Shortlisting machine learning-based stock trading recommendations using candlestick pattern recognition

Abstract

Quantitative stock trading systems apply automated, data-driven models to invest in the stock markets. In the last two decades, the machine learning community has deeply explored the use of per-stock machine learning models to forecast next-day stocks’ directions and generate profitable trading signals. On the basis of the experience of discretionary traders, some promising attempts to enhance classifier performance by integrating the knowledge extracted from Japanese candlestick charts have been made. However, machine learning-based trading systems tend to generate an excessive number of false signals and do not necessarily consider the information provided by candlestick patterns in the appropriate manner. To alleviate these negative effects, this paper proposes to decouple the machine learning and pattern recognition steps so that the trading system can generate a reduced number of double-checked trading recommendations. Specifically, it proposes to selectively filter out the machine learning-based trading recommendations that are deemed as potentially unreliable according to the recognized graphical patterns. To this aim, it explores various alternatives to combine pattern recognition strategies with different machine learning models, including various shallow and deep supervised models and autoregressive techniques. The experiments, carried out on different market exchanges and under different conditions, demonstrate the effectiveness of the proposed approach in terms of return of investment and maximum drawdown of the trading system.