Source: Chan et al. (2026) — “Using Machine Learning to Predict Extreme Earnings Changes”
I like finance papers that accidentally become papers about human behaviour. This one does that. It looks like a study about data, markets, or disclosure, but underneath it is really about how people decide what to believe.
There’s a question buried in the machine learning earnings prediction literature that this paper addresses directly: where exactly does ML add value over traditional models? The answer turns out to be highly concentrated — and it concentrates in exactly the place where the stakes are highest. The researchers show that predictive gains from machine learning are heavily concentrated in the tails of the earnings distribution. For moderate earnings changes — the kind that happen most of the time — ML models and traditional linear models perform similarly. The improvement from ML is marginal for the bulk of earnings observations. But for extreme earnings events — large jumps and crashes — the ML advantage is substantial. This makes sense when you think about what machine learning is technically doing. Linear models are well-suited for relationships that are approximately linear and gaussian. But extreme earnings changes tend to occur in non-linear regimes: they’re driven by specific business events, industry dynamics, and interactions between variables that don’t look like normal earnings variation. A Random Forest model that can capture complex non-linear relationships and interactions will do substantially better here. The hedge portfolio test confirms this economically: a portfolio based on the Random Forest’s predictions of extreme earnings events generates size-adjusted annual returns of around 12.84%.
My takeaway: AI and data do not remove judgment from finance. They move judgment to the part where people decide what the signal means, whether it is trustworthy, and whether they are brave enough to act on it.