Why do forecasts go wrong?

A lot has been written about forecasting and the need to improve its accuracy. Despite our best efforts, the fact remains that improving forecast accuracy continues to be an elusive goal and achieving 80% accuracy levels at the granular product-geography combination remains the holy grail.

Recent advances in sophisticated statistical models, big data and machine learning will definitely help make better forecasts in future in a pristine environment. At the same time, it’s equally important to investigate why forecasts eventually turn out to be substantially different from the actual demand despite our best efforts and intentions.

While competitors’ actions will continue to be unpredictable, making it difficult to forecast actual demand, this factor remains outside the organisation’s control. What’s more important is to control the other internal factors which prevent us from making accurate forecasts.

In my opinion, there are three main reasons within an organisation’s control which pull the forecasts away from reality. Understanding these factors will lead us to the right direction of solution.

1. Forecasting is done too infrequently. A lot of FMCG companies still operate on monthly forecasts, whereas fresh demand signals are now available on a daily basis. The practice of monthly forecasting started when reliable data at a granular level could only be collected once a month. We have progressed by leaps and bounds in terms of collecting reliable granular sales data at a daily frequency, however forecasting still continues to be done monthly.

2. Forecasting is done too much in advance. Companies typically forecast about three-four weeks before the start of the first forecast period. They would churn out estimates for the next three-six months and even roll-out a twelve months forecast at times. While it does help in Rough Cut Capacity Planning and ordering long lead time items, it need not be done in great detail. Since fresh demand signals and patterns are now available on a daily basis, we definitely miss out on many such important signals when we forecast so much in advance.

3. Forecasts are often used to measure sales performance. Many companies still focus on forecast achievement, whereas measurements should actually be done on sales growth and market share. If incentives and career progression are linked to forecast achievement, it becomes a political number, negotiated between the sales team and the top management. Top management looks at providing a stretch to the sales team and pitches for a forecast higher than the estimated demand. Sales team, on the other hand, wants to play safe and wants a forecast lower than the demand. Final forecast depends on who prevails in the end. In any case, the tussle takes it away from reality.

If forecasts are to be used by supply chain teams to move stocks to various locations, it is imperative that we get rid of these three shortcomings.

Latest demand signals need to be processed to update forecast on a daily basis, thereby improving the frequency of forecasts. Taking the latest signals into account would mean that today’s signals will change tomorrow’s forecast, which takes care of the problem of forecasting too much in advance. My final recommendation would be to take the human touch out of forecasting and process the demand signals in a pristine manner without any subjective bias creeping into it. Let the pattern of demand signals decide what the forecast should be.

Theory of Constraints based replenishment process has these three recommendations firmly embedded into it. If properly designed heuristics could convert the demand signals into action areas for the supply chain teams, should we even spend our efforts in forecasting demand at such a granular level?