1D CNNs for Stock Prediction

Article BigQuant

Summary

This article introduces one-dimensional convolutional neural networks as a way to process financial sequences. A convolution applies shared filters across local time windows, allowing the model to learn patterns that can appear at different points in a series. Pooling can reduce sequence length by summarizing each window with a maximum or average. The article describes a two-layer Conv1D model for stock-price prediction, with input organized as samples, time steps, and features.

For comparison, it uses the platform’s default StockRanker shallow-learning template with matching training and prediction periods and features. The article reports annualized return increasing from 109% to 118% and says the Sharpe ratio also improved. These are presented as backtest results, but the text supplies no dates, detailed risk or cost assumptions, statistical uncertainty, or independent replication. It characterizes the model as a demonstration, so the reported comparison does not establish that the approach will generalize to other datasets or trading conditions.

Key ideas

One-dimensional convolutions learn local patterns in sequential data by applying shared filters across time.
The article presents Conv1D inputs as samples, time steps, and features.
Pooling can shorten a sequence by retaining window maxima or averages.
A two-layer convolutional model is compared with a shallow StockRanker strategy using matching data periods and features.
The reported backtest improves annualized return and Sharpe ratio, but the evidence lacks detail and independent validation.

Using One-Dimensional Convolutional Networks for Stock Prediction

Summary

Key ideas

Tags