Convolutional Neural Networks for Trading Pattern Recognition

Summary

This article explains convolutional neural networks as a way to recognize patterns in price-chart data. It describes how convolution filters scan input sequences to produce feature maps, while subsampling retains local maxima or averages to reduce dimensionality and noise. The filters are learned during training rather than specified in advance. The article also outlines backpropagation through subsampling and convolution layers, then describes assembling those layers with a fully connected network for a final prediction.

The author implements and tests a convolutional model against a fully connected network. The reported comparison shows slightly lower prediction error and better target hitting for the convolutional model, with fewer but more target-proximate signals; the training times are similar. The proposed explanation is that convolution and subsampling preprocess chart data and reduce connections needed by the final perceptron. The evidence is specific to the article’s experiment, and the text does not establish that the result generalizes across markets, data, or trading conditions.

Key ideas

• Convolution filters scan price data for learned local patterns and produce feature maps.
• Subsampling compresses feature maps and can reduce sensitivity to the exact location of a pattern.
• Backpropagation through pooling routes gradients according to the pooling rule, while convolution weights are updated from input and error maps.
• The reported test modestly favors the convolutional model, but its findings are limited to the experiment described.

Tags

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