- cbind(prophetGrid, results[,1])
#best paramters found
DailyParameters <- prophetGrid3[prophetGrid3results == min(results[,1]), ]

dailyres=FcDecompositionModelByRegression(DataFrame = DailySales,TimeColumnName ='Time',

FeatureName = 'Sales',SplitDataAt = 1093,Frequency = 'day',

ForecastPeriods = 30,PlotIt = T,Holidays = holidays2,

seasonality.prior.scale = DailyParameters$seasonality_prior_scale, 

changepoint.prior.scale = DailyParameters$changepoint_prior_scale,

holidays.prior.scale = DailyParameters$holidays_prior_scale,

daily.seasonality=DailyParameters$daily_seasonality)

dailyres$Accuracy

1-abs(mean(dailyres$TestData$y) - dailyres$Accuracy[2,3])/dailyres$Accuracy[2,3]


##multiplikative

data(AirPassengers)
DF=ConvertTS2DF(AirPassengers)
model=FcDecompositionModelByRegression(DataFrame = DF,TimeColumnName ='Time' ,

FeatureName = 'Data',SplitDataAt = 120,Frequency = 'month',

ForecastPeriods = 24,PlotIt = T,seasonality.mode = 'multiplicative')

Mthrun/TSAT documentation built on Feb. 5, 2024, 11:15 p.m.