High-Frequency Volatility Forecasting with Realized Measures and Data Activity
Summary
The discussion surveys ways to estimate or forecast volatility from high-frequency market data. It points to realized volatility research based on tick observations and notes that high-low data can offer an alternative when using every tick is impractical. A cited study addresses efficient volatility estimation from high-frequency data, while another respondent describes wavelet filtering as a computationally quick approach that produced reasonable estimates in their experience.
A separate suggestion uses the standard deviation of data packets per second as a possible subsecond signal. The rationale is that packet activity can track quote or trade frequency, which may rise during more volatile or stressed conditions. This is an anecdotal observation from one practitioner's experience, not evidence of a broadly validated forecasting rule. The thread provides starting points rather than a comparison of methods, and it gives no defined sampling choices, performance results, or guidance on microstructure noise and out-of-sample validation.
Key ideas
- Realized volatility methods aggregate information from high-frequency price changes to estimate volatility.
- High-low observations can provide an alternative to processing all tick data.
- Wavelet filtering is proposed as a computationally efficient approach, based on one contributor's experience.
- Variation in data-packet arrival rates may serve as a subsecond proxy for quote or trade activity.
- The suggestions are not compared with common evaluation methods or supported by reported out-of-sample results.
Tags
Full text
# How to forecast volatility using high-frequency data? # How to forecast volatility using high-frequency data? There is a large literature covering volatility forecasts with high-frequency tick data. Much of this has surrounded the concept of "realized volatility", such as: - "Realized Volatility and Correlation" (Andersen, Bollerslev, Diebold, and Labys 1999) - "Modeling and Forecasting Realized Volatility" (Andersen, Bollerslev, Diebold, and Labys 1999) Other efforts have looked at high/low data to improve the forecast without including all the tick data. Robert Almgrem has a nice lecture on the subject as part of his "Time Series Analysis and Statistical Arbitrage" course at NYU. What's the best way for forecast volatility using high-frequency data? Note: A similar question was previously asked on Wilmott. ## Answer by Meh (score 8) https://quant.stackexchange.com/a/877 Relevant paper: Efficient Estimation of Volatility using High Frequency Data (Zumbach, Corsi, and Trapletti 2002) http://www.olsen.ch/fileadmin/Publications/Working_Papers/020221-efficientVolEstimator.pdf ## Answer by ast4 (score 7) https://quant.stackexchange.com/a/875 Personally, I've dealt with volatility estimation using wavelets with HF data. Estimations seem reasonable and also it's fairly quick computation wise compared to other methods. There's quite a bit of literature on the subject, I would recommended starting off with An Introduction to Wavelets and Other Filtering Methods in Finance and Economics ## Answer by noah (score 2) https://quant.stackexchange.com/a/12832 std(PPS) PPS = Packets Per Second (wiki article: network packets) The standard deviation of packets per second received from a liquidity source are directly related to the number of quotes per second, or the number of trades per second occurring on that liquidity source. Thus, the higher the number of network / data packets per second, the more volatility there will be on that specific venue, or the market as a whole. std(PPS) can be used as a leading indicator of volatility in a sub-second trading environment. This answer is based on my personal experience of analyzing multiple liquidity providers (data sources) in real-time to predict / mitigate volatility or stressed markets.
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.