1
$\begingroup$

Let's say we want to forecast sku-level demand one year and four months ahead, and we have daily demand data for the last 3 years.

Taking into account that most daily time series at sku-level contain many zeros. Is it better to group the demand by month and forecast 16 months ahead?

With the time series showing lumpy and erratic demand patterns, which model is appropriate for demand forecasting?

Taking into account that we only have 3 years of daily data, how would you split the dataset if we grouped it as monthly data? Would it be logical, taking into account that we do not have enough monthly data and that we are talking about product demand, to have the 33 months as train and the rest as test, and look for a model that predicts well at least the first 6 months?

$\endgroup$

1 Answer 1

1
$\begingroup$

It all depends very much on what you plan on doing with your forecast. If you use it for production planning, and your production plans are always frozen in monthly buckets, then you need a forecast on monthly granularity. This can be calculated either by aggregating to months and then forecasting, or by forecasting by day and then aggregating the forecasts. Try both with a holdout sample and see which one is more accurate. If you want to be fancy, do both and reconcile the forecasts, this often improves accuracy, but it may not be worth the additional complexity.

Conversely, if you need to forecast for daily supermarket replenishment, then you need daily forecasts. Either forecast on daily level, or aggregate-forecast-disaggregate.

Lumpy demands are typically forecasted using Croston's method (or the Syntetos-Boylan bias correction). This aims for an expectation forecast. If you need quantile forecasts for setting safety stock, you can add a distributional assumption, which may be questionable if your data are lumpy. Alternatively, you could do quantile regressions or similar. As above, what is appropriate depends on what you plan on using your forecast for.

For the split, I would usually recommend mimicking your "production" horizon. If you need forecasts for the next three months, then your test sample should contain at least three months. Consider rolling your test sample forward to get more information out of it.

You may want to take a look at Forecasting: Principles and Practice (2nd ed.) by Athanasopoulos & Hyndman and Forecasting: Principles and Practice (3rd ed.) by Athanasopoulos & Hyndman, or Boylan & Syntetos (2021), Intermittent Demand Forecasting: Context, Methods and Applications. Shameless self-promotion: I will be giving a workshop on demand forecasting at the ISF in July, which is very reasonably priced, and you would meet people with similar questions as yours.

$\endgroup$
1
  • $\begingroup$ What happens in the splitting is that by grouping the daily data into monthly I have only 36 months of data. Following the production approach, this would lead me to split the dataset into 20 months for training and 16 months for testing. Another easier question, if the production plans are always frozen in monthly periods, should the EDA or seasonal diagnostic be done using the daily data or monthly data? $\endgroup$
    – johndhd
    May 10, 2022 at 9:32

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.