CNN Candlestick Images for Next-Day Metals Direction

Summary

This project uses computer vision to classify the next trading day’s direction for selected precious and industrial metals. It turns fixed windows of daily candlesticks into 224-by-224 images and trains ResNet CNNs to predict whether the following day’s open-to-close move will be neutral, upward, or downward. The labels use a move threshold and a drawdown condition; input data are drawn from markets and related instruments that trade earlier, along with other prices and indicators.

The authors compare ResNet variants and two image window configurations, using a randomly held-out validation subset. They report that ResNet34 had the lowest error among the tested models and retain instruments with final error rates below their stated cutoff. Backtests and a short live period are described, including probability filtering and stop-loss use. The evidence is limited: the live sample covers few trading days, signals are unevenly distributed over time, and the text gives no detailed performance statistics or robust out-of-sample assessment. The method is a project example, not proof that the classifier will generalize across metals or market regimes.

Key ideas

• The project encodes fixed windows of daily candlesticks as images for CNN classification.
• Next-day direction is labeled from the following session’s open-to-close move, with neutral and directional classes.
• The tested ResNet models and image window sizes did not perform equally across instruments.
• Only instruments meeting a stated validation error cutoff were retained for trading signals.
• The reported live results cover a short period and provide limited evidence of generalization.

Tags

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