Abstract
Financial-news studies commonly aggregate articles by day or use broad event windows, obscuring the elapsed time between publication and a tradable market response. We ask whether timestamped news and minute-level prices can identify how information persistence varies by release session and volatility. We propose Differentiable Session-Aware Rapid Attention Decay (DiffSARAD), which jointly learns a continuous global half-life, session- and volatility-conditioned adjustments to it, and a three-class post-publication abnormal-return classifier. The representation combines conditional exponential weights with both weighted FinBERT sentiment composition and decayed news intensity, preserving half-life identifiability for single-news histories. Using 23,294 analyst-rating items for 23 equities from 2 January 2018 to 30 June 2020, executable 30-, 60-, and 120-min labels, and 18 expanding purged folds with strictly separated validation, the estimated base half-life is 17.4–23.0 h. Intraday coefficients are positive and post-market coefficients negative in all 18 folds at every horizon, indicating slower intraday and faster post-market decay, and the pattern recurs under a second processing version. Synthetic recovery experiments on the archived event-history design confirm that these parameters are recoverable and that near-zero coefficients are returned under a zero-effect generating process. News and decay features lower daily log loss versus price-and-market features under serial-dependence-robust inference, and verified nested ablations attribute small but significant loss reductions to each conditioning component. The news-count baseline remains the strongest classifier, and a two-basis-point cost stress turns net returns negative.