Benchmarking Cryptocurrency Forecasting Models in the Context of Data Properties and Market Factors

Abstract

Cryptocurrency price prediction presents a significant challenge due to the inherent nonlinearity of the market. In this thesis, we assess the performance of thirteen time series forecasting models in predicting the prices of twenty-one different cryptocurrencies across four specific time frames. Our analysis centers on how data characteristics and market conditions affect the precision of these models and explores the implications of both broadening the scope of training data and extending the forecast periods. Our findings indicate that TBATS, LightGBM, XGBoost, and ARIMA consistently deliver the most accurate results. We identify key factors influencing prediction accuracy, including market trends, heteroskedasticity, volatility, and market capitalization. Additionally, the choice of time frame markedly affects all models’ predictive accuracy. Contrary to expectations, we observe that increasing the volume of training data does not necessarily enhance the performance of deep-learning and RNN-based models. Our thesis offers a comprehensive benchmark of forecasting models within the cryptocurrency context, underscoring the conditions crucial for improving prediction accuracy.