The transformer didn't predict the return. It predicted its own confusion, and that's what mattered.
A new paper from Rutgers introduces WTrans: a signal built not from what a transformer thinks a stock will return, but from how uncertain the model is about that prediction. Alone, this uncertainty measure predicts nothing (t ≈ 0.5). But interacted with anomaly consensus direction (µ), it becomes one of the strongest signals documented in recent cross-sectional research.
Among stocks where anomalies agree it's a buy, high model uncertainty → higher future returns
Among stocks where anomalies agree it's a sell, high model uncertainty → lower future returns
The interaction µ × WTrans yields a Fama-MacBeth t-statistic of 8.68 and a FF6-adjusted DiD alpha of +1.30%/month
The conditional long-short factor has an R² of just 1.5% on FF6 — genuinely orthogonal to the entire factor zoo
The mechanism is limits-to-arbitrage through investor attention: the premium exists almost entirely in low-media-coverage and low-analyst-coverage stocks, and essentially vanishes where informed arbitrageurs are active.
The signal also beats cross-theme disagreement (σ) in a direct horse race — σ drops to t = −0.19 when WTrans enters the same regression.