0
$\begingroup$

I need to forecast sales of hundreds of products in a retail store. I have weekly sales data of each for 2.5 years (133 weeks). The value I want to predict is actually aggregated demand for each SKU for the next 3 months (to make inventory orders). The individual sales in each week is not important.

How do I go about this problem? Do I again aggregate the data to monthly and predict for 3 periods in future or do I leave it as weekly data? The issue is that aggregating the time series to monthly will reduce the data points to just 33.

The sales of each SKU is in itself very noisy so I used hierarchical forecasting with weekly data but there seems to be too much noise in the top category itself as seen in picture. weekly aggregated sales all SKUs weekly aggregated sales of all SKUs [top of the hierarchy]

Will I be able to do a good forecast for the time series in picture? I've tried arima with fourier terms (k=14) and they didnt give that good results neither did tbats. I'm thinking of doing monthly aggregation and using ets.

$\endgroup$

1 Answer 1

1
$\begingroup$

This is a very broad question. Either one of modeling and forecasting aggregate data, or modeling weekly data and aggregating the forecasts, could give you better results - we don't know. Alternatively, consider doing both and aggregating in the temporal dimension (take a look at MAPA).

In terms of whether you will be able to forecast your series well, How to know that your machine learning problem is hopeless? may be useful. In retail sales, it is typically most important to model promotions and calendar events, which will likely be easier to do on weekly level, especially if your week definition is aligned with the duration of promotions.

We have a number of earlier threads on retail forecasting.

$\endgroup$

Your Answer

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

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