Training Convolutional Networks on Trading Chart Images

Summary

The article presents a computer-vision workflow for classifying market chart screenshots with a convolutional neural network. It defines buy and sell image categories around directional price moves and daily extremes, then describes collecting hourly screenshots from MetaTrader 5, dividing labeled images into training and validation sets, and capturing a separate sequential test set. The network processes chart images rather than a long table of indicator values, and the author discusses inspecting intermediate feature maps to understand what visual patterns the model detects.

That inspection reveals that some chart features may be visually confounded, prompting a suggestion to add a flat-market category and refine labels. The article reports dataset sizes and describes a simple optimization workflow, but the provided material gives no out-of-sample classification metrics or evidence of profitable trading. Screenshots encode chart formatting and selected indicators as well as price, so results may depend on labeling choices and image setup. Turning model scores into a trading strategy requires additional validation and optimization.

Key ideas

• The proposed workflow uses chart screenshots as inputs to a convolutional image classifier.
• Training labels distinguish upward and downward moves and selected daily extremes.
• The author separates labeled images for training and validation and captures sequential screenshots for testing.
• Visualizing feature maps can expose confusion between chart patterns and motivate new classes.
• The article provides no reported out-of-sample metrics or demonstrated trading profitability.

Tags

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