Assessment of Crypto Market Heterogeneity using Clustering Methods

Abstract

Since the introduction of Bitcoin, the cryptocurrency market has witnessed significant growth in both market value and the number of available cryptocurrencies. However, this market is characterized by heterogeneity, with a wide variety of cryptocurrencies exhibiting diverse technologies, user bases, and use cases. This heterogeneity makes it challenging for investors to make informed decisions.
To address this issue, this thesis aims to assess the heterogeneity of the crypto market using clustering methods, enhanced by back-testing. By incorporating back-testing into clustering techniques, the thesis examines the historical performance of clusters, thus providing a measure of their validity. The study utilizes a five-year dataset, considering different time periods to evaluate the clusters’ consistency.
The clustering process involves three methods: K-means clustering, dynamic clustering using histograms, and TADPole clustering based on time-series volatility. The data is transformed into distinct distributions of returns, enabling the application of each clustering method. The optimal number of clusters for K-means clustering is determined using cluster validity indices.
Furthermore, association tests are conducted between the clusters and various variables, including market capitalization, trading volume, financial ratios such as Beta and Sharpe Ratio, and heavy-tail behavior. Fisher’s test is employed to assess the significance of these associations, providing additional insights into the cluster characteristics.
In cases where time-series clustering fails, the Hierarchical Risk Parity (HRP) algorithm is utilized to cluster highly correlated coins. The resulting clusters’ weights are evaluated, and a two-year rolling analysis is performed on the Sharpe Ratio to assess portfolio performance.
Finally, the thesis applies subsequent predictive modeling, utilizing random forest algorithm to predict future cryptocurrency prices. The cluster assignments serve as features in the models, enabling an evaluation of their importance and the validity of the clustering methods. The models are backtested on historical data to measure their predictive performance.
By combining clustering methods, association tests, and predictive modeling, this thesis offers insights into the heterogeneity of the cryptocurrency market. The findings provide investors with valuable information for making informed investment decisions in this volatile and diverse market.

Date of Award	Aug 2023
Original language	American English
Supervisor	IBRAHIM Elfadel (Supervisor)