Exploring Model Uncertainty in Cryptocurrency Volatility Prediction:log-Transformation,Time-Zone Sampling,and Model Specification
This paper presents a comprehensive comparison of various models for predicting cryptocurrency volatility.The findings highlight that the rough volatil-ity model demonstrates robust and reliable performance in forecasting out-of-sample volatility across multiple periods.Conversely,the heterogeneous autoregressive(HAR)model shows relatively weaker results.However,the log-transformed HAR model exhibits superior predictive capabilities.Additionally,the study emphasizes the sig-nificance of selecting appropriate time zone divisions,considering the impact of dif-ferent time zones on cryptocurrency market volatility.To address model uncertainty in volatility modeling,the paper introduces the method of model averaging using least squares.The results indicate that model averaging outperforms alternative ap-proaches by effectively balancing the strengths and weaknesses of different models,ultimately enhancing the credibility and stability of predictions in the cryptocurrency market.The study underscores the importance of considering the unique character-istics and historical performance of cryptocurrency volatility when selecting suitable volatility models.Furthermore,it emphasizes the need for careful evaluation of model performance across diverse datasets and prediction targets to mitigate uncertainty arising from blind application.