Source: Bertomeu, Cheynel, Liao & Milone (2024) — “Using Machine Learning to Measure Conservatism,” Management Science
At first this sounds like a technical finance question. Then you look closer and realise it is also a question about attention, incentives, and whether people are using information or just being impressed by it.
There’s a recurring tension in empirical accounting research between the clean theoretical models we write down and the messy reality of how firms actually behave. Linear regression models assume that accounting relationships — how earnings respond to returns, how accruals relate to cash flows — are approximately linear and stable. These assumptions are tractable and often produce publishable findings, but they import specification error that’s difficult to quantify. This paper, the published version of an earlier working paper on ML-based conservatism measurement, focuses on a specific mechanism: the role of non-linear relationships between observable firm characteristics and conservatism. The theoretical model of conservatism predicts that the same economic events should be recognised differently depending on firm-specific factors — industry, financial health, institutional ownership, reporting history. Linear models either ignore these interactions or handle them crudely with interaction terms. A neural network can learn them directly from data. The result of allowing these non-linearities: substantially cleaner measurement. The ML-based conservatism measure has better fit, fewer anomalous observations, and less temporal instability than conventional measures. More importantly, it gives a different picture of how conservatism has evolved over time. The paper’s framing of ML as a “noise filter” is the insight I find most transferable.
So I would not read this as a neat technology story. It is a messy information story. The tools improve, but people still have to decide what deserves attention. Annoying, but true.