Las Vegas oddsmakers have made Michigan State a 2½-point favorite over Purdue in this afternoon’s basketball game. How did they come up with that spread? Mostly it’s an algorithmic simulation, a multivariate statistical equation built up from analyzing massive amounts of data collected about all the teams’ past performances in all of their games. But the bookmakers make their money not by betting on the games but by collecting the vig — the house’s cut on each bet placed, typically 10% of the amount bet.
Suppose Purdue, in a mild upset, wins the game. Vegas will pay off on bets that went under the predicted 2½-point MSU margin of victory. But they’ll also collect from bets that went over the spread. As long as the total amount of money bet on the “under” doesn’t greatly exceed the “over,” Vegas breaks even on the betting money while still making a 10 percent profit on the vig.
The point-spread algorithm isn’t deterministic, cranking out a single prediction of which team will win and by how much. It’s probabilistic, generating a Bell curve’s worth of possible outcomes of varying degrees of likelihood. Each of the factors going into the oddsmaking algorithm has a plus-or-minus confidence interval associated with it.
E.g., in the 18 games it’s played so far this year MSU has scored an average of 83.8 points, while Purdue has yielded to its opponents an average of 68.2 points. So you’re first best guess might be that MSU will score halfway between its average score and Purdue’s average points allowed: (83.8 + 68.2)/2 = 76 points. But there’s also the variability to take into consideration: MSU has scored between 106 and 63 points, while Purdue has allowed between 89 and 46 points.
The same idea holds for pace of play, rebounding margin, shooting percentage, and all of the other relevant variables: an average score is surrounded by a plus-or-minus halo of variation. These probabilistic variables are combined and differentially weighted in an equation that, like its components, generates a probabilistic average and a range of variation.
Only one actual game between MSU and Purdue will be played this afternoon; when the buzzer sounds only one actual final score will be posted on the scoreboard; some betters will have won while others have lost. But from the algorithm’s standpoint the game that’s played this afternoon is only one of an infinite number of MSU-Purdue games that could be played — a practical application of multiverse theory. The algorithm runs a series of simulations on maybe a thousand hypothetical games, systematically tweaking the plus-or-minus variabilities in the model to generate what-if scenarios, then cranking out a final score for each simulated game. In some simulations MSU wins by a dozen; in others MSU loses by a dozen. What Vegas wants to know is the midpoint, where half of the simulated game results fall on left tail of the distribution, the other half on the right tail. That’s where Vegas wants to set the point spread.
But not so fast. Just because the algo cranked out the average point spread for its simulated games, that doesn’t mean that the betters are going to fall 50-50 on either side of that spread. People who bet on games might have a system and do analyses; they might even have an algorithm of their own that they’ve built or a service they’ve bought into to help them beat Vegas at its own game. But betters also play hunches, follow instincts, play favorites.
MSU has won 21 straight Big Ten games: aren’t they about due for a loss? And MSU has beat the Vegas point spread in 8 straight times: surely things are due to even out. That sort of thinking is called the gambler’s fallacy — that the longer somebody has been on a lucky streak, the greater the likelihood that the streak will come to an end. Still, just because it’s a fallacy from an empirical standpoint doesn’t mean that bettors stop believing it.
So there’s still a seat in the back room for the guy with the green eyeshade smoking a cigarette. Let’s say that the algo runs a thousand simulations of the game and the average prediction is MSU by 5 points. The savvy bookmaker might have reason to expect that gambler’s fallacy will play a role in the betting and consequently nudge the spread down a little bit. But the bookmaker needn’t rely solely on intuition and experience. There’s plenty of historic betting data to be mined; simulated bets can be placed online or in focus groups days before the actual odds are to be posted. Almost surely Vegas relies now on a self-learning AI to predict human betting patterns for each game, using its findings in tandem with those of the point-spread prediction algorithm.
The open-access college basketball algorithm website I pay attention to is Kenpom. According to that algo, MSU is about 9 points better than Purdue. However, the game is being played on Purdue’s home court, and according to Kenpom’s analysis the home court advantage is worth on average about 3½ points. So MSU should be expected to win by 9 – 3½ = 5½, covering the Las Vegas spread of 2½.
I have rooting interests in MSU, having graduated from there. On the other hand, I do tend to think that MSU has been playing over their heads lately and are due for a bit of a comeuppance. However, I acknowledge that there’s a certain amount of gambler’s fallacy bias creeping in there. So I’ll split the difference between Las Vegas and Kenpom.
My prediction: Michigan State by 4 — I’m taking the over.
UPDATE… Final score: Purdue 73, Michigan State 63. One of those wrong futures showed up in the present. Good thing I didn’t put my money where my mouth was.