Blog
Technical articles on the project's design, validation and results.
My pipeline said OK. The data had been wrong since January.
A monthly macro series stored as if it were daily, 147 wrong values in a core feature, and no alarm. How I found it by comparing the database against the source.
I trained 72 machine learning models to predict the S&P 500. None of them beat the baseline.
A full walk-forward evaluation of 72 models on SPY, and an honest look at the two that seemed to work — and why they probably didn't.