4

I've been tasked to generate a forecast for our newly operation business which has more than 500+ sku. Almost 90% of them are following intermittent demand pattern, with very few data points to train upon (maximum I recorded 80-90 data points for each product) I cannot train an ML model.

Is there a way or model where I can train such small data set using excel or any other tool?

Link to data Intermittent_DATA

Dan
  • 43
  • 5
  • 1
    In my experience, 90 points is enough for forecasting models such as ETS, but you might have to agregate your SKUs into families if the individual time series are chaotic. If you share the data, you will likely have more help with your problem. – Kuifje Nov 19 '21 at 12:04
  • I've added link to data – Dan Nov 19 '21 at 12:13
  • 1
    A very quick and dirty approach shoes that if you aggregate all of your data and plot the time series, there is no trend, nor seasonality. I am not sure there is much to be done with the data set. – Kuifje Nov 19 '21 at 12:31
  • If you're trying to forecast the lead time demand, you might consider Willemain et al. (2004). – SecretAgentMan Nov 19 '21 at 13:17
  • I agree with Kuifje. You might try regressing data for each SKU separately on indicators for day of the week, to see if there is a "seasonal" pattern within a week. Otherwise, your best bet may just be to fit a stationary distribution to each SKU (or to the daily sum across SKUs). – prubin Nov 19 '21 at 17:00

2 Answers2

2

You can consider aggregating demand by grouping some SKU's (group that makes more business sense) and by which you can have more points for a group. You can train (though data seems less) and then forecast output & later post-process it by splitting them across individual SKUs based on some business rule/logic.

anjikum
  • 984
  • 4
  • 8
0

You can try facebook prophet model, I have a similar demand forecast task, the data is also limited. Prophet can provide a good forecast for my case.

Chemmyyu
  • 151
  • 1