Authors: Athanasios-Fotios Athanasiou, Emmanouil Sofianos, Periklis Gogas, Theophilos Papadimitriou

Title: Directional Forecasting of Cryptocurrency returns using Support Vector Machines

Abstract

Cryptocurrencies are designed as a digital currency in the form of fiat money. Bitcoin and Ethereum are considered the two major such examples as they have the highest impact and market capitalization. The aim of this study is to directionally forecast the returns on Bitcoin and Ethereum. Efficient forecasting has important applications for investment portfolio managers.

The forecasting models were created using Support Vector Machines (SVM), a supervised learning methodology in the area of Machine Learning. In building the best forecasting model, both the linear and the radial basis function (RBF) kernels were used. The optimization of the model’s parameters was achieved through Cross-Validation, while a Coarse to Fine approach was used for parameter fine tuning.

We select the best Autoregressive model by evaluating 32 lags, in order to capture potential short-dependence. The initial 80% of the data are used for in-sample model optimization (training), while the remaining 20% is used to test our optimal models to out-of-sample data. Daily data from 11/9/2015 to 5/04/2018 were used.

For the Bitcoin, we reached a 73.6% and 75.87% in-sample and a 72.19% and 70.59% out-of-sample accuracy, using linear and RBF kernel respectively. In the case of the Ethereum, we achieved a 72.27% and 75.87% in-sample and a 68.45% and 70.59% out-of-sample accuracy, using linear and RBF kernel respectively. These models have shown a significant degree of generalization ability as that they are able to perform with similar accuracy both in-sample and out-of-sample.