National Post (Latest Edition) - - SPORTS - ED­DIE PE LLS

Mes­sage to hoops fans: This story could make you look bril­liant. A com­puter science pro­fes­sor at Univer­sity of Illi­nois has cre­ated a for­mula that pre­dicts NCAA tour­na­ment up­sets at dou­ble the suc­cess rate of some­one pick­ing at ran­dom — in­clud­ing, but not lim­ited to, those who throw darts at the bracket, or pick based on their favourite colour, the most fe­ro­cious mas­cot or the num­ber of vow­els in the coach’s last name.

This year’s up­set picks both come out of the South re­gion. They are No. 13 Buf­falo over No. 4 Ari­zona and No. 14 Wright State over No. 3 Ten­nessee.


The com­puter sci­en­tist who spear­heads this project, Shel­don Jacobson, says the com­puter mod­els only an­a­lyze po­ten­tial up­sets by 13, 14 and 15 seeds. A 16 seed has never beaten a 1; any­thing in­volv­ing 11s or 12s pro­duce “too much noise,” Jacobson says, mean­ing the rel­a­tive close­ness of the un­der­dogs with their fifth- and sixth­seeded op­po­nents in­ter­feres with the sta­tis­ti­cal model he uses to pre­dict the up­sets.

Jacobson and fel­low sci­en­tists pared down 115 pub­licly avail­able met­rics for ev­ery team in col­lege bas­ket­ball to 15 that have served as the best pre­dic­tors of up­sets in years past.

Some ex­am­ples in­clude ef­fec­tive pos­ses­sion ra­tio — es­sen­tially the num­ber of points a team scores per pos­ses­sion — along with av­er­age scor­ing mar­gin and op­po­nent’s 3- point shoot­ing per­cent­age.

Now for the science: The frame­work of these for­mu­las is called “bal­ance op­ti­miza­tion sub­set se­lec­tion” ( BOSS), which is an ar­ti­fi­cial- in­tel­li­gence al­go­rithm ( Google that if so in­clined). The Na­tional Science Foun­da­tion ini­tially funded Jacobson for a project that used ar­ti­fi­cial in­tel­li­gence to ex­plore so­ci­etal is­sues, such as whether gov­ern­mentspon­sored pro­grams to en­hance job skills ul­ti­mately lead to higher in­comes for work­ers.

Af­ter the fund­ing ran out, Jacobson sought uses for his cre­ation that could re­ally help peo­ple.

March Mad­ness gen­er­ates more than $ 10 bil­lion US a year in wa­ger­ing, much of which comes when play­ers chip in $ 10 or $ 20 and fill out brack­ets for their of­fice pools and col­lect points based on the num­ber of cor­rect picks.

Pick­ing the even­tual cham­pion — No. 1 seeds Vil­lanova and Vir­ginia started at 5-1 odds to win it all, with No. 2 Duke at 6-1 — al­ways helps. But some­times the real dif­fer­ence mak­ers are the cor­rect up­set picks in the early rounds. That’s when the Buf­fa­los and Wright States of the world beat Go­liath and briefly re­store faith in the gum­drops- and- lol­lipops no­tion that any­thing re­ally is pos­si­ble.

It’s not, Jacobson as­sures us.

Still, his web­site, brack­e­todds.cs.illi­, gets thou­sands of hits a day this time of year. Among his other bas­ket­ball-re­lated projects are pre­dict­ing which teams will make the tour­na­ment and where they’ ll be seeded. The com­puter didn’t do so well this year — it had Louisville and USC com­fort­ably in — in large part be­cause, as Jacobson says, the se­lec­tion com­mit­tee “keeps chang­ing the rules.”

“You had a team like Ari­zona State that got in de­spite some hor­ri­ble home losses to me­diocre Pac-12 teams, and then you have Louisville, which is the kind of team that typ­i­cally gets in but didn’t,” Jacobson said. But that was then. Once the brack­ets were re­vealed, Jacobson set the com­puter’s fo­cus to­ward pick­ing these up­sets. Its track record since 2003 is hardly per­fect, but still prob­a­bly bet­ter than yours.

Us­ing BOSS, the com­puter picks the two most likely up­sets each year. Last sea­son, not a sin­gle 13, 14 or 15 ad­vanced, so it got 0 per cent. Two sea­sons ago, there were three such up­sets — Iona and Buf­falo — but the com­puter didn’t pick ei­ther of them.

But in 2015, BOSS picked Ge­or­gia State and UAB and went 2 for 2.

And since 2003, 10 of its 26 se­lected games have re­sulted in up­sets. That’s 38.4 per cent, or dou­ble the ex­pected num­ber of cor­rect se­lec­tions a per­son would get by us­ing a “weighted ran­dom se­lec­tion method.” In other words, dou­ble what you’d get by pick­ing slips out of a hat, or choos­ing a team be­cause you like the fight song.

For those plac­ing faith in his science, Jacobson warns of the large gulf be­tween pre­dict­ing the fu­ture and fore­cast­ing what could hap­pen.

“No­body pre­dicts the weather,” he says. “They fore­cast it us­ing chances and odds.”

Sim­i­larly, he says, “ar­ti­fi­cial in­tel­li­gence looks at some out­comes that the hu­man eye can’t catch. The mod­els we use give some in­di­ca­tion of what the fu­ture may look like.”


Jacobson freely ad­mits he does not gam­ble on bas­ket­ball or any­thing else.

Nor does he fill out a bracket.


No. 13 Buf­falo is one of those teams to watch to pull off an up­set in the first round, pre­dicts a com­puter model.

Newspapers in English

Newspapers from Canada

© PressReader. All rights reserved.