Forecasting Cryptocurrency Returns with Dynamic Model Learning: An Analysis of Cryptocurrency Valuation

This paper examines the factors influencing cryptocurrency valuation. Given the characteristics of cryptocurrencies as assets, we utilize the trading volume of the corresponding cryptocurrencies, the US dollar index, the implied volatility of the KOSPI return, and the gold futures price as the factors of the cryptocurrency valuation model. In addition to the use of exogenous variables as predictors perceived as having power to forecast dynamic movements of returns, the forecasting model includes other cryptocurrency returns in the system to reflect dynamic co-movements. We use the Bayesian dynamic learning procedure that switches over multiple forecasting models at each point in time. The best performing model with a different group of predictors changes over time very frequently, meaning that not a single model seems dominant as the optimal forecasting model. While finding an optimal forecasting model is important, it is more practical to adopt a forecasting model that enhances portfolio gains from an investor’s perspective. To investigate the gains from the optimal portfolio relative to the benchmark portfolio, we calculate the Sharpe ratio. The benchmark forecasting model with no explanatory variable produces the lowest Sharpe ratio from all the forecasting models considered. Compared to the benchmark forecasting model, the inclusion of any predictor among the exogenous predictors increases the annualized Sharpe ratio. It is vindicated that our methodology of selecting the optimal forecasting model is equivalent to the optimal portfolio selection with the highest portfolio gains.