Choosing Sampling Frequencies to Reduce Microstructure Noise
Summary
The document asks when market microstructure noise becomes important and whether there is a sampling interval above which researchers can ignore it. One response presents five-minute sampling as a common compromise for intraday realized volatility: it can reduce noise-related bias while retaining information from high-frequency observations. It describes volatility signature plots, which compare realized variance across sampling frequencies to identify where estimates stabilize, and cites research that found little noise bias at five-minute intervals in its setting.
A second response cautions that this interval is not universal. The appropriate scale may vary with instrument liquidity, so it recommends comparing a naive volatility estimator, a range-based estimator, and a microstructure-aware estimator across reasonable intervals for each instrument. The estimates' convergence can guide a choice, with periodic reassessment if conditions such as tick size change. These are practical heuristics rather than a fixed threshold: realized variance is not robust to microstructure noise, and the sampling interval that works for one asset or purpose may not work for another.
Key ideas
- Five-minute sampling is a common starting point for limiting microstructure noise in realized volatility estimates.
- Volatility signature plots can show how realized variance changes across sampling frequencies.
- The suitable sampling interval depends on instrument liquidity and may differ across assets.
- Comparing multiple volatility estimators across candidate intervals can help identify a stable scale.
- Reassess the selected scale as market characteristics, including tick size, change.
Tags
Full text
# At what time step the microstructure noise start to kick in?
# At what time step the microstructure noise start to kick in?
When looking for papers on-line I often find things designed specifically to deal with micro-structure noise. I spent some time trying to understand / implement / test them for results that vary widely. I would be interested to know at below which timestep this micro-structure noise really starts to kick-in. Or conversely at above which timestep I can start to ignore these papers.
In Market Microstructure in Practice (Lehalle Laruelle Eds) the chronology of market speed start at 2MHz in the 70s and ends at 3.6Ghz at the starts of 2018. The SEC consulted market participants in 2010 about microstructure well after the GHz step that happened in the 00's. Does that mean that anything above the ms, s or minute is free from micro-structure noise ? (Or just that regulators were a bit slow on that concept ?)
## Answer by Pleb (score 4)
https://quant.stackexchange.com/a/69191
##### A 5-minute sampling frequency is a popular choice when avoiding the adverse effects of microstructure noise:
When dealing with eg. estimation of intraday volatility estimates there is a wide consensus that sampling at a 5-minute frequency will keep the bias induced by the microstructure noise to a minimum and still contain the added information inherent in high-frequency data.
The choice of a 5-minute frequency can be observed in the article of Hansen, P. R., & Lunde, A. (2006) that studies the market microstructure noise in high-frequency data and analyze its implications for the realized variance ($RV_t=\sum_{i=1}^n r_{i,t}^2$)*** under a general specification for the noise. They specifically construct volatility signature plots — by plotting frequency on the x-axis and realized variance on the y-axis — and observe for which frequency $RV_t$ stabilises (see Figure 1). Here, they find that the bias induced by microstructure noise is not present at a 5-minute sampling frequency for the realized variance estimator.
In conclusion, you can ignore the effects of microstructure noise in your data (and the papers dealing with this matter) when sparse sampling at a 5-minute frequency for most asset-classes.
While it might not be intraday volatility estimation you are dealing with, you can use the above as a rule of thumb for when microstructure noise becomes prominent in the data.
*** The realized variance estimator is not robust in the presence of microstructure noise.
## Answer by lehalle (score 2)
https://quant.stackexchange.com/a/69215
Not sure that I will go for a 5 min scale for any stocks; it really depends on the liquidity of the stock. For some 15 minutes is better, and for other 1h or 1:30 hours.
Of course it is easy to state this since "liquidity" is not well defined, and could be even be defined a circular way: this time scale is the one at which you "recover" liquidity. It is probably at this rhythm that one should take strategic trading decisions (i.e. trade scheduling) and keep the tactical ones (cross the bid-ask spread or not, wait further away in the book, etc) inside these natural time intervals.
In practice (to paraphrase the title of this "excellent" book... by the way, there is a second edition), I would implement three estimators of volatility
- the naive one
- the German-Klass (the minimum variance estimator using a polynomial formula of the OHLC)
- a microstructure one (you can go for the Uncertainty zones model of Rosenbaum et al. or for the Multiscale one of Zhang, Mykland and Ait Shalia).
And you compare them on each stock until you find a time scale at while they give similar results. That is not very complicated to code an automated scan of few reasonable scales and run it on your whole universe; you will end up with the correct time scale for each instrument. You can update this every year (or when the tick of the stock changes). I would moreover monitor Rosenbaum's $\eta$ to check the scale when it changes too much.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.