Meta-Labeling

Meta-labeling, from López de Prado, is a simple but powerful reframing: stop asking one model to do everything, and split the trade into two separate decisions. First, which way do I bet?, the side. Second, should I take this bet at all, and how much?, the size. A primary model (or even a plain rule) decides the side. A secondary model, the meta-model, decides the size by predicting whether the primary model is about to be right.

The elegance is that the secondary model's job is a clean yes/no question. Its label is binary: given that the primary took a position, was that position correct? So its training target is

$$m_{t} = \mathbb{1}\{\text{primary side} = \text{realized outcome}\},$$

which reads "1 if the primary model's direction matched what actually happened, 0 otherwise." The realized outcome usually comes from Triple-Barrier Labeling, did price hit the profit target before the stop? The meta-model then outputs $\Pr(\text{correct}\mid \text{features})$, a probability you use as a confidence filter and a size dial.

Why split the decision

Splitting the decision buys you three things:

• Higher precision without losing recall. Let the primary be aggressive, taking many positions (high recall, it catches most opportunities). The meta-model then vetoes the low-confidence ones, raising precision. Improving precision and recall separately tends to lift the overall F1 score more than forcing one model to balance both.
• Natural position sizing. The meta-model's probability maps straight to a bet size, big when $\Pr(\text{correct})$ is high, small or zero when low. That is a clean bridge to Position Sizing and fractional The Kelly Criterion.
• It bolts onto anything. The primary can be a machine-learning model, a technical rule, or even a discretionary/fundamental signal. Meta-labeling learns when to trust an existing signal without rebuilding it.

Worked example

Your primary rule goes long whenever a momentum filter fires. Over a test period it takes 100 long trades, of which 55 win, a precision of $55\%$. Decent side-picking, but you're acting on every signal, including a lot of marginal ones.

Now train a meta-model on features (volatility, spread, order-flow, time-of-day) with the binary label "did this particular long win?" Applied out-of-sample, it flags only the 60 trades where $\Pr(\text{correct}) > 0.5$. Of those 60, 45 win, precision jumps to $75\%$. The 40 trades it vetoed were mostly losers (10 winners, 30 losers), exactly the marginal bets you wanted to skip. You kept $45/55 \approx 82\%$ of the winners while cutting the losers roughly in half.

Then size with the probability instead of a flat stake: a trade with $\Pr(\text{correct}) = 0.8$ gets a larger position than one at $0.55$, feeding the probability into a fractional-Kelly sizing rule. Same primary signal, far better realized performance, purely from adding a size decision on top of the side decision.

Pitfalls

• Overlapping labels. Meta-labels inherit the overlapping label spans of Triple-Barrier Labeling, so they must be validated with purging and embargo, not naive k-fold, or the precision gain is an illusion.
• Threshold overfitting. The $\Pr(\text{correct})$ cutoff is itself a parameter; tune it on purged folds, not on the test set.
• Class imbalance. In trending or one-sided periods the "correct" class dominates; weight samples by uniqueness and check calibration of the probabilities before sizing on them.
• Feature leakage into the meta-model. The meta-model sees the primary's side; make sure that side is computed with information available at decision time only.

In interviews

Expect "how would you decide not just the direction of a trade but whether to take it and how big?" The crisp answer: a primary model picks the side, then a meta-model predicts the probability the primary is correct, and that probability filters low-confidence trades (raising precision) and sets the position size. Be ready to note that meta-labeling improves an existing signal rather than replacing it, that the probability feeds Kelly-style sizing, and that the labels overlap so validation must be purged and embargoed.