High-frequency trading : what are the risks and the returns for financial markets ?

Hi ! PARIS is the new Cen­ter on Data Ana­ly­tics and Arti­fi­cial Intel­li­gence for Science, Busi­ness and Socie­ty crea­ted by Ins­ti­tut Poly­tech­nique de Paris (IP Paris) and HEC Paris and recent­ly joi­ned by Inria (Centre Inria de Saclay).

The term refers to the pla­cing of buy and sell orders on the finan­cial mar­kets over an extre­me­ly short time frame, on the order of nano­se­conds, gene­ral­ly in reac­tion to events. For example, when the Fed decides to cut inter­est rates, high-fre­quen­cy tra­ders will place orders a nano­se­cond after the macroe­co­no­mic announ­ce­ment is made. 

Obvious­ly, humans can’t react on this time scale. High-fre­quen­cy tra­ding is cha­rac­te­ri­sed by two phe­no­me­na. First­ly, it is auto­ma­ted tra­ding : it requires the use of algo­rithms and com­pu­ters to place orders. Second­ly, it must react extre­me­ly qui­ck­ly to mar­ket events, whe­ther macroe­co­no­mic announ­ce­ments or changes within the mar­ket such as price movements.

If we look at the Paris Stock Exchange, for many years we had this image of tra­ders shou­ting their orders around the tra­ding floor. This has not been the case since 1986, when the tra­ding floor was repla­ced by a com­pu­ter RAM. Since then, tra­ders have been car­rying out their tran­sac­tions all over the world via online ter­mi­nals. This move towards auto­ma­tion began in the late 1960s and ended for equi­ty mar­kets in the ear­ly 1990s. This auto­ma­tion led to so-cal­led algo­rith­mic tra­ding and, over the last twen­ty years, to high-fre­quen­cy trading.

On the equi­ty mar­kets, it’s esti­ma­ted that around two-thirds of tran­sac­tions are car­ried out by high-fre­quen­cy tra­ders. This may seem high, but it’s not all that sur­pri­sing. It’s a bit like super­mar­kets : while they account for a large pro­por­tion of consu­mer tran­sac­tions in France, they are mere­ly inter­me­dia­ries. High-fre­quen­cy tra­ders are not the final hol­ders of secu­ri­ties, but the prin­ci­pals. They gene­ral­ly only hold the secu­ri­ties for a very short time.

It’s true that it has evol­ved consi­de­ra­bly. The tra­ders we often see in the movies, behind their com­pu­ters or on the phone giving their orders, have gene­ral­ly been repla­ced by com­pu­ters and algo­rithms. Today, most tra­ding algo­rithms are deve­lo­ped by com­pu­ter scien­tists. They are recrui­ted by spe­cia­li­sed com­pa­nies, par­ti­cu­lar­ly in high-fre­quen­cy tra­ding. These pro­files can pro­vide a defi­nite com­pe­ti­tive advan­tage, as they have exper­tise in wri­ting tra­ding-spe­ci­fic codes and have deve­lo­ped exper­tise in opti­mi­sing mar­ket access speed.

While high-fre­quen­cy tra­ding requires an algo­rithm with the right stra­te­gies, it is true that the mar­kets are well aware of them. It’s just that the time scale for put­ting them into prac­tice is much shor­ter. Opti­mi­sing this speed of access to the mar­ket is the­re­fore cru­cial, because reac­tion time is essen­tial : the ope­ra­tor with the shor­test reac­tion time has an unde­niable com­pe­ti­tive advan­tage. These tech­ni­cal skills are very spe­ci­fic and gene­ral­ly concern people who have trai­ned in com­pu­ter science, for example. If there is one area of the eco­no­my where we can talk about Big Data, it is the finan­cial mar­kets. Eve­ry day, they gene­rate an enor­mous amount of data in digi­tal form. Eve­ry tran­sac­tion, eve­ry order sub­mis­sion is a new piece of data. For high-fre­quen­cy tra­ding, it is the­re­fore neces­sa­ry to sub­scribe to data feeds that are sold by the exchanges. This requires high­ly advan­ced tech­ni­cal exper­tise and tech­no­lo­gi­cal invest­ments that are cost­ly both in terms of human capi­tal and computers.

Inves­tors like you and me don’t have enough capi­tal to access the oppor­tu­ni­ties these inves­tors do when making these invest­ments. The prin­ciple of high-fre­quen­cy inves­ting is to obtain modest pro­fits on each tran­sac­tion. These are very small pro­fits, but because the tran­sac­tions are repea­ted so many times, the pro­fits are very large very qui­ck­ly. If the invest­ment is large, risk-taking is mini­mal, because the stra­te­gies are well known and reac­tion time is opti­mi­sed to make the most of oppor­tu­ni­ties. The unpre­dic­table part is rela­ti­ve­ly small. As a result, high-fre­quen­cy tra­ding firms have expan­ded rapid­ly over the last ten years. The pro­fits that can be made are now much lower.

There are seve­ral famous cases of insta­bi­li­ty cau­sed both by algo­rithms and to the detriment of algo­rithms, inclu­ding the 2010 flash crash in the Uni­ted States. The stock mar­ket crash las­ted just over 30 minutes and was not direct­ly due to high-fre­quen­cy tra­ding, but it did cause these ope­ra­tors to stop their algo­rithms, which pro­ba­bly ampli­fied the crash. And then, as always with algo­rithms, there can be desi­gn flaws, bugs that were not anti­ci­pa­ted. This was the case in 2012 for Knight Capi­tal, a major US tra­ding firm and one of the four major high-fre­quen­cy tra­ding ope­ra­tors. One day, one of their algo­rithms deve­lo­ped a bug and Knight Capi­tal qui­ck­ly recor­ded a loss of 440 mil­lion dol­lars and almost went bankrupt.

One area we still don’t talk about very much is the risk of intru­sion by a plat­form or algo­rith­mic ope­ra­tor. Ill-inten­tio­ned hacking, for example by ter­ro­rists, could serious­ly des­ta­bi­lise the market.

We can expect the tra­ding indus­try to use these tools one day. Some ini­tial thought has alrea­dy been given to this issue in a field other than the finan­cial indus­try, name­ly online sales. Regu­la­tors, nota­bly in the UK and the US, have shown that there is col­lu­sion bet­ween algo­rithms. Take Ama­zon’s pri­cing algo­rithms, for example : they are foun­ded on arti­fi­cial intel­li­gence based on lear­ning first, a self-lear­ning pro­cess for repri­cing pro­ducts. But they don’t just raise or lower prices depen­ding on the context, they find ways to charge non-com­pe­ti­tive prices. If they are not coded for this, they dis­co­ver ways of making pro­fits by crea­ting impli­cit collusion. 

Opti­mi­sing the speed of access to the mar­ket is cru­cial, because reac­tion time is everything.

This has not yet become an issue for the finan­cial mar­kets, but it soon will, because it is part of the same ques­tion. In 2017, Bla­ckRock, one of the lar­gest Ame­ri­can asset mana­gers, began using arti­fi­cial intel­li­gence tech­niques to fore­cast returns and make port­fo­lio allo­ca­tions. The firm also relies on big data as part of its quan­ti­ta­tive and tra­di­tio­nal invest­ment stra­te­gies. As for the invest­ment bank JP Mor­gan, it has alrea­dy announ­ced the deve­lop­ment of a fore­cas­ting tool based on arti­fi­cial intel­li­gence to anti­ci­pate the deci­sions of cen­tral banks.

Interview by Jean Zeid