Meta-labeling, from López de Prado's work on financial machine learning, splits trading into two models: a transparent primary (the rules engine) that decides what qualifies, and a learned secondary trained only on the primary's own outcomes, estimating the probability each qualifying trade wins. The secondary filters; it never originates.
An end-to-end learned trader is unauditable, fragile out of distribution, and trained against the hardest target in finance: predicting returns from scratch. Meta-labeling keeps the auditable rules engine as the decision-maker and gives the model a narrower, honest question: among trades that already passed every rule, which resemble the ones that historically worked? The rules stay inspectable; the model adds selectivity; and the failure mode of a bad model is fewer trades, not strange ones.
Each row is one completed rule-qualified trade: the market state captured at entry (volatility measures, strike deltas, scores, days to events, structure economics) and a label from the realized outcome. Honest labeling is most of the craft: barrier-aware labels respect the actual exit rules rather than pretending every trade ran to expiry, cross-validation must be purged and embargoed so overlapping trades never leak between train and test folds, and simulated fills stay out entirely, since mid-price fills would teach an edge that does not exist. Small datasets stay in simple regularized models; complexity is earned by row count, not ambition, which is why a tiered model ladder, logistic regression first and gradient boosting only past hundreds of labeled trades, is the standard shape.
Because it bounds what a wrong model can do. A secondary that can only remove or shrink trades degrades, at worst, into missed opportunities; a model that can originate trades imports every pathology of learned trading the architecture was chosen to avoid. The same asymmetry applies to its outputs: a calibrated win probability shown next to the rules' own score is context a human can weigh, and disagreement between the two is a flag worth reading, not an order. Options Scanner implements exactly this shape, with the model gate off by default and its accumulating forecast-versus-outcome record as the evidence for ever tightening it.