Using a Neural Network to Predict Returns from a Pine Strategy Score

Summary

The document describes adapting a fixed-weight, multi-indicator trading strategy by feeding its combined indicator score into a small neural network. Indicator weights vary by classified market state, but are initially hand-set. The proposed model uses the score from the prior bar to predict the next bar’s return, and trades long or short when its predicted return exceeds stated positive or negative thresholds. Stop-loss and take-profit rules are also retained.

The author reports that the implementation collects samples from successive candles, retrains periodically, and shows falling mean squared error during training. They describe prediction correlation with realized returns as weak and overall profitability as mediocre, attributing limitations to the use of a single feature, noisy short-term futures prices, and unstable samples. The document offers a practical implementation account, but does not provide enough detail about validation design, out-of-sample testing, costs, or risk-adjusted results to establish that the model has predictive or trading value. The reported observations are specific to this experiment.

Key ideas

• The approach replaces fixed trade thresholds with neural-network predictions based on a weighted indicator score.
• Market-state rules determine indicator weights before those scores are passed to the model.
• The model is trained to predict the next candle’s return from the previous candle’s score.
• Trading thresholds trigger directional positions, with stop-loss and take-profit mechanisms retained.
• The author reports weak predictive correlation and mediocre profitability, citing limited features and noisy data.

Tags

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