2
$\begingroup$

I work at a company where we plan 60 days ahead. We have 5 years of historical sales data recorded daily. The data exhibit trends, weekly seasonality, and yearly seasonality.

Currently, I'm training the model with all historical data to forecast the next 60 days. However, I'm finding it difficult to achieve a good fit for the Christmas/New Year's Eve period, which sees a peak in sales.

Since I only need to forecast 60 days ahead (July and August), I'm wondering how important it is to achieve a good fit for periods (like the end of the year) that my model will not be forecasting.

Ideas that came to my mind:

  1. Use only data from the months that will actually be forecasted
  2. Try to smooth periods of high demand that are not part of the forecast period

However, I'm unsure if these ideas are good practices or even recommended for my situation.

Improve this question
$\endgroup$
2
  • $\begingroup$ Given the unit of analysis, without controlling for seasonalities (day of week, month of year, etc.), parameters for non-seasonal observations will be biased. $\endgroup$
    – user78229
    Commented Jun 27 at 6:03
  • $\begingroup$ Have you tried to add the special calender days as features? $\endgroup$
    – picky_porpoise
    Commented Jun 30 at 6:33

1 Answer 1

Reset to default
3
$\begingroup$

If by "fit" you mean the in-sample fit, beware. In-sample fit is not a good indicator of out-of-sample forecast accuracy.

It is not surprising that your forecasts (or fits) for periods of high sales are worse, because higher sales usually also come with higher variance. This effect is usually neglected in forecasting, unfortunately.

If your products show yearly seasonality, then it is probably better to use as much historical data as possible, while accounting for your calendar events in some way, either by using features or by just cutting these periods out of the training sample if your model allows for this. Do use a model allowing for if you also have intra-weekly seasonality. One caveat is that if your Christmas sales are really strong, they might be picked up as spurious seasonality in spite of covering them with features, so keep that in the back of your head.

If your sales are not seasonal, then you may find the approach you suggest workable.

Yet another approach would be to do both approaches and take the average of the two forecasts. Averaging very often improves forecasts.

And no, whatever you do, you really don't need to worry too much outside Christmas about how your model performs around Christmas (except for that possible spurious seasonality mentioned above).

Improve this answer
$\endgroup$

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.