How do I forecast during Recession?
In a client meeting on Friday – the 13th, I encountered a “scary” statement!! The category manager told me that his gut forecast was more accurate than the one generated by his ‘expensive forecasting system’ for last quarter or so. The symptom was recent and the forecast was going away by as much as 40%!!! Could this be symptom of a recession? How do I forecast during such times? Complex algorithms are far more powerful in finding out hidden patterns and extrapolates them beyond the capacity of human mind. Then why would such powerful models fail to detect a recession which is so obvious?
Forecasting systems have typically 12-36 months data. This works well in identifying patterns during regular times but not during recession. The demand falls dramatically during such times. Even before the system detects the dramatic drop in demand, probably a quarter or two has already passed!! The result is burgeoning inventories in the warehouse.
The key to forecast during recession is to detect it first. The first real indication comes when Forecasting systems miss three consecutive forecasts. During recession they continue to over forecast. The flip side of this rule is that if you forecast only once a month or a quarter, you still lose the plot! The only way to capture is to manually compare last 6 weeks forecast. During recession, demand for consumer products reduces dramatically. Mapping end consumer demand is the best indicator of slowdown. This means, for retailers, they need to look at their weekly sales of baskets and for Consumer Product companies – look at secondary or tertiary sales data, or syndicated sources for detecting the rate of slowdown.
The second step to this is to recalibrate the forecasting system. It could be done through variety of ways. One of the most effective ways to do it is to reduce the history window used in forecasting. Instead of 24 months, 3-6 months data would provide better picture. It is possible that some of the algorithms may not work with this little data. But then, if human mind can forecast better than the complex algorithms during recession, why not use simpler algorithms with less history – bootstrapping, weighted moving average, exponential smoothing etc. These algorithms are not as accurate as ARIMA, Regression or Nerual Networks, but performs better than them with a very short history. Typically, recession lasts anywhere between 6-8 quarters. These models should be deployed during such times. If the system is already using these algorithms, all the coefficients needs to be recalibrated every 2-3 weeks.
Most of the companies do forecast at a distribution center level and at a product level. During recession, customers become more price sensitive and are likely to switch the channels for buying the same product. Channel forecast becomes a balancing factor in manage inventory and satisfying consumer demand at a low cost. Distribution centers serve all channels and they are likely to experience different variability during recession. If the channel level forecast is brought into the picture, it will iron out inventory mismatch between distribution center and end consumer demand resulting into significant cost savings, higher fill rates and improved customer experience. After all, demand forecasting is all about improving customer experience, and there is no better time to do so, when everyone is going wrong in doing so… recession!!