Undefine Chart Patters: An Image-Based Approach

Academic literature of asset technical analysis and price series pattern recognition typically involves characterising heuristic chart patterns with piecewise linear functions and filtering characteristic price sequences with local regression. We propose an image-based approach to encode price series as a two-dimensional density matrix and analyse price geometries with convolution filters. The method scales local prices based on volatility within a sliding window and creates a density array via one-hot encoding and aggregation. By comparing convolutional network classifiers and autoregressive networks, we show that the image-based price representation improves pattern-related feature interpretability. However, forecasting future price movements remains challenging due to global insignificance of consistent patterns. Period-adaptive pattern filters are necessary to denoise series and separate patterns prior to image-based encoding and forecasting.