stockprecog predicts P(trade succeeds) cross-sectionally on Brazilian (B3) equities — and it is deliberately a study of honest invalidation. The headline finding is modest and partly negative, and that's the point. The contribution isn't a magic alpha; it's the anti-self-deception evaluation ruler, built on the López de Prado / Advances in Financial Machine Learning methodology: point-in-time data, causality-audited features (zero look-ahead), deflated Sharpe ratios, and every trial logged to a JSON file rather than cherry-picked.
What the ruler actually said
Price-derived signals — momentum, volatility, fractional differentiation, Amihud microstructure, CUSUM/SADF regime detection — are statistically real but economically sub-marginal: gross deflated Sharpe around 0.41 collapses to roughly −1.1 net of costs at a 10-day rebalance. The result is robust in the ways that matter: decile sorts don't rescue it, longer horizons don't rescue it, and even a spread-only cost floor sits near zero for large caps. If a pipeline this careful says a signal is sub-economic, that verdict is the result.
The README opens with the sentence a recruiter should read twice: this is not a trading system and never claims to be. It's a worked example of how to evaluate a strategy without fooling yourself — 26 tests keep the pipeline honest about its own mechanics.
Why publish a negative result
Because negative results with rigorous methodology are rarer and more useful than optimistic backtests, and because publishing forces the discipline. The study is archived on Zenodo with a citable DOI (10.5281/zenodo.20706701) and the dataset tooling is reproducible end to end.
To be clear about positioning: I'm not a quant and I don't present myself as one — this lives in my portfolio as evidence of method: how I measure, how I resist the urge to believe my own results, and how the same "test, don't infer" reflex from my engineering work behaves when pointed at finance.