Combining CNN Feature Extraction with RNNs for Stock Signals

Summary

The article describes a hybrid neural network for classifying the direction of a stock’s daily candle. A one-dimensional convolutional network processes sequences of open, high, and low prices to extract features; a recurrent layer then models their temporal order. The author outlines data standardization, chronological train-test splitting, sequence creation, one-hot encoding, early stopping, and conversion of the trained model to ONNX for use in an MQL5 trading robot.

The example uses Tesla data and defines the positive class as a close above the open. The reported test accuracy is 54%, with class-level precision and recall also provided, so the result is modest and does not demonstrate a profitable strategy. The article warns that this combined architecture can be computationally costly and prone to overfitting, especially on a simple prediction task. It does not establish robustness across other assets, periods, or trading costs, and its signal should not be read as evidence of market-beating performance.

Key ideas

• A CNN can extract local features from price sequences before an RNN models their temporal relationships.
• The example classifies whether the daily close is above or below the open using Tesla price data.
• Chronological splitting and early stopping are used in the described training workflow.
• The reported test accuracy is 54%, which does not establish trading profitability.
• The author identifies overfitting and computational cost as key limitations of the hybrid model.

Tags

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