When and how to adjust statistical forecasts in supply chains? Insight from causal machine learning

Abstract

ABSTRACT Empirical studies have discovered that most statistical forecasts in supply chains are subjected to judgemental adjustments during a forecast review. Although such a practice requires significant management effort and frequently reduces forecast accuracy, many organisations prefer this approach as part of their Sales & Operations Planning process. This study aims to identify the optimal policy to achieve significant accuracy improvement from forecast review. We focus on a practical situation where managers periodically review forecasts from the statistical software and compare them with judgemental forecasts from the sales and marketing functions. Managers must decide whether to disregard the judgement and continue with the existing forecast or revise the statistical forecast based on the judgement. To determine the best course of action, we conducted a numerical experiment using data from five supply-chain companies wit more than 12,000-point forecasts. The experiment considered three alternative actions: “do-nothing”, “follow the judgement”, and “simple-average”. Using a causal machine learning method, namely a policy tree, we develop a set of decision rules that maximise the expected accuracy gains given the variation in forecasting features. The result proposes a simple yet effective policy to recommend suitable actions based on two identified key features: “judgment direction” and the “accuracy of statistical forecasts”. The policy was tested against real-world data and achieved remarkable accuracy with roughly a 3–11 percentage points improvement over the baseline. Our findings offer valuable insights for managers to customise their forecast review policy based on their unique environment.