Demand forecasting sometimes look like a somewhat stagnant field: most methods used in the industry nowadays were already commonly available 30 years ago, says Joannes Vermorel.
Joannes Vermorel is the Founder of Lokad, a company that specializes in forecasting as a service. Vermorel is passionate about high dimensional statistics, machine learning and cloud computing. He wrote this guest blog on quantile forecasting:
In most large companies, my experience indicates that not much has been improved at the forecasting level for the last 10 years or so. Yes, forecasting software have now a better user interface, are easier to install, and benefit from higher quality data as input, but when looking at the statistical engine that produces the forecasts, in the vast majority of situations, it’s nothing but the good old dozen of classics: linear regression, exponential smoothing, moving average, Box-Jenkins, ARIMA, Holt-Winters, a bit of custom seasonality indexes, …
When I did create Lokad a few years ago, my goal was to bring something new to forecasting, and not to settle for status quo. Over the years, we have brought many small incremental improvements that over-perform the classics, but each 0.1% of extra accuracy has been an uphill battle.
Recently though, with a bit of luck, we did stumble upon a new way of looking at forecasts, namely quantile forecasts. Now, we have just put quantile into production.
Quantile regression has been around for decades among academic circle, but for some reasons, it failed to be noticed among supply chain practitioners. However, benchmarks that we have performed across the customer base of Lokad indicate that as far inventory optimization is concerned, quantiles forecasts are making “classic” forecasts obsolete in most situations.
In fact, “classic” forecasts suffer from a very indirect process to transform forecast values into inventory levels: this process is called safety stock analysis. This analysis relies on the core assumption that the forecast error is normally distributed. However, observations we have collected over dozens of retail and manufacturing businesses indicates that the “normal distribution” approximation is very loose at best. As a result, no matter how good are the forecasts, the weakest link is the safety stock analysis.
In contrast, quantile forecasts represent a direct approach to estimate reorder points. Instead of first producing forecasts, and then extrapolating those forecasts into inventory metrics, the forecasting engine directly produces the inventory metrics (the quantile equals the reorder point). Hence, it becomes possible to directly optimize the forecasting models to fit inventory constraints, that is, the lead time and the desired service level.
My take is that 10 years from now, most companies serious about inventory optimization will have “gone quantile”. Lokad has just a short head start.