[Paper Review] Explainable Deep Convolutional Candlestick Learner

Junhao Chen, Samuel Yen-Chi Chen+2 more|arXiv (Cornell University)|Jan 8, 2020

Market Dynamics and Volatility3 references4 citations

TL;DR

This paper proposes a GASF-CNN model for explainable candlestick pattern recognition using local search adversarial attacks to reveal how the model identifies patterns like Morning Star and Evening Star. The method demonstrates that the model learns human-like perception of candlestick patterns, achieving a 64.36% average attack success rate, confirming its interpretability and reliability for financial vision applications.

ABSTRACT

Candlesticks are graphical representations of price movements for a given period. The traders can discovery the trend of the asset by looking at the candlestick patterns. Although deep convolutional neural networks have achieved great success for recognizing the candlestick patterns, their reasoning hides inside a black box. The traders cannot make sure what the model has learned. In this contribution, we provide a framework which is to explain the reasoning of the learned model determining the specific candlestick patterns of time series. Based on the local search adversarial attacks, we show that the learned model perceives the pattern of the candlesticks in a way similar to the human trader.

Motivation & Objective

• To develop a deep learning model that recognizes candlestick patterns in financial time series with high accuracy.
• To address the lack of interpretability in existing deep learning models for candlestick analysis.
• To provide a framework that explains how the model identifies patterns by mimicking human visual reasoning.
• To validate the model's interpretability using adversarial attacks that reveal critical features in the input.
• To establish a foundation for trustworthy, explainable AI in financial trading systems.

Proposed method

• The paper proposes a GASF-CNN model that encodes time series data into Gramian Angular Summation Field (GASF) matrices for improved pattern representation.
• It applies a modified local search adversarial attack to perturb the GASF matrix by scaling diagonal elements within a range [0.8, 1.2] to test model robustness.
• The attack process iteratively modifies elements along the diagonal, preserving overall structure while testing model sensitivity to key pattern components.
• Perturbations are retained only if they cause a misclassification, ensuring adversarial examples are valid and meaningful.
• The attack is applied across multiple iterations with periodic resets to explore diverse adversarial directions.
• The method evaluates model behavior by analyzing which input features (e.g., specific candlesticks) are most critical for classification.

Experimental results

Research questions

• RQ1 How do deep convolutional neural networks perceive candlestick patterns in a way that aligns with human traders?
• RQ2 Can adversarial attacks based on local search reveal the critical features used by a GASF-CNN model in candlestick recognition?
• RQ3 To what extent does the model rely on the diagonal structure of the GASF matrix to identify multi-bar candlestick patterns?
• RQ4 Does the model’s decision-making process mirror human visual perception of patterns such as Morning Star and Evening Star?
• RQ5 Can the proposed attack framework serve as a reliable explanation tool for financial vision models?

Key findings

• The GASF-CNN model achieves a 64.36% average success rate in local search adversarial attacks, indicating high sensitivity to specific input perturbations.
• The attack success rate is highest for the Evening Star pattern (82.5%) and Morning Star (42.1%), reflecting differences in pattern complexity and robustness.
• The model primarily relies on the diagonal elements of the GASF matrix, corresponding to the last three bars, to identify multi-bar candlestick patterns.
• Adversarial examples show that small perturbations to the 8th and final bars can render patterns like Morning Star and Evening Star unrecognizable, confirming their critical role.
• The results confirm that the model learns to perceive candlestick patterns similarly to human traders, focusing on structural and contextual cues such as body size and wick length.
• The framework successfully explains the model’s reasoning, demonstrating that it captures human-like visual intuition in financial time series analysis.