14
$\begingroup$

When dealing with trade data, for example from TAQ, a common problem is that of determining whether a trade was a buy or a sell. The most commonly used classifier is the Lee-Ready algorithm (Inferring trade direction from intraday data, 1991). Unfortunately, this method is known to be inaccurate: Lee and Radhakrishna (Inferring investor behaviour: evidence from TORQ data, 2000) report that Lee-Ready incorrectly classifies 24% of the trades inside the spread.

How to improve on Lee-Ready's recipie? What are the best algorithms for trade classification?

$\endgroup$
  • $\begingroup$ On every trade there is both a buyer and a seller. When you talk of a buy or sell do you mean which side was the aggressor? ( removed liquidity). In the case of intraspread prints even that gets pretty blurry since you have to consider internalizers and pay for order flow models.... $\endgroup$ – PabTorre Sep 28 '13 at 5:28
  • $\begingroup$ Have a look at the VPIN papers for their method. $\endgroup$ – user2763361 Oct 2 '13 at 14:12
3
$\begingroup$

Don't use TAQ. The reporting times of the trades can be a few seconds delayed. Use the exchange feeds. There you can see which order crossed the spread. The only issue that you will run into is hidden orders. In that case you simply can't tell. (eg: The mkt is 30.00x 30.02 and then you see 30.01 trade. You don't have any way to tell if there was a hidden offer or bid.

For trades from dark pools they are really not possible to track since they all go to the ADF with a potential delay and there's no published book.

$\endgroup$
2
$\begingroup$

There's a lot of literature about Bulk Volume Classification (BVC), which takes a slightly different approach by estimating proportions of bars that are buy vs. sell (rather than classifying individual ticks).

See Easley 2015 for a comparison BVC to tick rules like Lee-Ready in 3 futures markets. Note that futures are probably the easiest market to correctly classify individual trades and thus represent the sternest test re: classification accuracy of BVC vs. tick rules.

I don't think there's a general consensus as to whether BVC outperforms aggregated tick rules in determining VPIN. You'd have to read the below and make your own decision.

For BVC: Panayides 2014

Against BVC: Andersen 2013, Chakrabarty 2014

$\endgroup$
2
$\begingroup$

Check out this job market paper "Inferring trade directions in fast markets" (3rd in this PhD thesis).

The author uses all possible information on quotes changes to match trades with the corresponding quote. The algorithm is said to outperform the common Lee and Ready (1991), Ellis et al. (2000) and Chakrabarty et al. (2007) algorithms.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.