Explaining Forecasts to Build Trust and Increase Adoption

Our client, a national restaurant chain, relies on sales forecasts to appropriately plan for the future. Short-term forecasts inform operational decisions regarding inventory management and labor scheduling, while long-term forecasts enable strategic planning. Leveraging a vast amount of historical sales data, our client implemented new machine learning (ML) models with a goal of improving forecast accuracy over prior methods.

While these models led to better predictive accuracy, their complexity made them difficult to interpret: Why did this model make this prediction? This lack of interpretability resulted in obstacles to trust and adoption among those who were expected to use the forecasts to make decisions. These obstacles were amplified when forecasts deviated from expected ranges without explanation. Lacking insight into the model’s “thought process,” team members found it easier to rely on their own judgment rather than the forecasts.

Elder Research developed an innovative solution that provides detailed explanations for each forecast by highlighting the primary components driving the forecast up or down. The underlying structure of the client’s predictive models made traditional model explanation methods impractical, so we tailored the solution to our client’s forecasting methodology.

For instance, if the forecast is higher than expected, the explanation can provide insight—perhaps identifying that summertime seasonal effects are leading to the increase. Here, the model has learned from historical data that summertime typically yields higher sales, and the explanation can identify and communicate this. The explanations will ultimately be displayed alongside the forecasts in an interface that allows restaurant managers to interactively see and interpret the forecasts. With forecasts demystified, end users can combine this information with their own expert knowledge to make more informed decisions.

Elder Research helped the client navigate a challenging problem, crafting a custom solution that explains the forecasts made by a suite of complex forecasting models. As a result, the client can harness the predictive capability of cutting-edge machine learning methods without compromising trust or stakeholder buy-in.