Andreas Park PRO
Professor of Finance at UofT
Discussion of Identifying High Frequency Trading activity without proprietary data
Paper by Bidisha Chakrabarty, Carole Comerton-Forde, and
Roberto Pascual
Discussion by Andreas Park
Northern Finance Association Meeting, September 2021
What's the paper doing?
- nice: the title says exactly what the paper is doing
- results:
- measures that we think should identify HFT identify HFT
- Hasbrouck and Saar's Strategic Runs seems to work best in capturing situations when HFTs are active
- measures that we think should identify HFT identify HFT
Comments on Measures
- Number of messages
- Number of cancellations
- Number of revisions \(+\) cancellations
- Orders cancelled within 0.1 sec
- Updates at the top-of-book
- StDev of midquote over 0.1sec intervals
- Number of responses within 0.1 sec of a top-of-book price improvement
- Strategic runs as in Hasbrouck and Saar
- Revisions seem special here and akin to other markets' cancellations
- not sure why this is called monitoring intensity vis-a-vis cancellations
How computed as HFT-specific?
- sounds like someone tries to fill a position and thus cancels and re-submits
- I can see this is an algo execution strategy, but why HFT? What is the HFT trying to do, which HFT does this fit to?
- Why not scaled by trade or volume?
- Argument given: Yao & Ye 2018 (cite missing from bib) say that sometimes high HFT has low message-to-trade
- but you have the data: test it!
- NSE seems to be a quasi-monopolist - changes the LP updating needs
- Number of messages
- Number of cancellations (exclude IOC cancellations?)
- Number of revisions and cancellations
- Orders cancelled within 0.1 sec
- Updates at the top-of-book
- StDev of midquote over 0.1sec intervals
- Number of responses within 0.1 sec of a top-of-book price improvement
- Strategic runs as in Hasbrouck and Saar
Results in a nutshell
Charles Jones: "All good papers have the key results in table III"
Results in a nutshell
Three correlations of interest
- among the public data measures
- among the HFT-data measures
- HFT-data vs public data
statistical pre-requisite: standardization, including intra-day seasonality
all between 40-80%
New step: squeezing out a little more
all previous measures are somewhat order-submission based,
so should (and do) catch message-intensive strategies
HFTs have many strategies
- some are message intensive, like market making
- others are not (cross-market arb)
what captures these?
seems like correlation to me ... how's that identified?
smaller things
- Maybe "mess" is not the best variable name...
- Report averages only - all stocks are for main index
\(\to\) large firms in a small sample (60 firms)
\(\to\) size differentiation isn't too meaningful
- limited order posting \(\not=\) liquidity provision
- most folks use stock-day panel in their analysis, and the correlation at this level is therefore most critical
Literature:
- Boehmer, Li, Saar (RFS 2018)'s PCA analysis links the first principal components to message-intensive strategies \(\to\) more discussion vis-a-vis your result is needed
- "Referee 2 walks into a bar and states that this is not the library he would have built": Devani, Tayal, Anderson, Zhou, Gomez, Taylor "Identifying Trading Groups" IIROC 14-0210 (2014) use PCA and SVM to identify characteristics of HFT with msg-intensive coming out on top (inventory not so much). Discuss relation.
- Ist NSE a quasi-monoplist? Affects the HFTs strategies that we might expect to see.
- tl;dr syndrome:
- who wins the horse race, which one should one use?
- What metric are you using to assess winning?
Overall
- Results confirm what we were all hoping for: using messages proxies well for HFT presence.
- Some more work is needed to identify aggressive HFT strategies.
@financeUTM
andreas.park@rotman.utoronto.ca
slides.com/ap248
sites.google.com/site/parkandreas/
youtube.com/user/andreaspark2812/
NFA 2021 Discussion
By Andreas Park
