Supply Chain View

Let’s face it, we never think our forecasts are accurate enough. We suspect there must be a better way out there. Often this is down to the simple fact that forecasting is never precise. How can it be? I often think that one of the most important eureka moments for a supply chain professional is the realisation that a good measure of the likely error in a forecast is as important as the forecast itself – more on that later.

Forecasting solved?

We might be tempted to believe that modern planning and forecasting systems have relieved us of the need to think too carefully about choosing forecast methods. But a frightening proportion of enterprises large and small still operate from some sort of spreadsheet or home-brewed database system (sound familiar?). Or we run an ERP system with fairly rudimentary forecast logic. Even if we do have a fully-functioned planning system that will deseasonalise, fit trends, find auto-regressive models, etc, etc… – how do we set the parameters that control these system? Are they set correctly now? Does anyone dare touch them?

Forecasting Principles website

The Forecasting Principles website is a useful resource I have recently discovered. The tone is rigorous and academic, but the content is extremely wide-ranging. It includes book and paper reviews, and an excellent taxonomy of forecasting methods.

Many of the approaches are not applicable to supply chain problems, but time-series methods are very well covered. In addition there is information about the M-series forecast competition, a thorough evaluation of published methods and proprietry systems against various sets of time series.

It also includes this excellent factoid:

Societies have been suspicious of forecasters. In A.D. 357, the Roman Emperor Constantius made a law forbidding anyone from consulting a soothsayer, mathematician, or forecaster. He proclaimed “…may curiosity to foretell the future be silenced forever.”

Some basic forecast advice

For what it’s worth, here are a few things I’ve learned from numerous implementations and improvement projects.

1. If you manage a large number of SKUs, your planners are going to struggle with very sophisticated forecast methods, even if they are heavily automated.

   Complicated methods, although they can produce very good results, need special care and attention. Here’s what D.S.G. Pollock warns about the popular Box-Jenkins or ARMA method in a lecture on forecasting:

   Even when the data are truly generated by an ARMA process, the sampling errors … can lead one to identify the wrong model. … When the data come from the real world, the notion that there is an underlying ARMA process is a fiction, and the business of model identification becomes more doubtful.

2. For very slow movers, no method is going to be a measurable improvement on a weighted moving average (because there are just too few observations) – focus improvement activity on stocking rules, rather than forecasts.
3. If your forecast method radically changes the forecast for a SKU from period to period, then it’s (a) too sensitive to the historical data (too complicated a model perhaps?) and (b) really going to mess up your suppliers. Move to something simpler.
4. Use robust statistics – this means the capping of any very large values in the demand history. For most supply chains this is completely legitimate: customer propositions and expectations are different for very large quantities, even if this is not formalised. If it can be formalised, then demand categorisation (see below) is the way to go. In any case a robust stats approach will make it easier to discern trend and seasonal patterns in demand.
5. You are using demand and not sales to forecast, aren’t you? Remember to take into account lost sales effects, and to move any sales out of forecast requirements and into planned requirements.

Links:

http://www.forecastingprinciples.com
http://www.qmw.ac.uk/~ugte133/courses/tseries/8idntify.pdf