Machine Learning for Bitcoin: Insights from Tick-Level Trades and Limit Order Book Microstructure

This paper uses machine learning models to study the high frequency trading of Bitcoin. Our out-of-sample empirical tests show that machine learning models, which use tick-level trade and limit order book information, outperform the auto-regressive benchmark models using only past log returns of bitcoin prices. A trading strategy using the sign of the predicted log returns as the buy/sell signal each period produces a significant higher return than the buy-and-hold strategy of investing in Bitcoin. These findings suggest that microstructure information has predictive content and that machine learning can extract it more effectively than linear benchmarks.