NMR · 2026-10-30 20:24 UTC

One Prompt Generates 1,400 Features

Unity dataset researchers used Claude to generate 1,400 new features from stock historical data by applying four transformations (own-history rank, moving-average crossover, weeks since peak, weeks since last change) to 350 base features. A single prompt to Claude Opus 5.5 created a feature engineering pipeline that increased an example model from 1,040 to 2,440 total features, achieving high Neutral Benchmark Contribution on validation data. The entire feature generation process cost approximately $30 in tokens and produced features spanning every stock back to 2003.

Loading the story on Tree News…