Enhancing Model Explainability in Financial Trading Using Training Aid Samples: A CNN-Based Candlestick Pattern Recognition Approach

Abstract

Candlestick pattern recognition is a widely adopted technique in financial trading, leveraging visual patterns to analyze price movements. Deep Convolutional Neural Networks (CNNs) have exhibited remarkable accuracy in this domain. However, the increasing demand for transparency and explainability in CNN-based models raises concerns regarding their applicability in trading decision-making. This paper addresses these concerns by presenting a framework that enhances the explainability of CNN-based candlestick pattern recognition models. Our approach introduces an innovative data augmentation method to generate training aid samples, facilitating the model’s learning process within human domains. By incorporating this framework, traders gain valuable insights into the decision-making process, comprehending the rationale behind the model’s predictions. Our proposed approach exposes the inherent “black box” nature of CNN-based models, improving their interpretability and empowering traders to make well-informed decisions based on transparent and understandable insights. This advancement holds significant potential for enhancing decision-making processes in financial trading and fostering trust among traders.