The Trouble with VPIN
Skip to content
Finance & Accounting Apr 2, 2012

The Trouble with VPIN

New economic indicator fails to live up to its promises

Based on the research of

Torben Andersen

Oleg Bondarenko

Stock markets in the United States on May 6, 2010, were not having a good day. By early afternoon, concerns over the European debt crisis and an upcoming jobs report had driven most major indices solidly into negative territory. But as bad as things looked, they were about to get a lot worse.

Add Insight
to your inbox.

We’ll send you one email a week with content you actually want to read, curated by the Insight team.

At 2:41 PM, prices for E-mini S&P 500 futures — the world’s most liq­uid equi­ty index con­tract — start­ed plung­ing. By 2:44 PM, algo­rithms used by high-fre­quen­cy traders to buy and sell that con­tract were going crazy, sell­ing more than they were buy­ing and threat­en­ing to vapor­ize liq­uid­i­ty. Mere sec­onds lat­er, those same algo­rithms bought and sold over 27,000 con­tracts in just 14 sec­onds, yet net­ted only 200 addi­tion­al con­tracts. The mar­ket was going haywire. 

One sec­ond after that fren­zy, trad­ing in E-mini S&P 500 futures was halt­ed for five sec­onds, forc­ing the com­put­ers to take a breather. By 3:00 PM, mar­kets recov­ered from the crash. In just 20 min­utes, the Dow Jones Indus­tri­al Aver­age had lost and then regained near­ly 1,000 points.

The inci­dent had tak­en a psy­cho­log­i­cal toll. Investors were spooked. And worse, no one seemed to know how it had hap­pened. It would take the Secu­ri­ties and Exchange Com­mis­sion and Com­mod­i­ty Futures Trad­ing Com­mis­sion five months to release a report that drew only ten­ta­tive con­clu­sions. Since then, some pre­ven­ta­tive mea­sures have been put in place but many experts are not con­vinced they are enough.

Imag­ine people’s relief when a new method was announced that could pre­dict immi­nent flash crash­es. Devised by two well-respect­ed econ­o­mists and the research head of a hedge fund, the mea­sure, called VPIN, or vol­ume-syn­chro­nized prob­a­bil­i­ty of informed trad­ing, mon­i­tors imbal­ances in trad­ing — when sell­ers out­num­ber buy­ers or vice ver­sa — and pur­ports to peak before prob­lems arise. Its inven­tors hail it as supe­ri­or to exist­ing mar­ket indi­ca­tors like VIX, the wide­ly watched volatil­i­ty index.

The trio behind VPINDavid Easley and Mau­reen O’Hara, both econ­o­mists at Cor­nell Uni­ver­si­ty, and Mar­cos López de Pra­do, head of high-fre­quen­cy trad­ing at Tudor Invest­ment — believe the mea­sure has the poten­tial to become a crit­i­cal finan­cial indi­ca­tor and have filed for a patent. Fur­ther­more, they are urg­ing reg­u­la­tors to use VPIN as a watch­dog signal. 

There is every rea­son to think that might hap­pen — in addi­tion to her posi­tion at Cor­nell, O’Hara also serves on a pan­el con­vened by the SEC and the CFTC to inves­ti­gate the flash crash. She and her co-authors strong­ly believe that VPIN could alert mar­ket reg­u­la­tors to an impend­ing crash like that which occurred on May 62010.

The prob­lem is, not every­one agrees.

Dis­sect­ing VPIN

Tor­ben Ander­sen is one of those peo­ple. His research focus­es on mar­ket volatil­i­ty and asset pric­ing, two fac­tors that are cen­tral to under­stand­ing the flash crash. His inter­est in pecu­liar mar­ket events is what led him to pick up the VPIN paper.

It has pol­i­cy rel­e­vance,” says Ander­sen, a pro­fes­sor of finance at the Kel­logg School of Man­age­ment, of the work­ing paper. What also piqued his inter­est was the authors’ claim that VPIN could pre­dict mar­ket imbal­ances and short-term volatil­i­ty bet­ter than VIX. I have oth­er papers where I crit­i­cize VIX,” he says, but I still think it’s fair­ly good. Maybe you could make it a lit­tle bit bet­ter, but it’s good.”

