Forecasting Volatility with Machine Learning and Rough Volatility: Example from the Crypto-Winter
Forecasting Volatility with Machine Learning and Rough Volatility: Example from the Crypto-Winter
This paper extends a volatility forecasting framework combining machine learning (LSTM) and rough volatility models to the cryptocurrency market during the 2022 crypto-winter. Using 5-minute intraday data from 213 coins on Binance, the authors compare parametric models (AR, HAR, RFSV, QRH) with non-parametric LSTM networks. They find that a universal LSTM trained on the pooled dataset outperforms traditional models calibrated per coin, suggesting a universal volatility formation process. The RFSV model, with only one parameter (Hurst index), also performs well, and incorporating the Zumbach effect via the QRH model yields similar performance with five parameters that are not asset-dependent. A key finding is the inverted asymmetric volatility in cryptocurrencies: positive returns increase future volatility more than negative returns, opposite to stock markets. This asymmetry is captured by the LSTM's sensitivity to past returns. The results confirm the robustness and universality of the volatility formation mechanisms across different asset classes, with implications for risk management and trading strategies in cryptocurrencies.
Highlights
- 1Extends a volatility prediction framework combining LSTM and rough volatility to cryptocurrencies during the 2022 crypto-winter.
- 2Shows that a universal LSTM trained on a pool of 213 coins outperforms traditional models like HAR and AR.
- 3Demonstrates that a parsimonious rough volatility model with Zumbach effect achieves similar performance with only five non-asset-dependent parameters.
- 4Discovers inverted asymmetric volatility in cryptocurrencies (positive returns increase volatility more than negative returns), opposite to stocks.
- 5Provides evidence for universality of volatility formation mechanisms across cryptocurrencies.
Methods
- MLong Short-Term Memory (LSTM) neural networks with input lengths of 7 or 30 days, using volatilities and returns as inputs.
- MRough Fractional Stochastic Volatility (RFSV) model with Hurst parameter estimation for volatility forecasting.
- MQuadratic Rough Heston (QRH) model incorporating Zumbach effect for parsimonious parametric forecasting.
- MLocal sensitivity analysis via gradient computation to study asymmetric volatility effects.
Results
- RLSTM models with return inputs and longer windows (30 days) significantly outperform HAR and AR models.
- RRFSV model outperforms HAR and is competitive with LSTM without returns, but underperforms LSTM with returns.
- RThe universal LSTM trained on all coins performs similarly to LSTMs trained on subsets (e.g., top 50 coins), supporting universality.
- RPositive returns increase future volatility more than negative returns for cryptocurrencies (inverted asymmetry), contrasting with stocks.
- RThe parsimonious QRH model with universal parameters achieves prediction performance close to coin-specific calibrated models.
Analyze Paper
Generate insights from "Forecasting Volatility with Machine Learning and Rough Volat...".