Imaging Time Series for Deep Embedded Clustering: a Cryptocurrency Regime Detection Use Case

Abstract

Following the recent trend of data-centric AI, we propose a clustering method to offer additional insight into precious and yet less-explored cryptocurrency price time series. While invaluable efforts have been conveyed in the domain of time series clustering, we integrate and harmonize some of the best practices in the field, namely Gramian Angular Field (GAF), Variational AutoEncoders (VAEs), and Deep Embedded Clustering (DEC). We use time series to image transformations as a preprocessing step for VAE to reduce dimensionality. After performing K-means clustering on VAE’s latent space, we provide DEC with cluster centroids from the previous step and retrain our network to do the clustering task. We evaluate the proposed method with the Bitcoin Tick-bar price dataset from 2017 onwards. Results demonstrate that our method leads to financially interpretable clusters and can improve Silhouette Score up to 10 percent compared to non-imaged time series.