The goal is to predict out-of-stock situations, either quantitatively (the gap) or qualitatively (out-of-stock likely to happen in next few weeks). Background:
We have existing demand planning process, through heuristics and software packages (using time series models). Most of the time they perform well, and recorded out-of-stock is rare (~ 8% of total orders rejected due to out-of-stock). Hidden cases might exist, meaning customer did not even place the order after knowing we ran out of stock.
Out-of-stock happens "in clusters". Out of a two year period, we have only 3-4 out-of-stock periods, each lasting one to two weeks. They did not happen when the demand was the highest, because we planned for that.
- Time series can capture the average demand, however they cannot model extreme / unexpected cases, which cause out-of-stock. Using upper confidence interval can cover some extreme cases, however this significantly increases warehouse cost.
- Not enough data for machine learning - essentially only 3 out-of-stock cases because they happen "in cluster".
Please suggest how I can handle this - any input welcomed...