Forecasting

No doubt about it – despite being critical for effective business planning, accurate forecasting is complex and challenging and still remains elusive for many organizations. However, as the recent issue of APICS Magazine points out in their article “Outlook Warm and Sunny”, one can create good forecasts through the proper combination of judgmental and statistical methodologies and use them to identify new market opportunities, anticipate future demands, effectively schedule production, and reduce inventories.

What’s interesting about this article is that it is well known that neither technique on it’s own can be very effective. Most of us lack the ability to accurately judge future demand due to limitations in human cognitive abilities, the restricted amounts of information we have at our disposal, and unknown causal relationships. Similarly, statistical forecasts are limited with respect to the models they are based on. Although a statistical model is much more accurate than any intuitive model we could come up with, it is built on assumptions and causal relationships which may change over time. The best example of a statistical model gone bad is Nike’s $400M failure in 2000 due to demand forecasting software. Nike relied exclusively on automated forecasts without any judgmental checks, but the newly implemented models were not yet fine-tuned and accurate enough to be deployed in a fully automated mode.

The best forecasts are those that leverage the strengths of both judgmental methods and statistical methods. However, as the author points out, well-established rules must be followed in order to effectively combine these techniques.

The following table summarizes the strengths and weaknesses of each approach.

Judgmental Forecasts
Strengths Weaknesses
Responsive to latest environmental changes

Can include “inside” information

Can compensate for “one-time” or unusual events

Human cognitive limitations.

Biases

Statistical Forecasts
Strengths Weaknesses
Objective

Consistent

Can process large amounts of data

Can compute many variables and complex relationships

Slow to react to changing environments

Only as good as model formulation and available data

Can be costly to model “soft” information

Require technical understanding

According to the article, judgmental and statistical forecasts can be combined in different ways to take advantage of their individual strengths but the most popular method in practice appears to be the managerial adjustment of statistical forecasts where managers adjust the statistical forecast in a “managerial override”. Managerially adjusted forecasts can often improve forecast accuracy by including information not available to the statistical model. However, if performed incorrectly, adjustments can cause inaccuracy due to inherent human bias. Thus, established rules should be followed for effective adjustments.

The rules outlined by the article are the following:

Only practitioners with domain knowledge should adjust statistical forecasts.
Judgmental adjustment is more likely to improve accuracy when the adjustment is based on domain knowledge. Generally, only domain practitioners will be aware of the relevant contextual information that should be used to adjust a forecast.
Adjust statistical forecasts when there are known changes in the environment.
The adjustment should compensate for specific events not captured by the statistical model or time series. It should not be based just on intuition or bias.
Structure the judgmental adjustment process.
Use a documented or computationally consistent methodology. This will allow you to repeat successes and insure that failures are caught, corrected, and not repeated.
Document all judgmental adjustments made and measure forecast accuracy.
Records must be kept of all adjustments made, and the reasons therefore, and the results of the forecast must be measured so the process can be improved over time and the underlying statistical models updated when relevant observations are made.

When good, quantifiable, and historical data is available, reliance should be placed primarily on statistical forecasts. Only when the domain practitioners know of relevant contextual events or information not contained in the model should judgment be used to adjust the forecast.