Which volatility model is best for predicting Realised volatility, especially in crypto?
The Heterogeneous Autoregressive model of Realized Volatility (HAR-RV) is the most widely recognized industry benchmark and reliable starting point for predicting realized volatility in cryptocurrency markets Verified Answer #1, Verified Answer #2, Verified Answer #3. While no single model is universally superior across all assets and market regimes, HAR-RV consistently outperforms traditional GARCH-family models at short-term horizons because it is specifically designed to model realized measures derived from high-frequency intraday data Verified Answer #4, Verified Answer #5, Verified Answer #6.
The HAR-RV model is effective in crypto because it captures "long-memory" properties and volatility persistence by decomposing volatility into daily, weekly, and monthly components Verified Answer #1, Verified Answer #2, Verified Answer #7. Empirical studies show it generates lower forecast errors, such as RMSE and QLIKE, for the majority of cryptocurrencies at 1-day ahead horizons Verified Answer #1, Verified Answer #5.
For practical application in cryptocurrency, enhanced versions of the HAR model are often preferred to account for the market's unique characteristics Verified Answer #4. Recommended variants include HAR-J, which incorporates jump components, and HAR-RS, which accounts for downside semivariance to manage asymmetric risk and sudden price shocks Verified Answer #2, Verified Answer #5. Log-HAR-RV-J is also cited as a superior practical upgrade for crypto specifically Verified Answer #4.
Alternative models may be used depending on data availability and asset liquidity Verified Answer #8. Realized GARCH is considered a strong alternative when a full return-volatility system is required Verified Answer #8. If only daily returns are available instead of intraday data, asymmetric models like GJR-GARCH or EGARCH are the best fallbacks, though they are generally less accurate for predicting realized volatility than HAR-type models Verified Answer #4, Verified Answer #8.
Machine learning and hybrid models represent the current frontier for maximizing accuracy in high-liquidity assets like Bitcoin Verified Answer #2, Verified Answer #5. Frameworks utilizing Random Forest, XGBoost, or Neural Networks (such as LSTM) can improve predictions by capturing non-linear relationships and regime switches that linear econometric models may miss Verified Answer #6, Verified Answer #7.