Skip to content
All library documents

Using HAR-RV to Forecast One-Minute Bitcoin Volatility

Article Quant Q&A · Author: Jan Sila
Read the original

Summary

The document raises a modeling question about applying the heterogeneous autoregressive realized volatility model to forecast one-minute realized volatility in an active Bitcoin market. The proposed regressors are realized volatility measured over recent intervals ranging from the prior minute to the past hour, intended to represent volatility persistence across multiple time scales. The author uses log quote mid-prices, following the setup described in Corsi's HAR-RV paper, and reports having previously tried an EGARCH model.

The post asks whether the one-minute forecast horizon is unusually noisy or otherwise unsuitable, and why published work at that frequency may be scarce. It offers no answer, test results, or comparison of forecasting performance. Accordingly, it serves as a research question and model specification rather than evidence that HAR-RV works at this resolution. Microstructure noise, measurement choices, and the Bitcoin data sample would need to be examined before drawing conclusions.

Key ideas

  • The proposed task is forecasting one-minute realized volatility with a HAR-RV model.
  • The suggested predictors summarize realized volatility over several recent time scales.
  • The application uses Bitcoin quote mid-prices and follows a published HAR-RV setup.
  • The document asks whether one-minute volatility forecasts are too noisy but supplies no evidence resolving the question.
  • Model suitability depends on sampling, measurement choices, and empirical comparison.

Tags

CryptoVolatilityMarket microstructureBacktesting
Full text
# HAR-RV model for predicting 1-min volatility


# HAR-RV model for predicting 1-min volatility












I would like to use HAR-RV model (Heterogenous AutoRegressive - Realized Volatility) to predict a 1 minute realised volatility using the HAR model. As regressors I intend to use RV of previous minute, 10 minutes, 30 minutes and 1 hour to allow for some multiscaling as mentioned in Corsi's paper A Simple Approximate Long-Memory Model of Realized Volatility.

I was just wondering why is there no literature for 1-minute HAR? Is it heavily noisy and inappropriate model? Data I use is quite active Bitcoin series for which I'm using log-quote mid-prices as in the original Corsi paper. Before I used e-GARCH but looking at the results I do not suppose it is the best way forward.

Many thanks for suggestions.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.