How should I determine what frequency should I use for doing microstructure research using intraday data? For some reason, there seems to be general consensus of using 5 minute interval, but is there any advantage to this over higher frequencies if e.g. 1 minute data is available?

I think the answer is driven by asking yourself a few questions :

• If you are a practitioner at what frequency are you able to trade and want to trade ? (you are limited by this so no need to go to higher frequencies than that in any case)
• What effect(s) do you want to study ?
• What is a common frequency used by practitioners or academics for the question you have. Maybe which frequency is best becomes the question you want to study. You can form an opinion/intuition ex-ante as well.
• Potentially computing time and resources (tick by tick can grow to a lot of data very quickly)

I have seen people use from a 1 hour aggregation to the tick by tick full book level.

A few examples of academic paper with different aggregations : https://www.scheller.gatech.edu/directory/faculty/lee_s/pubs/Jump2-1-12.pdf (analyst estimates impact on price, done at 30 min intervals)

If you specify which type of questions you want to answer, others might have more specific suggestions on which interval to use.

• Thank you for your answer. I am trying to estimate the impact of certain events on intraday returns - I have data of the events and returns on one minute frequency and want to estimate the impact of event on both returns and volatility. Based on my results so far, I am thinking the reason why nobody hasn't really done similar things with one minute data is because at least GARCH seems to be really struggling with all the 0 returns that are present in one minute data, so I might have to switch to a different volatility modelling tool or switch to 5 minute data instead...
– Ana
Commented Mar 2, 2016 at 22:57
• If you want to study volatility I think you will need several periods to have a realized volatility metric so it would make sense to look at say 15-30 min volatility (computed with one minute returns). See for example finance.martinsewell.com/stylized-facts/volatility/…). Commented Mar 3, 2016 at 7:58
• Another observation : if you have "many" 0 returns at the minute level I would double check the data, unless you are working with relatively illiquid instruments it seems a bit odd, especially around an event... I might be off but it's worth double checking IMO. Commented Mar 3, 2016 at 7:59
• It is more of a market wide event, and many of the 0 observations are outside the event window, but are none the less important observations to quantify the variance outside of the event window in order to be able to quantify the change in volatility and how certain parameters affect this. Many 0 returns are indeed interesting, the market I am working with should be relatively liquid but the quality of data should be fairly decent - I should mention that I am using mid-prices in my calculations, which causes a lot more 0 observations wrt. close-prices that include bid-ask bounce in returns.
– Ana
Commented Mar 5, 2016 at 12:13
• Another problem GARCH really seems to be having, even more so than the 0s, is the outlier observtions during the event window. Thank you for the articles, I'll look into them!
– Ana
Commented Mar 5, 2016 at 13:51

I think it all about your choice of data you want to research on, this is because you can always switch to finer intervals if 5 min is not working for you. So high frequency will range between tick, 1 sec, 1 min and 5 min bars. As a researcher you have the autonomy to choose the data set you need given the research objectives that you have.