Thinking Fast, Trading Late: System One Decision Models vs the Market
Dehkordi et al.
A pre-registered, blinded test of language-model readings of company news finds that a signal which replicated on validation fell to an information coefficient of 0.020 on the blinded test. Commercial models identified masked companies in 62% to 74% of cases.
Source published · In the Weekly Review of 8 November 2026
Filed under Quantitative finance, Large language models
The authors propose OLIVIA, a protocol that rules out look-ahead, memorisation and selection: point-in-time inputs with company names masked, selection on validation data only, an analysis plan frozen by cryptographic hash and a test period unblinded once. They apply it to models reading news about S&P 500 stocks.
A fine-tuned open-weights model replicated on two validation quarters, with daily information coefficients of 0.055 and 0.065, but fell to 0.020 with a t-statistic of 1.7 on the January to July 2026 test. Commercial models’ readings match the overnight move to the open, not the move after it, so the news is priced before a trader can act. No pre-registered strategy survived costs and a deflated Sharpe ratio, and masking hid names but not companies, against a 28% baseline.