麻豆传媒

It鈥檚 been a long Sunday down at the bar. My favourite American football team, the New York Giants, is playing the Chicago Bears. It鈥檚 November, week 10 of 17 in the National Football League, and contending teams are starting to separate themselves from the pack, spurring on argument 鈥� and betting 鈥� on who will make the play-offs. It all peaks with the Super Bowl, the most bet-on championship game in all of sport, on 4 February.

If you want to make a fortune betting on sports, time was that your best hope would be a judicious combination of skill, intuition and luck. Not any more. These days your competitive edge could come from a more rigorous source. One man believes his computer models are now at the point where they might predict the outcomes of real matches accurately enough to tip the odds in your favour, using nothing more than statistics and a series of networked PCs.

Why is it that so few sports gamblers make money on the games? As any fan will tell you, there are too many variables. On any given Sunday, a star player can break an ankle, or a gust of wind can blow the ball off-course. So when bets are laid, bookies make money and almost everyone else loses.

Try telling that to Stephen Oh, the creator of a sports forecasting service called Accuscore. Oh鈥檚 speciality is computer simulation of American football, quantifying everything from first downs to touchdowns, fumbles to interceptions, taking full advantage of the rigid structure of the game. There have been other efforts to do this, but Oh鈥檚 approach is the most comprehensive. He has been selling his service for more than a year, and so far he鈥檚 been right. Right enough for me to stake $220 on his predictions for the week, and now I鈥檓 set to watch the real games play out on TV.

In the late 1990s, Oh was a graduate student in population genetics at the University of Michigan, Ann Arbor. He wrote computer programs to model the migrations and interrelationships of ancient peoples based on the genetic make-up of present-day populations. By quantifying things like rate of mutation, selection pressure and the advantages conferred by certain genes, he created software models that mimicked how genes migrate, mutate, and increase or decrease in frequency over many generations.

Specifically, Oh worked on the question of whether humans sprang up in Africa about 100,000 years ago and spread across the continents, or if they evolved across the world and then integrated into the familiar Homo sapiens. As he worked towards his degree, however, Oh tired of the academic debate, which he felt had little hope of being resolved.

Disheartened, he left Michigan in 1999 and worked instead as a financial consultant. But Oh鈥檚 true calling was as a sports geek, and in his spare time he analysed his favourite football team, the Baltimore Ravens. 鈥淭he coach, Brian Billick, was really conservative. All he wanted to do was run the ball,鈥� says Oh. 鈥淗e thought that if they passed a lot, they鈥檇 lose the ball more on interceptions and get blown out. I wanted to see if he was right.鈥�

So in 2002, Oh began writing code that simulated football teams instead of human genomes. Genes were replaced with players; mutation rate and selection pressure became the probability that a pass would be completed or that a player would run for a touchdown. Each generation became a discrete 鈥減lay鈥� in a game. The rules of American football, in which each play depends cleanly on the ones before it, lent themselves to modelling the game this way. And so the idea behind Accuscore was born.

From his first analysis, Oh knew he was onto something. He simulated Ravens games against the one team they couldn鈥檛 seem to beat: the New England Patriots. The code produced a slew of differing outcomes, just as you鈥檇 expect from natural variance 鈥� players won鈥檛 perform the same every game. The more simulations he ran, though, factoring in tendencies such as a quarterback鈥檚 performance near the end of games or whether a running back tends to drop the ball in the rain, the more statistical patterns began to emerge 鈥� and the more Oh says his simulations began to resemble what happened when his team played the Patriots for real.

鈥淏illick was right and wrong,鈥� he says. 鈥淚f the team passed more, it raised their probability of losing by bigger margins. But it also increased their probability of winning from 35 to 37 per cent. That may not mean anything to a scientist, but in sports it鈥檚 all about incremental edge. Even a 1 per cent edge is important.鈥�

On that point, Oh is in line with the father of quantitative sports forecasting, Bud Goode. In 1961, Goode successfully predicted the outcome of the US national championship game in college football. Over the years, he went on to predict professional games and has been hired by 21 teams to provide advice, much of it based on a statistic he invented: the yards per pass attempt differential (YPA), the number of yards a team accrues from passing the ball divided by the number of times it elects to pass, compared with what the team yields to its opponents.

The long haul

I asked Goode what he thought of Oh鈥檚 attempts to reinvent statistical football analysis. He hadn鈥檛 heard of Accuscore, but said he doubted Oh could find anything that would serve him better than YPA had served Goode. When I told him that Oh鈥檚 company markets Accuscore as a gambling tool (it already has 13,000 subscribers) he was unimpressed. 鈥淔ootball has nothing to do with the odds,鈥� he scoffed. 鈥淵ou might get lucky a few times, but you just can鈥檛 win over the long haul.鈥�

Probably good advice, but Oh has been bent on beating the odds from an early stage. In 2004, he and a small band of believers started Orion Data Analysis, the company based near Los Angeles, California, that runs Accuscore. Their aim was not just to predict the winners of games, but to beat the 鈥減oint spread鈥� that bookies set each week. This simply consists of the team the bookies expect to win plus a predicted margin of victory. To place a bet you then either back the underdog to do better than the bookies鈥� prediction (a smaller losing margin or even a win), or else back the favoured team to exceed the expected margin. Either way, the bookies give you even odds.

