Source: van Binsbergen, Han & Lopez-Lira (2023) — “Man versus Machine Learning: The Term Structure of Earnings Expectations and Conditional Biases”
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.
Analyst earnings forecasts are optimistic. This has been known for decades — the literature on analyst bias is extensive. What this paper adds is a way to measure the bias precisely, in real time, using a machine learning benchmark that provides a statistically optimal unbiased forecast. The idea: train a machine learning model to produce unbiased earnings expectations based on publicly available information. Then define analyst bias as the difference between what the analyst forecasts and what the ML model forecasts. This “conditional bias” can be measured for each analyst, each firm, and each point in time. Several findings follow from this measurement approach. Analyst forecasts are, on average, biased upward — the ML model’s expectations are lower. The bias increases with forecast horizon: the further out the forecast, the more optimistic analysts are relative to the unbiased benchmark. This is consistent with analysts providing strategically optimistic long-run forecasts to maintain management access and investor relationships, while gradually revising down as the reporting date approaches. The capital market implications are significant. Firms with the most upward-biased analyst forecasts — those where analysts are most optimistic relative to the ML benchmark — are associated with negative cross-sectional return predictability.
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.