So I start­ed read­ing it,” Ander­sen says of the VPIN paper. It’s just — I can’t get my hands on this thing. It’s such a com­pli­cat­ed beast. Not in its con­struc­tion, but in its mix­ing of all these dif­fer­ent con­cepts … And when we start­ed look­ing at it sys­tem­at­i­cal­ly, in terms of fore­cast­ing it per­forms much worse than VIX.”

In say­ing we,” Ander­sen is refer­ring to him­self and his co-author, Oleg Bon­darenko, a pro­fes­sor at the Uni­ver­si­ty of Illi­nois at Chica­go. Togeth­er, the two took VPIN apart to math­e­mat­i­cal­ly ana­lyze each com­po­nent. When they were fin­ished, they con­clud­ed that VPIN was not bad, per se, but it could not do what its cre­ators had claimed. We can’t com­plete­ly get the results they got,” Ander­sen says.

To cal­cu­late VPIN in its sim­plest form, you group con­sec­u­tive­ly trad­ed con­tracts — say 50,000 in a row — into bins, regard­less of time or date. Group­ing sequen­tial trades in that way is called trad­ing time, and depend­ing on mar­ket vol­ume it can vary sub­stan­tial­ly with respect to clock and cal­en­dar time. In the next step you ana­lyze how many min­utes those 50,000 trades spanned, down to one-minute incre­ments, also known as time bars. If trad­ing is hap­pen­ing at a furi­ous pace, it is pos­si­ble all 50,000 trades in a bin could be squeezed into one minute, or time bar. 

After iden­ti­fy­ing the indi­vid­ual time bars, you assign each bar a buy” label if there were more con­tracts bought than sold in that span or a sell” label if more were sold than bought. Time bars labeled as buy” are val­ued +1, while time bars labeled sell” are val­ued –1. You then con­struct a vol­ume-weight­ed aver­age of the buy-sell indi­ca­tors for the time bars and take the absolute val­ue of that num­ber. Final­ly, you merge that bin with the 50 bins pre­ced­ing it, per­form some more math­e­mat­i­cal wiz­ardry, and presto — you have cal­cu­lat­ed VPIN for that minute.

Remov­ing Bias

Accord­ing to Ander­sen, the prob­lems with VPIN are numer­ous. One issue is the way in which it mix­es trad­ing vol­ume and time. How many con­tracts that are trad­ed in a minute is very high­ly cor­re­lat­ed with volatil­i­ty. When volatil­i­ty is high, peo­ple trade more, and so these min­utes will con­tain many more con­tracts. As a result, there will be many less min­utes in the vol­ume buck­et, and that will bias the mea­sure towards the extreme in a com­plete­ly mechan­i­cal way,” he says.

The con­fla­tion of vol­ume and time also caused anoth­er prob­lem. Because trades are first grouped sequen­tial­ly and then next by reg­u­lar clock time, the sep­a­ra­tion between two days’ trad­ing ses­sions is obscured. For exam­ple, if there are not enough trades from one day to com­plete a group, then trades from the next day are used until the prop­er num­ber, say 50,000, is reached. 

As a result, VPIN is high­ly depen­dent on when exact­ly you start count­ing trades. If you start count­ing one day lat­er than some­one else, your groups will con­tain dif­fer­ent trades and your VPIN will be dif­fer­ent. It is also vital­ly impor­tant that you have all the trades for that time peri­od. Any miss­ing trades also will shift the con­tents of the groups, poten­tial­ly lead­ing to very dif­fer­ent results.

At first, Ander­sen and Bon­darenko could not repro­duce Easley, O’Hara, and López de Prado’s find­ings. We had to start in ten or fif­teen dif­fer­ent places in the past in order to repli­cate their results,” he recounts. Ander­sen and Bon­darenko also real­ized they were using a dif­fer­ent data source than Easley, O’Hara, and López de Pra­do did. Upon inspec­tion, it was evi­dent that our trad­ing vol­ume was on aver­age a lit­tle bit big­ger than theirs,” Ander­sen says.

