Whenever I see results of an experiment that include Holt-Winters method, I shrug. You should not use it, and here is why.

Holt-Winters was developed in 1960 by a student of Charles Holt, Peter Winters (Winters, 1960). He extended Holt’s exponential smoothing method (the method that introduced a trend component) to include a seasonal component. The method has performed well in many situations, but it was originally developed for a specific type of data: trend-seasonal. Furthermore, the original method implied that the noise had an additive form, while the seasonality was multiplicative (which is a weird combination).

Since then, lots of things have happened in forecasting, one of the biggest being the development of the ETS framework by Hyndman et al. (2008). That framework covers 30 possible models for time series with different types of Error, Trend, and Seasonal components. Holt-Winters is a method that aligns with only one of the models from the framework.

Now, if you have data without a trend, we know that the trend-based models will perform poorly, overfitting the data. Similarly, with seasonal models on non-seasonal data or other misalignments between the model and the data. This is a “horses for courses” sort of thing: you should use the model that suits your data, not the one you can easily find in a default package. And this means that Holt-Winters works as intended on one out of roughly 30 possible types of time series. Sticking only to it is similar to applying SARIMA(0,2,2)(0,2,2) to all your data, and then wondering why it performed poorly.

Interestingly enough, data scientists that do not know forecasting very well typically compare their ML approach with ARIMA with automatic order selection (usually the one developed by Hyndman & Khandakar, 2008) and explicitly with Holt-Winters instead of trying ETS. I don’t know why they do that. The only explanation I have is that they probably just are not aware of the more modern approach to exponential smoothing.

If you want to learn more about exponential smoothing, my book is available online, and Kandrika Pritularga and I will deliver a workshop on that topic at ISF.

Correction: Rob Hyndman correctly pointed out that “the seasonal method now known as Holt-Winters was actually developed by Holt, and is described in Holt (1957)“, not as I stated above in Winters (1960). Winters (1960) was the published paper describing the method, while Holt (1957) was an unpublished report for the Logistics Branch of the Office of Naval Research (ONR).