A DNN Stock Selection Workflow Using Seven Factors

Summary

This article introduces dense neural networks and backpropagation, then outlines an A-share stock selection experiment. A DNN uses connected layers of neurons; training sends predictions forward and propagates errors backward to update weights. The example uses seven factors to predict five-day returns, with missing-value handling, exclusion of extreme factor observations, and standardization before fitting. Its network has two hidden layers, dropout, and a linear output for the return estimate.

The described sample trains on 2010–2015 data and tests on 2016–2019 data. Hidden layer widths are 256 and 128, dropout is set to 0.1, and training runs for five epochs with mean squared error as the evaluation measure. The article says a backtest outperformed a benchmark, but the underlying chart, detailed performance statistics, factor definitions, and transaction assumptions are not available in the text. The result is therefore preliminary: feature selection, model design, and validation choices remain open, and the reported comparison alone does not establish robust out-of-sample performance.

Key ideas

• A dense neural network connects each layer to the neighboring layers and learns weights through backpropagation.
• The example predicts five-day stock returns from seven factors after cleaning and standardizing the data.
• The model uses two hidden layers, dropout, and a linear output layer.
• The reported benchmark comparison lacks detailed metrics and trading assumptions, so it is not enough to establish robustness.

Tags

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