The data both groups of researchers used are trades of E-mini S&P 500 con­tracts on the Chica­go Mer­can­tile Exchange — the same instru­ment that pre­cip­i­tat­ed the flash crash. Easley, O’Hara, and López de Pra­do obtained their data from the real-time data feed of a hedge fund, while Ander­sen and Bon­darenko received his­tor­i­cal data direct­ly from the CME Group. I’ve sub­se­quent­ly spo­ken to the guys at the CME Group. This par­tic­u­lar con­tract is only trad­ed elec­tron­i­cal­ly, and all the trades are record­ed in their sys­tem,” Ander­sen recounts. That’s the only com­plete his­tor­i­cal record for this data.”

figure-1-small.png Fig­ure 1. Minute-by-minute data for the E-mini S&P 500 futures index lev­el, the VPIN mea­sure con­struct­ed from one-minute data, the S&P 500 volatil­i­ty index, VIX, and the vol­ume of trad­ed con­tracts of the E-mini S&P 500 futures on the CME for May 6, 2010. Ver­ti­cal green lines indi­cate the tim­ing of the flash crash.”

When Ander­sen and Bon­darenko were final­ly able to find the prop­er start­ing point, they ran into oth­er prob­lems. The most alarm­ing was that VPIN spiked after the crash, not before, hint­ing that it may be a reac­tive met­ric rather than a pre­dic­tive one (Fig­ure 1). Also trou­ble­some was that VPIN did not hit an all-time high around the time of the flash crash. In fact, it crossed the same thresh­old as dur­ing the flash crash on at least two oth­er occa­sions, nei­ther of which cor­re­spond­ed with an errant market.

Revamp­ing VPIN

Despite his reser­va­tions, Ander­sen does not think VPIN is a fatal­ly flawed mea­sure. With some changes, he says, there is some encour­age­ment that you might be able to say some­thing use­ful” with VPIN. If you mea­sured more sen­si­bly, it lines up much more with what actu­al­ly hap­pened” in the flash crash, he adds.

By tak­ing the absolute val­ue of trad­ing imbal­ance each minute, VPIN ignores infor­ma­tion about the direc­tion in which the mar­ket is moving.

The first thing Ander­sen and Bon­darenko sug­gest is using signed mea­sure­ments, both with­in and across the time bars. By tak­ing the absolute val­ue of trad­ing imbal­ance each minute, VPIN ignores infor­ma­tion about the direc­tion in which the mar­ket is mov­ing. It may also pre­vent alter­nat­ing peri­ods of buy­ing and sell­ing from can­cel­ing each oth­er out.

For exam­ple, if the sell­ing indi­ca­tor out­paces the buy­ing indi­ca­tor over one vol­ume buck­et by a mar­gin of 0.5 but revers­es by the same amount over the next, both buck­ets are scored 0.5 by VPIN. The result is an aver­age imbal­ance of 0.5, despite the over­all per­fect bal­ance of buy­ing and sell­ing. In con­trast, an unsigned VPIN mea­sure would report a change of zero (0.50.5), pro­vid­ing observers with bet­ter infor­ma­tion on cumu­la­tive imbal­ances or lack thereof.

Ander­sen and Bon­darenko would also do away with the switch between trad­ing time and cal­en­dar time. Start­ing with trad­ing vol­ume and stick­ing with trad­ing vol­ume — as opposed to start­ing with trad­ing vol­ume and then study­ing trades per minute — a mod­i­fied VPIN would not be arti­fi­cial­ly biased toward extreme val­ues dur­ing peri­ods of stress.

Prepar­ing for the Next Flash Crash

On Sep­tem­ber 20, 2010, the SEC and CFTC released a joint report of their inves­ti­ga­tions. They laid blame for the May 6 flash crash at the feet of a mutu­al fund, lat­er iden­ti­fied by the Wall Street Jour­nal as Wad­dell & Reed Finan­cial of Over­land Park, Kansas. At around 2:32 PM that day, a trad­er from Wad­dell & Reed had start­ed a sell pro­gram to unload 75,000 E-mini S&P 500 con­tracts worth about $4.1 bil­lion, an enor­mous sale for that instru­ment. The rate of sell­ing was pegged to 9 per­cent of trad­ing vol­ume in the pre­vi­ous minute, which meant as vol­ume ramped up the pro­gram dumped even more con­tracts onto the market.

