At last year’s Gartner Supply Chain Executive Conference analyst Tom Enright summarised the essence of successful demand forecasting with the following statement: “Demand forecast accuracy depends on your ability to recreate the environment in which historical demand occurred.”
While seemingly innocuous, this statement represents a significant departure from conventional thinking. The traditional approach is to employ algorithms that try to match past and future demand patterns. The simplest example is so-called “naïve forecasting”, as in “If we sold 120 units last month, then we will likely sell 120 units this month.” Adding regression analysis doesn’t change the fundamental approach. “Because demand has been increasing an average of 10 additional units per month, then it will likely do so in the future” is simply projecting acceleration rather than velocity. Employing “best fit” ranking to identify an algorithm that matches backward-looking history is a more exotic version of the same fundamental approach.
Instead, Enright suggests that the entire approach needs to be turned on its head. We should not be looking to match historical results. We should be looking to match the historical environment. We should not look to match the numbers that come out. We should look to recreate the demand drivers that generated those numbers in the first place.
Enright says that poor forecast accuracy stems from firms’ inability to recreate the historical demand environment. According to Gartner’s benchmarking service, average forecast error for most companies is in the low 70s (MAPE, Unit-Location level, 30 day lag) and even “good” forecasting rarely gets beyond the high 70s (using the same criteria).
In his latest research, Enright sums up the way forward: Capture additional demand inputs to recreate the environment in which demand occurred. Leverage the proliferation of new data opportunities to better understand the past and predict future demand (How Demand Management Disruption Will Power the Future Supply Chain, 22 November 2018).
Enright says that common statistical forecasting inputs such as historical demand do not create a sufficiently comprehensive set of attributes to accomplish this. So he says, “Improving demand accuracy is now intrinsically linked to the use of analytics to recreate the environment in which historical demand occurred.” He suggests including inputs to demand calculations such as demand transfer, competitor pricing, weather, social commentary, and shipping and returns policies. In short, any capturable element that influences how customers purchase, whether in B2B or B2C environments.
Enright adds that employing these data sources may require new solutions to model the demand. He says, “The sheer volume of structured and unstructured data currently available is greater than most current demand technology can absorb and use for effective insight and decision making. In order to utilize this data to improve demand planning and forecasting, structured and unstructured data needs to be utilized in a different way than it is today.”
ToolsGroup calls this “demand modelling”. Demand modelling works from the bottom-up, as opposed to top-down, breaking the demand components into a series of internal and external factors—the demand stream—and looks at how each impacts. It looks at the specific factors driving demand at a granular and daily level for individual SKU-Locations. It considers external demand-shaping factors— like new product introductions, trade promotions, end-of-aisle displays, price reductions—that have an impact at the most detailed level, such as at the store, and incorporates them into the forecast.
In modelling demand, the signal is determined by "decomposing" the data into a signal and a noise portion. It makes the granularity of this baseline demand as detailed as possible. The more detail, the more signal is preserved—and the clearer the signal can be identified from within the noise. The most commonly used granularity is individual sales order-line, daily by item and by ship-to location. Additional information—seasonality, promotions, market intelligence—further fine-tunes the separation of signal from noise for a “quieter,” more accurate forecast.
Enright adds that this new set of data inputs may require the use of machine learning algorithms to learn from richer historical data to sense demand from a wider variety of data inputs. He concludes, “Using technology to analyse more data sources will produce a more accurate demand plan.”