Abstract
Bitcoin is the world’s largest cryptocurrency by market value and provides opportunity for investors to benefit from this relatively new asset class through portfolio diversification. Nevertheless, the price volatility of Bitcoin proves to be challenging to deal with especially for long-term investors.
As such, in this paper we aim to provide a machine learning solution that can benefit from the daily price swings by classifying whether to go long, go short, or stay away from the Bitcoin market in the next 24-hour window, using hourly historical data. Our approach involves a carefully thought-out design of features and target labels, which we believe contributed to the decent overall results of the study, where we present how the use of a simple artificial neural network architecture with two hidden layers is sufficient to achieve 318% total return over the three years covering January 2018 to December 2020. This translates to 61% CAGR, which is a massive outperformance relative to a buy-and-hold investment strategy’s 28% CAGR over the same period.
Notably, this outperformance was achieved with relatively lower risk, as drawdowns were contained despite major downturns in Bitcoin price in late 2018 and late 2019. Overall, we find the model we present here to be robust during Bitcoin crashes, but still has large room for improvement in terms of maximizing returns during upswings in Bitcoin price.
