Please rotate your device to landscape mode to view the charts.

(Re-)imag(in)ing Price Trends

Journal: Journal of Finance

Date: 20231201

Author: Jiang, Jingwen; Kelly, Bryan; Xiu, Dacheng

Abstract:
We reconsider trend-based predictability by employing flexible learning methods to identify price patterns that are highly predictive of returns, as opposed to testing predefined patterns like momentum or reversal. Our predictor data are stock-level price charts, allowing us to extract the most predictive price patterns using machine learning image analysis techniques. These patterns differ significantly from commonly analyzed trend signals, yield more accurate return predictions, enable more profitable investment strategies, and demonstrate robustness across specifications. Remarkably, they exhibit context independence, as short-term patterns perform well on longer time scales, and patterns learned from U.S. stocks prove effective in international markets.

Link: Google Scholar

Background and Context

Research Focus

This study explores whether machine learning image analysis of stock price charts can predict future returns more accurately than traditional technical analysis methods.

Methodology

The researchers use a Convolutional Neural Network (CNN) to analyze stock price chart images containing open, high, low, close prices and trading volume from 1993-2019.

Innovation

Rather than testing predefined patterns like momentum or reversal, the CNN automatically identifies predictive price patterns from raw image data.

Superior Performance of CNN-Based Trading Strategies vs Traditional Methods

  • The CNN strategy (I5/R5) achieves more than double the performance of traditional technical trading strategies
  • Equal-weight portfolio Sharpe ratio of 7.15 compared to next best performer TREND at 2.92
  • Shows CNN's ability to identify more profitable trading patterns than human-designed indicators

Successful Application of CNN Across Different Time Horizons

  • CNN strategies maintain strong performance across different investment horizons
  • Shows strongest performance in weekly trading with declining but still significant results for longer periods
  • Demonstrates robustness of CNN pattern recognition across time scales

International Transfer Learning Success

  • CNN model trained on U.S. data performs better when applied to international markets than models trained locally
  • Shows 56% improvement in Sharpe ratio using transfer learning vs local training
  • Demonstrates global applicability of identified price patterns

Performance Across Market Size Categories

  • CNN strategy remains profitable when restricted to largest 500 stocks
  • Demonstrates robustness to liquidity constraints and trading costs
  • Shows practical applicability for institutional investors

CNN vs Linear Model Performance

  • CNN outperforms simpler linear models even when using the same image-based data representation
  • Demonstrates value of deep learning's ability to identify complex patterns
  • Shows importance of both data representation and model architecture

Contribution and Implications

  • Introduces novel application of deep learning to technical analysis that outperforms traditional methods
  • Demonstrates that price patterns have predictive power across markets and time horizons
  • Provides evidence that visual representation of financial data can improve predictive modeling

Data Sources

  • Performance comparison chart based on Table I showing weekly strategy results
  • Time horizon comparison based on Tables I and II showing weekly, monthly and quarterly results
  • International transfer results based on Table X comparing local vs transfer learning performance
  • Market size analysis based on Table IV comparing full sample vs top 500 stocks
  • Model comparison based on Table IX comparing CNN vs logistic regression approaches