Replacing Fixed Indicator Weights with a Neural Return Predictor

Summary

This article describes adapting a Pine strategy that assigns fixed weights to technical indicators according to market state. The proposed workflow rebuilds indicators in Python, classifies market conditions, computes a weighted score, and feeds features into a small neural network to predict returns. Trading decisions then depend on prediction thresholds, with stop-loss and take-profit logic retained. The text describes periodic retraining and collecting samples from successive candles.

The author reports that training ran and mean squared error declined, while predictions showed some relationship to realized returns. Overall trading results were described as mediocre, with limited predictive accuracy attributed to a single-score input, noisy short-term futures prices, and possible overfitting. The code excerpt includes a larger technical-feature vector, so implementation details do not align perfectly with the prose description of a single input. No robust out-of-sample results or transaction-cost analysis are provided. The experiment is best read as a prototype, with feature quality, validation, online training, and risk management still requiring work.

Key ideas

• The experiment converts a market-state-dependent indicator score into an input for return prediction.
• A neural network’s predicted return is used to decide whether to open or reverse a position.
• The article reports declining training loss but only weak predictive association and mediocre overall results.
• Short-term price noise, limited features, small samples, and overfitting are identified as concerns.
• The prototype lacks robust out-of-sample and transaction-cost evidence.

Tags

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