outlier, Forecast, demand, statistical model, History, supply

03 Feb 2012

Automatic Outlier detection - Blessing or Curse?

One of the puzzling questions that Demand Planners ask in our training workshops is why their software produces a flat forecast 90% of the time.  An expensive software that took an army and a couple of years to implement typically suggested a constant model or moving average model.  This resulted in a flat forecast.

Although the naked eye can graphically see (if graphs are made available to the user) a nice seasonal pattern, the expert selection in the software produced a constant forecast of eternity.  There are many tricks underlying this final result – some of them known and some of them hidden.

One of the culprits is the outlier detection process.  The software can intelligently detect outliers for a given setting and outlier detection method.  Typically you use a K-factor to develop tolerance bands around the ex-post factor to identify outliers.  What are K-factors and how to leverage the K-factor settings to produce good forecast models?

We observed in a variety of cases, people use low k-factors that then throw out all seasonal peaks and troughs.  A low k-factor is super vigilant.  It does not allow any pattern to escape through to the modeling engine.  All the engine sees is just a cluster of a few data points that are closely scattered around the ex-post forecast or just a historical mean.

See the picture below.

A k-factor of 1 will eliminate all patterns seen in the demand profile.  It just keeps a fraction of the original data set that all points to the historical mean as a violently accurate forecast.  This has nothing to do with the power of the statistical engine available to the software.

Leave a Reply