Meta-Labeling Machine Learning for Stop-Loss and Take-Profit Signals

Summary

This project outlines a process for predicting trade exits with machine learning. It samples market data into dollar bars, estimates volatility from each bar’s high and low using Parkinson’s method, and generates initial buy or sell signals from a moving-average crossover combined with a return-strength threshold. Stop-loss and take-profit outcomes are labeled by checking which of two horizontal return barriers or a vertical time barrier is reached first; the vertical barrier is set at the completion of two dollar bars.

The author compares predicting the directional signal directly with meta-labeling, which predicts whether to take a base strategy signal. Logistic regression using the latter approach reportedly reached 64.85% accuracy, while direct classification had 54%; the conclusion gives a separate overall accuracy figure of 65.86%. These figures are reported without dataset details, out-of-sample validation, costs, or risk-adjusted performance, so they do not establish trading profitability or robustness. The project itself notes that further study could improve the model.

Key ideas

• The workflow samples prices into dollar bars and estimates volatility with the high-low Parkinson measure.
• A moving-average crossover strategy supplies candidate trade directions.
• Exit labels depend on which of the upper, lower, or vertical barriers is reached first.
• Meta-labeling is used to classify whether to act on the base strategy’s signal.
• Reported classification accuracy does not establish net profitability or out-of-sample reliability.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.