In 1994, the IBM Micro-electronics Division, itself a fortune 100 size firm, put in place a major effort to create best in class supply chain planning process and software including demand planning(DM), central planning, available to promise, et al. I was fortunate to be an original member and had the opportunity to work extensively on all key components including DM (1996-97) – created an estimate of demand – a forecast.  With 20 years of experience in computational methods at that time, I thought DM would be a walk in the park. I quickly learned it was illusively complex.

One of the critical challenges was finding the “best way” to mix and match judgment and statistical forecasts, where there are typically multiple judgments. In 2015, we now refer to this as effective collaboration.  I went to my math books full of best optimization methods such as non-linear least squares, branch and bound, dual simplex, maximum likelihood…none were helpful. Fortunately, IBM had contracted with Supply Chain Consultants (now Arkieva) to lead the effort and brought with them some collaboration methods that have proven successful in previous DM applications they had undertaken.

Fast forward 20 years and Arkieva has continued to build on this early leadership and just published a paper “Practical Considerations in Forecast Value Added Analysis” in the summer 2015 issue of FORESIGHT with some easy to follow, but brilliant insights to generate successful collaboration.  Key points made in the article are:

  • The Forecast Value Added (FVA) concept is designed to determine which if any steps in the forecasting process—particularly those steps that impose judgmental overrides to statistical forecasts—improve forecast accuracy, and which do not.
  • The statistical forecast is typically passed on to collaborators given little or no guidance other than to “improve on it if you can” or “check it for reasonableness.” This sets the folks responsible for improving the forecast in direct competition with statistical forecasting. A better practice would be to view the statistical forecast as one of the collaborators, not the competition.
  • When there are multiple levels of overrides—from sales, marketing, management—averages or weighted averages of these can be taken to create a consensus forecast. FVA analysis can determine if the consensus forecast proves to be better, not just whether the individual inputs improve the forecast. In fact, it is possible that each individual collaborative input can make things worse, while the consensus forecast shows overall improvement.
  • The right question, then, is not whether each of these inputs adds value, but whether each of these inputs can be combined to effectively integrate analytics and planner expertise into a better forecast.
  • An adaptive weighting scheme is suggested for calculating consensus forecasts. In this scheme, the weights assigned to the collaborators change over time to reflect changes in their relative forecasting performance.

As Dr. Karl Kempf (Senior Fellow Intel) stresses the successful interplay between human expertise with analytics generates smarter solutions faster.

Like this blog? Please share with colleagues and also follow us on LinkedIn or Twitter and we will send you notifications on all future blogs.