Bloomberg Businessweek - USA (2019-05-27)

W

e’vecome toexpect a lotfrom intelligent

machines:Spotifypickssongsforyourplaylist.

Yourphonehaslearnedtorecognizeyourface,

evenona badhairday.Andself-drivingcars,

we’repromised,arejustaroundthecorner.Asthetechnol-

ogygetsbetter,it usuallybecomesmorecommonplace.But

here’ssomethingyou’lllikelyneverhaveaccessto:a comput-

erizedstockpickerthatoutsmartsthemarket.

It’snotforlackofinterestonWallStreet’spart.Theeffort

toscientificallymodelmarkets,whichbeganinthemid-1980s,

hasabsorbedthetalentsofsomeofthebrightestgraduatesof

mathandcomputerscienceprograms.A handfulofsecretive

hedgefundmanagers—includingRenaissanceTechnologies,

PDTPartners,andD.E.Shaw—havecarvedoutextraordinary

returns.Buttellingly,manyoftheleadingoperationstoday

arethesameonesthatdominatedscientificmodelingdecades

ago.Andyouprobablyaren’trichenoughorconnected

enoughtoinvestwiththem.

Onereasonmachineinvestingremainsanelitedomainis

obvious.Bydefinition,mostinvestorscan’tbeattheaverage,

andeverycomputerthatmomentarilyfindsa winningfor-

mulawillsoonfaceotherstryingtooutwitit.Butit turnsout

thatinvestingis alsosimplyharderthan,say,predictingyour

nextAmazonpurchase.“It’soneofthemostdifficultprob-

lemsinappliedmachinelearning,”saysCiamacMoallemi,

a professoratColumbiaBusinessSchoolanda principalat

BourbakiLLC.Herearejustsomeofthedevilishproblems

financialengineersaretryingtocrack:

➡Or,in quantspeak,it’snonstationary.

Anexampleofstationarydatamight

bethedistancebetweenyourleft

eyeandyournose.Unlessyouhave

plasticsurgery,it’sa constant.If a

machineis fedhundredsofpicturesof

you,it willbeabletoidentifyyouwith

highprobability.

Infinancialmarkets,data

canchangedramaticallyandin

unprecedentedways—forexample,

wheninterestratesturnednegative

acrossmuchofEuropeandJapan

in 2013.Othershiftscanbemore

mundane.In 1998 pricingofU.S.

stockswenttodecimalsfrom

fractions.Thatwasn’thardfor

computerstoadjustto,butit might

haveflusteredsomeofthehuman

traders.“Itchangedsomestructure

in themarketandprobablysome

behavior,too,”saysGlenWhitney,a

formerresearcheratRenaissance.

➡Stocksmoveallthetime,andnot

alwaysforanydiscerniblereason.

Mostmarketmovesarewhat

economistscallnoisetrading.①Togo

backtotheimage-recognitionanalogy,

imaginea computertryingtoidentify

peoplein photosthatweretakenin

thedark.Mostofthedatain those

picturesis noise—uselessblackpixels.

What’smore,asdatasetsgo,the

historyofstockpricesis relatively

thin.Sayyou’retryingtopredicthow

stockswillperformovera one-year

horizon.Becauseweonlyhave

decentrecordsbackto1900,there

areonly 118 nonoverlappingone-year

periodstolookatin theU.S.Compare

thiswithFacebookInc.,whichhas

anendlesstroveofstufftocomb

through—itprocesses 350 million

picturesa day.Andin image

recognition,simpletrickssuchas

rotatingthephotooralteringcolors

canincreasetheamountofdata;it’s

difficulttoartificiallyincreasethesize

ofa financialdataset.

➡Anobvioussignal—forexample,to

buystocksonthefirstdayofevery

month—isnotofmuchuse.If that

workedin thepast,it wasprobably

justa fluke,andevenif it isn’t,it’sgoing

tobequicklydiscoveredandtraded

awaybyothers.Soresearchershave

focusedonveryfaintsignals,ones

thatmightpredictthefuturepricewith

only51%certainty.“Wewerelooking

forpatternsthatarejustontheedge

ofdetection,”Whitneysays.Most

investorscan’ttakeadvantageof

suchpatterns.Tomakethemwork,

moneymanagershavetocombine

thousandsofbets②andmagnify

themwithleverage—investingwith

borrowedmoney.

THEDATAKEEPSCHANGING

THERE’SMORENOISETHANSIGNAL

THEEDGEYOU’RELOOKINGFORISREALLYSMALL

Prediction canbeimprovedonlysomuch,forcingelite

quantitative managers to look for other advantages. In invest-

ing, one profitable problem to solve is transaction costs.

The obvious transaction cost is the fee the broker charges.

But there’s also something called slippage, which accounts for

the quoted price—$135 for a share of IBM Corp., for example—

being relative to the number of shares you want to buy. You

might be able to buy only 100 shares at $135; to buy 1,000 shares

would require bidding a higher price to attract new sellers. The

average cost might then be $136. The only way to know the true

price, with slippage, is to transact in the market.

Teaching a machine to anticipate transaction costs helps

in two ways. First, the edge required for a trading signal to be

profitable might go from 51% to 50.5%. The second advantage

①
Economists have shown that stock prices fluctuate much more than news or changes in fundamentals would explain

② A 51% edge, applied over 10,000 trades, makes a profitable week all but certain. Assuming the edge is really there

Bloomberg Businessweek

WHERE THE MONEY IS

May 27, 2019