Predictive Power of Machine Learning in Cryptoassets

The work attempts to forecast the sign of the price change for cryptoasset time series through classification. The main purpose is to find evidence concerning market efficiency of the cryptoasset markets, potential trading strategies, and differences between the modelled assets. Supporting vector machines, random forests, and multilayer perceptron models are used. An additional model aggre- gates the results of the previous three. Bitcoin, Ether, XRP, and Binance Coin are the modelled cryptoassets. The input variables include transformed daily closing prices up to five lags, trading volumes, volatility, and moving averages. Random forest models perform the best, followed by supporting vector ma- chines, and multilayer perceptrons. Aggregation does not produce improved forecasting performance. The two older assets, Bitcoin and Ethereum, are found to be less forecastable than the newer, Binance Coin and XRP. Dif- ferences between the assets exist as exhibited through forecastability. Higher classification accuracies are not found to imply better trading performance. JEL Classification C15, C69, G13, G14, G17 Keywords cryptoassets, machine learning, forecasting, cryptocurrencies Title Predictive Power of Machine Learning in Cryp- toasset