Skip to content
Machine Learning

Labeling: The Triple-Barrier Method and Meta-Labeling

Account required to view full content

The previous lesson showed how to validate a financial model without leaking. This one is about the input that decides whether the model is worth validating at all: the label. In a quant problem you choose what “the right answer” even is, and most candidates choose badly. Labeling a fixed number of days ahead ignores volatility and the path the price took to get there, so the model learns from outcomes a real strategy would never have realised. The triple-barrier method fixes this by labeling the outcome of an actual trade with a profit target, a stop loss, and a time limit. Meta-labeling then adds a second model on top that decides whether to act on the first model’s call and how large to bet. This lesson defines both, sets the barriers from volatility, labels a price path by hand, and works a meta-labeling confusion matrix to show exactly how it lifts precision. It builds on the overlapping-label problem from the cross-validation lesson and on the precision and recall ideas from the evaluation-metrics lesson.