Accuscore simulates each game 10,000 times using a probabilistic approach called a Monte Carlo model. Each player on a side is represented by up to 70 parameters that reflect his strength, speed and skills in different situations. During a given play each key player鈥檚 possible action 鈥� whether a pass thrown and caught for a touchdown, or a fumble by a running back 鈥� is marked by a probability. Each probability depends on things like a player鈥檚 tendencies in different situations and the particular opponent鈥檚 style of play. As the play unfolds, the simulation produces a tree of these probabilities, each branch of which is a possible outcome (see Diagram).

Once the play is over, the conditions are reset with new probabilities for actions in the next play, based partly on what happened on the previous play. The result of all the simulations is a most likely final score 鈥� for instance, Bears 21, Giants 20. Then if the bookies鈥� point spread is Giants favoured by 1 point, you would bet on the Bears; if it were Bears by 3, you would bet on the Giants.

After almost two seasons of running, Oh says Accuscore is beating the point spreads offered in Las Vegas, Nevada, 56 per cent of the time. Getting it right little more than half the time may not sound like much, but betting on the spread is an even-money proposition. Even when you factor in the bookies鈥� tax, you only have to win about 52.5 per cent of the time to make a profit.

Oh admits that Accuscore has not seen enough games yet to claim that its success is statistically meaningful. 鈥淥ne season is not really enough to get over the variability,鈥� he says. But he points out that of 200 Accuscore users recently surveyed, about 70 per cent said they were up 鈥渁 little or a lot鈥� in their betting. While there鈥檚 no guarantee 鈥� this is gambling after all 鈥� Oh says, 鈥淚t鈥檚 way better than if you put your money in a hedge fund.鈥�

So far Accuscore鈥檚 subscribers, who each pay $500 a year for the service, seem to agree. What鈥檚 more, ESPN and ABC Sports, two major US media enterprises, see fit to post the Accuscore predictions alongside those of their analysts, and so far Accuscore is holding its own.

Whether Accuscore stands up over many seasons of play remains to be seen, so there鈥檚 plenty of room for the sceptics. Among them is Hal Stern, a professor of statistics at the University of California, Irvine, whose interest in the numbers behind sports led him to his present career. In the mid-1990s he and a graduate student predicted the outcomes of 110 football games, based on statistical analyses of the teams鈥� past performances.

Beating the Vegas bookies

Stern鈥檚 predictions beat the Vegas point spread 65 times, a success rate of 59 per cent. Despite his success, he admits a personal bias against the efficacy of sports predictions. 鈥淚 have a strong belief that you can鈥檛 do it, and we didn鈥檛 do well enough.鈥� He says his interest was in the intellectual challenge, and that it required an extraordinary amount of effort.

Of Oh鈥檚 approach, Stern says, 鈥淚t seems almost insane to try and simulate the whole game. Then again maybe that鈥檚 his edge.鈥� He adds, 鈥淚f you can simulate it well enough, then it鈥檚 possible that you could make money on it, but I remain largely sceptical. If you could predict games, why would you sell the advice?鈥�

Back at the bar it鈥檚 half-time, and with the Giants leading the Bears 13-10, I鈥檓 in a bit of a quandary. On the one hand, I鈥檓 a Giants fan; I want my team to win, and so do the oddsmakers, by one point. But in order to put Accuscore to the test I鈥檝e put $15 on their vision of the future 鈥� Bears by 1. What鈥檚 more, the results have already come in on the 14 other games played today, and I鈥檝e won seven and lost seven. If the Giants win tonight, I鈥檓 out a total of $30. If they lose, I almost break even.

Surrounded by a rowdy bunch of gamblers and casual fans, I find myself with a front-row seat for the TV, and watch in agony as the Giants unravel in the second half. In the end Accuscore proves right, but vastly understated: the Bears don鈥檛 win by 1, they win by 18, with the final score 38-20.

As a fan I鈥檓 dejected, sure, but overall I bet more than $200 on games I knew little about and came away losing less than $10. As a gamble it was a mild failure, but in terms of statistics, I鈥檝e landed right in the middle. A parting shot from Goode swirls in my head: 鈥淚f you bet a few games, you can win, but that鈥檚 not a meaningful win, is it?鈥�

He鈥檚 right, of course. But the way I see it, being down a little after a few games isn鈥檛 a meaningful loss either. It can be chalked up to statistical variance. True success is measured over the long term, and that鈥檚 something neither Accuscore nor I can lay claim to until we have many seasons under our belt.

This goes against Goode鈥檚 advice that gambling is the road to ruin. But me? I鈥檓 thinking about those 70 per cent of subscribers who said they were up. After watching the Bears trounce my team I鈥檝e still got most of my money, and with a little more patience, there is always hope that technology could set me on a different path: easy street.