The supply chain approach to planning and procurement management

Hewlett-Packard Journal, Feb, 1997 by Gregory A. Kruger

Consider a part used in j products and used [k.sub.i] times in product i, where i = 1, 2, ..., j. Let DE represent the forecasting or demand error. Then:

[DE.sub.part] = [k.sub.1][DE.sub.product1]

[k.sub.2][DE.sub.product2]

[k.sub.3][DE.sub.product3]

... [k.sub.j][DE.sub.productj]

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] www The big problem with this approach is the assumption of independence of forecasting errors among all the products using the part. If, for example, when one product is over its forecast there is a tendency for one or more of the others to be over their forecasts, the variance calculated as given here will underestimate the true variability in part demand uncertainty.

The second approach to estimating forecasting uncertainty for common parts is to explode product-level forecasts into part-level forecasts and product-level customer demand into part-level demand and measure the demand uncertainty directly at the part level. For a part common to j products we simply measure the forecast error once as the difference between the part forecast and actual part demand instead of measuring the forecast errors for the individual products and algebraically combining them as before. Any covariances between product forecasting uncertainties will be picked up in the direct measurement of the part-level forecasting errors. Clearly, this is the preferred approach to estimating part demand uncertainty, since it avoids making the assumption of forecast error independence among products using the part.

Estimation of Demand and Part Delivery Uncertainty

The whole approach to safety stocks and inventory management outlined here is dependent upon the basic premise behind any statistical sampling theory-namely, that future events can be modeled by a sample of past events. Future demand uncertainty is assumed to behave like past demand uncertainty. Future delivery uncertainty is assumed to behave like the supplier's historical track record. This raises two issues when estimating the critical inputs to the safety stock equations: robust estimation and business judgment. Both of these issues are extremely dependent upon the chosen frame of reference, that is, whether we are measuring real-time customer demand or part-level consumption on the factory floor.

From a sample size perspective we would like to have as much data as possible to estimate both demand and delivery uncertainty. However, in a rapidly changing business climate we may distrust data older than, say, six months or so. If I am measuring demand uncertainty as the deviations between real-time customer orders and the forecast, do I want to filter certain events so they do not influence the standard deviation of demand uncertainty and hence safety stocks? It may be good business practice not to allow big deals to inflate the standard deviation of demand uncertainty if those customers are willing to negotiate SRT. In statistical jargon, we want our estimates going into the safety stock equation to be robust to outliers. Naturally, if the demand uncertainty is measured as part consumption on the factory floor versus planned consumption, data filtering is not an issue. It is possible that an unusual event affecting parts delivery from a supplier may be best filtered from the data so that the factory is not holding inventory to guard against supply variability that is artificially inflated.