This paper introduces a novel approach to algorithmic trading by directly using 2-D stock bar chart images as input to a deep Convolutional Neural Network (CNN), bypassing traditional time series data analysis. The proposed CNN-BI model represents one of the first attempts to apply computer vision techniques to raw financial chart images for trading decisions, marking a significant departure from conventional quantitative finance methods that typically rely on numerical time series data.

The methodology involves generating sliding windows of 30-day bar chart images from Dow 30 stocks and training a deep CNN model specifically designed for this task. The researchers tested their approach across two distinct periods (2007-2012 and 2012-2017) to evaluate performance under different market conditions. Results demonstrate that the CNN-BI model successfully outperformed the Buy and Hold strategy, with particularly strong performance in trendless or bear markets. As a preliminary study, the authors acknowledge room for improvement but highlight the promising nature of this unconventional approach and its potential for integration into ensemble trading systems combined with other strategies.