Vol­cker, Greenspan, Ber­nanke, Yellen … and Hal?

▶ Ma­chine learn­ing may soon help cen­tral bankers de­cide pol­icy ▶ “The ca­pa­bil­ity is here. The big­gest hur­dle is … cul­tural”

Bloomberg Businessweek (North America) - - Contents - -Christo­pher Con­don

Ar­ti­fi­cial in­tel­li­gence (AI), which is al­ready steer­ing cars, could help steer the world’s big­gest economies conomies in the next half- decade. Bri­tain’s ri­tain’s cen­tral bank has been de­vel­opvelop­ing com­puter al­go­rithms for fore­cast­ing eco­nomic con­didi­tions and help­ing de­ter­minene in­ter­est rate pol­icy. Other mon­e­tary au­thor­i­ties are close be­hind. “The ca­pa­bil­ity il­ity is here,” says An­drew Lo, di­rec­tor of MIT’S Lab­o­ra­to­ry­ory for Fi­nan­cial En­gi­neer­ing. . “The big­gest hur­dle is the cul­tur­al­ral bar­rier. You’ve got a lot of f cen­tral bankers who are not as open pen to tech­nol­ogy.”

The Bank of Eng­land, un­der nder the di­rec­tion of Chief Econ­o­mist ist Andy Hal­dane, has qui­etly be­comeme a pace­set­ter in ex­plor­ing the pos­si­bil­i­ties of AI. Paul Robinson, who heads the bank’s two-year-old Ad­vanced An­a­lyt­ics unit, says the goal is to as­sist rather than re­place hu­mans. He says “many” cen­tral banks are at roughly the same stage of re­search and pre­dicts that AI will make a mean­ing­ful

con­tri­bu­tion to mon­e­tary pol­i­cy­mak­ing “cer­tainly within five years.”

Im­prove­ments would be wel­come. Econ­o­mists are, by their own ad­mis­sion, no­to­ri­ously bad at mak­ing pre­dic­tions. Con­sider the fore­casts for 2015’s U. S. gross do­mes­tic prod­uct is­sued by Fed­eral Re­serve pol­i­cy­mak­ers at the end of 2014. All 17 over­es­ti­mated the even­tual rate of growth, the clos­est by 0.2 per­cent­age point, the fur­thest by 1.3 per­cent­age points. The ac­tual num­ber was 1.9 per­cent.

Ma­chine learn­ing al­lows a com­puter to ac­quire a skill for which it hasn’t been ex­plic­itly pro­grammed. Google’s self-driv­ing car learns to drive by de­tect­ing pat­terns in vast amounts of driv­ing data. Hedge funds, such as Two Sigma In­vest­ments and Re­nais­sance Tech­nolo­gies, are al­ready us­ing AI to help make in­vest­ment choices.

At cen­tral banks the prin­ci­pal task is to set an in­ter­est rate on short-term bor­row­ing that guides the econ­omy to a sweet spot be­tween un­em­ploy­ment and in­fla­tion. Be­cause rates work with a lag, do­ing so de­pends on fore­cast­ing eco­nomic con­di­tions 6 to 12 months down the road.

One thing that makes mon­e­tary pol­icy tricky is that a rate change will al­ter the con­di­tions you’re try­ing to pre­dict. And long-es­tab­lished con­nec­tions among ec eco­nomic vari­ables can change. For ex­am­ple,ex the in­verse re­la­tion­ship be­tween un­em­ploy­ment and in­fla­tion—what econ econ­o­mists call the Phillips cur curve—seems to have disap ap­peared in re­cent years. R Robinson, of the Bank of Eng­land, con­cedes that AI works well when the struc­ture of the econ­omy i is “in­vari­ant,” or sta­ble, bu but is “less use­ful when it d does The un­der­goU.S. Fed­eral shifts.” Re­serve is mov­ing grad­u­ally on AI. It uses com­puter mod mod­els, in par­tic­u­lar one called FRB/US (pro­nounced “fer­bus”), to help with fore­casts. FRB/US is a “self­con­tained set of equa­tions, data, pro­grams, and doc­u­men­ta­tion,” ac­cord­ing to the Fed’s web­site. It’s use­ful for gen­er­at­ing an­swers to spe­cific what-if ques­tions: What will hap­pen to un­em­ploy­ment if 10-year Trea­sury yields rise by 2 per­cent­age points?

Un­like a ma­chine-learn­ing sys­tem, FRB/US doesn’t learn on its own. “For the fore­see­able fu­ture, the best ap­proach will in­volve a com­bi­na­tion of em­pir­i­cal rigor cap­tured in mod­els, to­gether with hu­man judg­ment,” says David Wil­cox, di­rec­tor of the Fed’s di­vi­sion of re­search and sta­tis­tics.

Some AI cham­pi­ons, such as Google Chief Econ­o­mist Hal Var­ian, are also skep­ti­cal about AI’S abil­ity to make eco­nomic fore­casts, but for dif­fer­ent rea­sons. As he sees it, the tech­nol­ogy is ready, but the data—the co­pi­ous sup­ply of raw num­bers that AI pro­grams sift through to reach con­clu­sions—are want­ing. “The data sets are so small. GDP is re­leased quar­terly, so 50 years of data is only 200 ob­ser­va­tions and only seven re­ces­sions,” he wrote in an e-mail.

AI will soon have a lot more data to chew on. Web scrap­ers such as MIT’S Bil­lion Prices Project are al­ready comb­ing the in­ter­net for real-time price points rel­e­vant to in­fla­tion.

Econ­o­mists and com­puter sci­en­tists agree there will al­ways be a role for hu­man be­ings in cen­tral bank­ing. As Michael Feroli, chief U. S. econ­o­mist at Jpmor­gan Chase, puts it: “I don’t see why, in prin­ci­ple, you couldn’t have a com­puter set mon­e­tary pol­icy. Hav­ing it tes­tify be­fore Con­gress is an­other mat­ter.”

The bot­tom line The ma­chine-learn­ing branch of ar­ti­fi­cial in­tel­li­gence could help cen­tral bankers set in­ter­est rates within five years.

Newspapers in English

Newspapers from Canada

© PressReader. All rights reserved.