A spectral relevance analysis approach to pattern recognition of financial time series

• High-frequency Bitcoin time series are encoded as Gramian Angular Fields for deep learning-based forecasting. • Spectral clustering of relevance maps uncovers distinct patterns with superior forecastability. • Introducing a novel methodology for identifying patterns associated with excess returns using Spectral Relevance Analysis. Understanding patterns in financial time series is crucial for improving prediction accuracy in algorithmic trading and risk management. This paper presents a novel AI-based computer vision approach for classifying financial time series. Historical price sequences are transformed into Gramian Angular Difference Field (GADF) images and fed into a convolutional neural network (CNN) for pattern recognition. To interpret the CNN’s decision-making process, we apply Spectral Relevance Analysis (SpRAy), enabling the identification of distinct clusters based on relevance maps. Clustering the images according to their relevance profiles reveals groups with significantly higher predictive performance compared to the full dataset. The corresponding relevance patterns highlight favorable price movement structures and are identified via the associated clusters.