Light-touch forecasting: A novel method to combine human judgment with statistical algorithms

Abstract

Forecast adjustments are an indispensable component of the forecasting process, but what is the most effective and efficient method to make these adjustments? Understanding how to effectively blend human forecast adjustments with statistical methods is of great importance as, even with the increased possibilities of AI, we can still not include all information available to a human planner. We address this issue by developing a method that builds on literature showing that some adjustments are consistently (in)accurate. More specifically, in two extensive case studies containing more than 3.5 million forecasting decisions, we confirm that planners are accurate in adjusting a statistical forecast in the right direction and determining the magnitude of downward adjustments, and they are inaccurate in determining the magnitude of upward adjustments. Leveraging these results, we introduce a novel method called light-touch forecasting, which attains performance levels similar to those of more traditional forecasting methods while minimizing the involvement of human planners. Furthermore, an online experiment shows an efficiency gain of 38% in terms of time spent on planning compared to traditional judgmental forecasting. We thus optimize the forecasting process by using the strengths of planners while avoiding their weaknesses.