I made some ML models for a demand forecasting problem. My predictors were the time-related features derived from the DateTime column (as in 'week of the year', 'day of the month', 'day of the week', 'month', 'quarter' etc) and an extra regressor( a categorical column with 10 levels which I hot encoded). In every single feature importance I checked, only the time-related features were the 'contributing' and not my other features. Is it something common in Time Series Forecasting, the time-related being the only contributing features and the extra-regressors 'pushed out'? And if yes, why is this happening?


1 Answer 1


I don't think it can be classified as common or uncommon in general, maybe for certain types of problems. But, it can certainly happen. In time series problems, it's not surprising to have date/time related features to play important roles on future predictability, because they may be certain patterns, periodicity or high autocorrelation.

However, for example, if we're predicting the cash amount that will be withdrawn from an ATM for a certain region, and let's say we have a binary feature saying if tomorrow there is an event or not in that region, certainly it won't be an unimportant feature for most cases (unless the event can be explained with time related features as well).

So, it's not possible to generalize and it highly depends on the problem. And, it's not surprising as well.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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