Encoding OHLC Candles as Images for CNN Return Classification

Summary

The article describes turning rolling OHLC windows into sparse, binary candlestick images for a convolutional neural network. Each sample uses 32 time steps, normalized and mapped onto a 128-by-128 grid; three columns represent each candle’s open, high-low range, and close, with a fourth column as spacing. The model uses convolution, pooling, dropout, and a dense output layer to classify whether the following five-day return is positive.

The experiment uses 100 older-listed stocks, samples windows every eight bars, and separates training and evaluation data from a test period beginning in 2015. Training and evaluation appear stronger than out-of-sample performance, and filtering training examples to the most extreme return groups provides little improvement. The author identifies image sparsity as a possible limitation: active pixels average under five percent of the grid. The write-up is exploratory and supplies no convincing evidence of predictive performance or detailed controls for leakage, costs, or broader robustness.

Key ideas

• OHLC windows can be rendered as binary candle images and supplied to a CNN.
• The experiment labels each sample by the sign of its subsequent five-day return.
• The described model combines convolution and pooling layers with dropout and a dense output.
• Reported training and evaluation behavior does not carry over well to the later test period.
• Sparse active pixels are proposed as one possible reason for weak out-of-sample results.

Tags

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