Building a Stock-Ranking DNN Strategy with Price and Volume Factors

Summary

This report describes adapting a platform’s deep neural network template for stock selection. The model uses 98 price and volume factors and three fully connected layers. The author uses a Calmar ratio label, about ten years of training data, and a universe drawn from sectors recently reaching new highs. The backtest rotates among four stocks every five days, with no market timing overlay, and reserves a later period as a test set.

The report offers implementation choices rather than a controlled comparison: it does not provide detailed performance results for the author’s experiments, and the cited platform result is reported without enough information here to assess its robustness. The author emphasizes fixing model parameters for reproducibility, preserving recent data for testing, and the risk of overfitting or underfitting. Suggested next steps include trying classification labels, changing network settings, revising risk controls, and testing higher frequency inputs. The central caution is that strong historical backtest performance does not establish generalization to future markets.

Key ideas

• The strategy applies a three-layer fully connected network to 98 price and volume factors for stock selection.
• The author uses a Calmar ratio label and roughly ten years of training data.
• The stock universe is filtered to sectors that have recently reached new highs.
• The backtest rotates among four holdings every five days and omits a market timing overlay.
• The report stresses reproducibility and out-of-sample testing because neural networks can overfit.

Tags

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