Meta Labeling: Filtering Signals and Sizing Bets
Summary
Meta labeling adds a secondary classifier to a primary model that already proposes a trade direction or classification. The primary model is tuned for high recall, accepting some false positives; the secondary model then estimates whether those proposals are likely to be correct. In trading, its output can help decide whether to act and how much to wager, while the primary model retains control of the long or short side.
The article illustrates the setup with MNIST images of handwritten threes and fives. It labels primary-model predictions as correct or incorrect, trains the secondary model on the original features plus primary predictions, and combines their outputs. The reported out-of-sample confusion matrix shows fewer false positives and improved performance metrics. This is a toy classification demonstration rather than evidence of trading profitability. It does not establish that the approach improves live results, and its claims about reduced overfitting and more reliable outcomes are presented without a detailed trading test.
Key ideas
- A primary model can prioritize recall and propose the direction of a trade.
- A secondary classifier can filter primary predictions that are likely to be wrong.
- Meta labels mark whether the primary model's predictions were correct.
- The secondary model's probability can inform whether to trade and how to size a position.
- The MNIST example demonstrates classification mechanics, not trading profitability.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.