Prices of cryptocurrencies have shown large fluctuations. However, they have been progressively adopted as an asset class by financial institutions. A question naturally arises: is the crypto class exceptional in terms of its volatility dynamics and predictability?
In this study, we have selected instruments from a wide range of categories for comparison with cryptocurrencies. We show that from both the perspectives of persistence and mean reversion, cryptocurrencies is similar to instruments in other asset classes. We then examine volatility predictability using the heterogeneous autoregressive (HAR) model and Random Forests (RFs). In-sample and out-of-sample model performances are evaluated using the correlation coefficient between observed and predicted. In both cases, the correlation decreases with increasing prediction horizon. The out-of-sample correlation drops more rapidly than the in-sample case, and the RF model in general outperforms the HAR model. From the correlation-prediction horizon curves, we see that the crypto asset class behaves in line with instruments in other asset classes. We conclude that the volatility of cryptocurrencies is on average as predictable as for other instruments.
Persistence and mean reversion
In order to assess uniqueness or characteristics of crypto volatility, we have selected instruments from a wide range of categories for comparison. The evolution of volatility over time is complex, but from the vast literature on this subject, two properties are widely recognized: persistence and mean reversion. We investigate both properties for cryptocurrencies in the context of the reference instruments.
- We use PACF here as the manifestation of self persistence. The Bitcoin/USD cross (XBTUSD) has a close to 0.4 lag-1 value, comparable to the high values for SPX, GT10 and CL1. It takes about 10 days for the XBTUSD PACF to fall below a low reference level of 0.05, again comparable to other instruments.
- Increments of squared weekly returns against their starting values show decreasing trends, from positive increments when starting values are low to negative increments when starting values are high. The trends are also approximately linear, consistent with the Cox-Ingersoll-Ross stochastic volatility model.
Volatility predictability
- Predictability using HAR model: The correlation coefficients between observed and predicted decrease with increasing prediction horizon. Among the 10 cryptocurrencies, XBTUSD shows higher correlations than for other coins at most horizons. The fixed income instruments show higher correlations than for other instruments. For XBTUSD, its correlations are within the range for other asset classes.
- Predictability using Random Forests (RF): We explore RF to see if we can improve prediction accuracy by drawing information from other instruments and by relaxing the linearity assumption. Backtesting results show that RF outperforms HAR with only a few exceptions.
Conclusion
From the correlation-prediction horizon curves, we see that the crypto asset class behaves in line with many instruments in other asset classes. Their persistence and mean-reversion properties are also comparable. We conclude that the crypto volatility is on average as predictable as for other instruments.
Download the white paper to learn more about predictability of cryptocurrencies.