Today was one of those days spent entirely on labeling logic. Turns out that my old “3% up in 24h = bull” rule was trash. Too many false positives in choppy markets. So I moved to a rolling z-score of returns over ATR-normalized periods. It’s noisier, but far more responsive to market context. The lesson? Don’t cheap out on labels. Models only learn what you show them, and if the labels are dumb, the predictions will be dumber. [[ML]] [[Labels]] [[Serendipity]]