2
$\begingroup$

I just read Demand-Driven Forecasting: A Structured Approach to Forecasting(Wiley and SAS Business Series) and have a few doubts in Holt-Winters Model:

1) Unlike OLS Regression Modeling technique or ARIMA, no assumptions were checked before running Holt-Winters. For instance, in ARIMA, we first make the data stationary before running ARIMA or in OLS, we check normality, auto-correlation etc. However, as per the book, no test was conducted before and after running Holt-Winters. We just calculate MAPE and check if it is acceptable. So, can someone confirm if there are any tests that we should do before and after running Holt-Winters?

2) Can Holt-Winters model incorporate causal factors such as Price, Promotion, Marketing etc?

3) When we are running Holt-Winters multiplicative or additive model, we don't need to explicitly adjust for seasonality before?

4) How to treat for missing values or outliers in time series forecasting? We obviously cannot remove missing values rows as the time interval will break. Hence, for missing values, I have thought to use CAGR for that period(weekly, monthly or yearly) in the last 5 years and use it to estimate missing value. Can someone confirm if this correct?

Also, I believe we shouldn't remove outliers and rather look why it is there in the first place. It might be because of some external factors like Price Change or Promotion which is causing the spike. This will ofcourse be not detected by Holt-Winters. So, if it is a rare promotion, then we should remove that value and insert the normal demand again by using CAGR of past data in order to maintain the historical pattern. Hence, the spike will be gone and the historical trend will be intact, thus increasing our MAPE. Can someone confirm if this correct?

$\endgroup$
0
$\begingroup$

1) Holt-Winters doesn't require stationarity the way an ARIMA model does, so you don't need to perform any of the steps you mention. However Holt-Winters is specifically designed for seasonal data, so don't bother applying it unless you have good reason to think your data is seasonal.

2) In the basic formulation of Holt-Winters, no, it can't. More recent implementations of Holt-Winters use the state space approach to time series modeling, and it might be possible to do so using the state space approach - although even that is still in debate. BSTS is another state space model with a structure similar to Holt-Winters, and that approach does allow for causal factors. You should check that.

3) No, you don't need to. The whole point of Holt-Winters is that it accounts for seasonality, so you don't need to adjust for it.

4) See here. Holt-Winters can't handle missing values, but other methods can. You should be careful with Outliers, since you don't know whether an outlier is a "real" outlier, a seasonal spike, or a causally driven event. For example in a sales time series, your spike might be a true outlier, or it might be an event like Black Friday, or it might be the result of a promotion and advertising campaign.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.