Forecasting Cryptocurrency Time Series Using Adaboost-Based Ensemble Learning Techniques

Abstract

In the domain of time series, cryptocurrency data is one of the most complex data available with its inherent volatility and non-linearity, thus making the task of analysis and prediction exceptionally challenging. The econometric models such as ARIMA and ARMA fail to capture the non-linearity of data putting forth the need to adopt other models for forecasting. In this paper, data of five cryptocurrencies-Bitcoin, Ethereum, Litecoin, XRP, and Stellar is analyzed in an effort to predict the next day closing prices of the respective cryptocurrencies. This goal is achieved by applying the ensemble learning technique of AdaBoost to boost the weak learners namely MLP, ELM, SVR, and LSTM all of which individually suffer from the problem of overfitting. Comparing the results produced by these combinations with the individual techniques, it can be proved that boosting gives significantly better performance accuracy as compared to the individual learning methods. Adaboost-LSTM gives the minimum MAPE of 3.6569, 6.1932, 13.1040, 11.0626, and 5.1058 for Bitcoin, Ethereum, Litecoin, XRP, and Stellar, respectively. Also, the consistency of the results produced was tested by applying the same techniques on the opening prices of cryptocurrencies. Adaboost-LSTM combination produced the minimum MAPE of 2.0087, 2.5135, 8.8680, 2.0351, and 3.8633 for the five cryptocurrencies, respectively.