Cryptocurrency Portfolio Construction Using Machine Learning Models

Abstract

Multivariate time series prediction is one of the main challenges and has been widely studied in quantitative finance. Prediction outcomes are the prerequisites for active portfolio construction and optimization and play a significant role in developing an efficient trading strategy. Since inception, cryptocurrencies have broad market acceptance and constitute an asset class characterised by high returns, high volatilities, usage, transaction speed and low correlations. Many hedge fund managers, proprietary trading desk in large banks and boutique trading firms has spent considerable efforts in forecasting cryptocurrency prices. This study is about the portfolio construction of nine most capitalised cryptocurrencies: binancecoin, bitcoin, bitcoincash, chainlink, EOS, ETH, Litecoin, MCO and XRP (More details on cryptocurrencies and the market capitalization can be seen in any crypto currency exchanges like binance, Upbit, Bitfinex, Coinbase). ARIMA, convolutional neural network and long short-term memory methods are proposed to forecast the cryptocurrency prices. In addition, multiple portfolios are constructed using equal weighted portfolio, modern portfolio theory, cointegrated pairs, Kelly criterion and risk parity which make financial or economic sense and then portfolio performance measures are used thereby determining the best portfolio to hold. We find that the portfolio constructed using cointegrated pair has outperformed and the annualised returns are further maximised using a pair trading strategy.