Comparing Holt Winters Implementations in R – Part 1

About 30% of the available data is help for testing and we use the initial 70% of the data to build our Holt Winters models. Both the function HoltWinters() and hw() try to optimize the alpha, beta and gamma values by minimizing the residuals but both does not give the same values. Hence the differing fits and predictions. The output of the following code stores the forecasting charts in your working directory. Each chart shows the actual, fit and forecast with 95% confidence drawn in orange lines. The accuracy measures that gets printed on the console are computed for the training data and the MAPE is for the predictions made by the respective models.

Holt Winters - hw()

Forecast Projections of Holt Winters Implementation – hw {forecast}


Holt Winters Projections from base 'stats' Package

Holt Winters Projections from base ‘stats’ Package

Just by looking at the fitted curve of the training data, one can argue that HoltWinters() performed equally well if not better than hw(). But when you look at forecasted values, clearly in this case, the predictions of hw() performed better than HoltWinters(). Lets look at the accuracy measures..

Compare accuracy of HoltWinters() vs hw()

Compare accuracy of HoltWinters() vs hw()

Interestingly, the MAPE of fitted values was slightly better for HoltWinters() over hw(), but the MAPE when calculated for forecast vs actual, the hw() function of ‘forecast’ package clearly outperforms. 

pigs <- dmseries("")
data <- pigs
data.ts <- as.ts(data)
#plot(data.ts, type = 'b', col = "blue", main = colnames(data.ts))
#Training and Test Split
split 	        <- ceiling(0.7 * length(data))
dataTrain 	<- ts(data[1:split], frequency = 12, start = c(1980, 1))
dataTest 	<- ts(data[c((split+1) : nrow(data))], frequency = 12, start = c(1991, 1))
actual    	<- unclass(dataTrain)
actualFull 	<- unclass(data)
#metd values => HoltWinters, hw
forecastGen <- function(metd){
			if(metd == "HoltWinters"){
				method    	<- "HoltWinters {stats}"

				mod 		<- HoltWinters(dataTrain, optim.start = c(alpha = 0.99, beta = 0.001, gamma =0.001)) 			
				model 	        <- forecast.HoltWinters(mod, h=(length(data)-length(dataTrain)))				
				actualFull 	<- actualFull[13:length(actualFull)] #drop the first 12 obs 	

		               actual          <- actual[13:length(actual)]         #drop the first 12 obs
				i = 4
			if(metd == "hw"){
				method 	<- "hw {forecast}"
				model 	<- hw(dataTrain, initial = "optimal",h=(length(data)-length(dataTrain)))				
				i = 2
			filename 	<- paste(method, ".png", sep = "")
			fit 		<-$fitted)[,1] # Fitted Values
			fitV 		<- unclass(fit)			
			fc		<- unlist(model[i])
			names(fc) 	<- NULL			
			fitForecast <- (c(fit, fc))
			comp 		<- cbind(dataTest, fc) 			
			compFull 	<- cbind(actualFull, fitForecast)	
			cat("Accuracy of Model:",metd,"\n")	
			mape		<- round(mean(abs((comp[,2]-comp[,1])/comp[,1]))*100, 2)	
			cat("\n",paste("MAPE of Predictions =" ,mape), "\n\n")
			upper 	<- c(actual,model$upper[,1])
			lower 	<- c(actual,model$lower[,1])
			#PLOT and SAVE
				plot(actualFull, type = 'l', col = 'black', lwd = 2.5, main = method, ylab = colnames(data.ts), xlab = "Time Intervals", ylim = c(min(lower),max(upper))) 
				lines(fitForecast, type = 'l', col = 'red', lwd = 2)
				lines(fitV, type = 'l', col = 'blue', lwd = 3)
				lines(upper, col = "orange", lwd = 0.5)
				lines(lower, col = "orange", lwd = 0.5)
				legend("bottomright", inset=.01, c("Actual","Fitted", "Predicted"), fill=c(1,4,2), horiz=TRUE, cex=0.75  )
#call the functions on HoltWinters() and hw()
forecastGen(metd = "HoltWinters")
forecastGen(metd = "hw")

If you are more interested in theoretical views you might find some related discussions here. If you are a beginner to R Programming language, this should help you to get upto speed. Author: Selva Prabhakaran Selva Prabhakaran Sanjeevi Julian


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s