Building a CNN–LSTM Model for A-Share Stock Selection
Summary
This tutorial outlines a machine-learning workflow for selecting Chinese A-share stocks with a convolutional neural network. It describes defining a stock universe and input factors, extracting and splitting data, turning observations into rolling sequences, training a model, generating predictions, and evaluating a strategy through backtesting. The example architecture reshapes factor inputs for a Conv2D layer, transforms the result for an LSTM layer, applies dropout, and uses a dense output for an up-move probability.
The article also explains layer roles and configuration choices, including weight initialization, activation functions, input dimensions, convolution settings, and dropout. It gives no performance results or detailed dataset, labeling, validation, or trading-cost methodology, so it serves as an implementation overview rather than evidence that the approach is profitable. Its descriptions of some neural-network settings are simplified, and the example’s feature and model choices would need independent validation against leakage, overfitting, and changing market conditions.
Key ideas
- A CNN stock-selection workflow proceeds from defining a universe and factors to training, prediction, and backtesting.
- Rolling windows convert factor observations into sequences that can be passed through a neural network.
- The example combines convolutional processing, an LSTM layer, dropout, and a dense probability output.
- Initialization, activations, layer dimensions, and convolution settings affect model behavior.
- The document provides no reported performance evidence or detailed safeguards against leakage and overfitting.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.