High-fre­quen­cy trad­ing firms ini­tial­ly picked up the con­tracts with the inten­tion of quick­ly turn­ing them around. But the sub­se­quent glut of con­tracts caused the price to drop fur­ther. The high-fre­quen­cy traders’ algo­rithms pan­icked, spark­ing the 14 sec­onds of fevered trad­ing in which 27,000 trades were made but only 200 posi­tions were added — what reg­u­la­tors called a hot pota­to.” CME com­put­ers then stepped in, halt­ing trad­ing for five sec­onds, enough for every­one to catch their dig­i­tal breaths. The mar­ket start­ed to recov­er from there, though major indices still end­ed the day with sub­stan­tial losses.

It has been near­ly two years since the flash crash. The SEC has imple­ment­ed cir­cuit break­ers” that halt trad­ing when prices start flail­ing, along with lim­it-up, lim­it-down” con­trols to pre­vent indi­vid­ual stock prices from trad­ing out­side spec­i­fied bands. Still, fears linger that despite these defens­es anoth­er such flash crash could sweep the market.

Easley, O’Hara, and López de Pra­do are urg­ing reg­u­la­tors to use VPIN as an ear­ly warn­ing tool. Ander­sen, as you might imag­ine, is less san­guine about VPIN’s util­i­ty. For now, though, his hands are tied. What’s hold­ing us back is that reg­u­la­tors have col­lect­ed impor­tant data about indi­vid­ual firms’ trad­ing activ­i­ty that Amer­i­can schol­ars may access, but I can’t as I’m only a res­i­dent,” he says. But that does not mean the inquiry is over. The reg­u­la­tors have this data,” he says, adding, Peo­ple are active­ly work­ing on it.”

I think there’s hope that you could come up with some use­ful mea­sures relat­ed to VPIN. Exact­ly how suc­cess­ful they will be, I don’t know.”

Relat­ed read­ing on Kel­logg Insight

The VIX, CIV, and MFIV: Mea­sur­ing up the accu­ra­cy of option-based pre­dic­tors of volatility

Cur­ren­cy Exchange Rates: Micro effects of macro announcements

Jumps in the Mar­ket Make for Jumpy Investors: Short-lived mar­ket events can have long-term effects on the appetite for risk

Featured Faculty

Torben Andersen

Nathan S. and Mary P. Sharp Professor of Finance, Department Chair of Finance

About the Writer

Tim De Chant was science writer and editor of Kellogg Insight between 2009 and 2012.

About the Research

Andersen, Torben G. and Oleg Bondarenko. Forthcoming. VPIN and the Flash Crash. Journal of Financial Markets, October 2012 (in press).

Read the original

Suggested For You

Most Popular


How Are Black – White Bira­cial Peo­ple Per­ceived in Terms of Race?

Under­stand­ing the answer — and why black and white Amer­i­cans’ respons­es may dif­fer — is increas­ing­ly impor­tant in a mul­tira­cial society.


Don’t Let Com­pla­cen­cy Derail Your Career

How to hone your learn­ing agili­ty and take good risks.

Most Popular Podcasts


Pod­cast: Our Most Pop­u­lar Advice on Improv­ing Rela­tion­ships with Colleagues

Cowork­ers can make us crazy. Here’s how to han­dle tough situations.

Social Impact

Pod­cast: How You and Your Com­pa­ny Can Lend Exper­tise to a Non­prof­it in Need

Plus: Four ques­tions to con­sid­er before becom­ing a social-impact entrepreneur.


Pod­cast: Attract Rock­star Employ­ees — or Devel­op Your Own

Find­ing and nur­tur­ing high per­form­ers isn’t easy, but it pays off.


Pod­cast: How Music Can Change Our Mood

A Broad­way song­writer and a mar­ket­ing pro­fes­sor dis­cuss the con­nec­tion between our favorite tunes and how they make us feel.