Warranty Provisions: Machine-Learning Versus Human Estimates

Abstract

This study employs machine learning to shed light on the accuracy of discretionary accounting estimates and the causes of human estimation errors. Using proprietary data from a large European manufacturing firm, we implement a set of prediction models to gauge a pervasive and economically relevant accounting estimate: the warranty provision. We find that machine learning models consistently outperform human experts when compared on the basis of individual warranty obligations. This gap widens when estimates are aggregated across homogeneous classes of products, as the machine makes relatively fewer and less severe overstatements. Applying model interpretability techniques and conducting a series of semi-structured interviews, we identify misspecifications of the managerial estimation model, specifically aggregation bias and anchoring to historical cost, as the primary causes of the larger human errors. Moreover, the interview evidence suggests that various firm-level factors, such as learning frictions, auditors’ preferences for process continuity, and strategic considerations, are important determinants of the design and continued use of misspecified estimation models in practice.