Source: Bertomeu, Cheynel, Liao & Milone (2021/2024) — “Using Machine Learning to Measure Conservatism”
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.
Accounting conservatism — the tendency to recognise bad news faster than good news in financial statements — is one of the central concepts in financial reporting research. There’s a large body of evidence that it exists, that it varies across firms and time periods, and that it has real economic consequences. The standard approach relies on linear regression models that compare how quickly earnings reflect positive versus negative market returns. These models work in the aggregate but have well-documented weaknesses: they’re sensitive to specification choices, they produce anomalous observations, and they’re unstable across time. This paper asks whether machine learning can do better. The answer is yes — but in ways that are more interesting than just “better fit.” The researchers use neural networks to extend the differential timeliness model, allowing the relationship between market returns and earnings to be non-linear and to depend on observable firm characteristics in more complex ways than traditional models can capture. The resulting ML-based measures of conservatism outperform conventional measures on several dimensions: better fit, fewer economically nonsensical observations, less year-over-year noise. One finding particularly stood out to me: the ML measure reveals a secular decline in accounting conservatism over the study period that the linear models partially obscure.
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.