Meta-Labels

A momentum signal fires 1,000 times per year but only 40% are profitable. Instead of discarding the signal, train a second model to predict which of those 1,000 trades will work.

The Problem

Raw trading signals typically have acceptable recall (they catch most real moves) but poor precision (many false positives). Tuning the primary model to improve precision degrades recall. Meta-labeling decouples the two: the primary model generates candidates, the meta-model filters them.

The Pattern

WRONG

from sklearn.ensemble import GradientBoostingClassifier

# Use raw signal directly - many false positives passed through
signal = primary_model.predict(X)  # 1=buy, -1=sell, 0=hold
positions = signal  # Every signal becomes a trade