Business: Rise of the bet bots
There’s an arms race brewing in the sportsbook as bettors and bookmakers adopt AI to give them an edge
The machines are coming.
That sentence used to sound like some bit of techno-paranoia. Now it’s just a statement of fact. Artificial intelligence is being heavily researched and developed by the leading Silicon Valley brains of our day.
“It’s inevitable,” Louis Rosenberg says. “Nature proves it’s possible. If you connect networks of neurons together, you can build an intelligence. It’s inevitable, and I think it’s in the 25- to 75-year time range. It’s not a thousand years out in the future.” But before Skynet sends a bunch of Terminators to Earth, our new robot overlords might be more interested in your wallet.
Rosenberg is the CEO and chief scientist of Unanimous AI, a 4-year-old San Francisco company that champions “Swarm AI.” Unanimous aims to replicate the way flocks of birds, schools of fish, and swarms of bees amplify their intelligence through participating in feedback loops, whether they’re avoiding predators or choosing a new home. But in this case, the swarm is emergent software intelligence, and the goal is to help humans make better decisions. Swarm AI works by soliciting individual input, weighing each person’s confidence in his decision, and running it through a neural network that processes those individual inputs into new insights that boost the intelligence of the group as a whole. The ability to adjust for feedback from participants in real time sets Swarm AI apart from mere crowdsourcing.
Unanimous garnered national attention in 2016, when the firm was challenged to pick the Kentucky Derby superfecta, or the first four finishers in a horse race, in order. Drawing on 20 experts for raw input, Unanimous predicted Nyquist, Exaggerator, Gun Runner, and Mohaymen. Those four horses hit the wire, in that order. It paid $542.10 on a $1 bet. Rosenberg made $11,000 on his $20 ticket.
Rosenberg uses sports betting as something of a test case for his technology, releasing picks regularly on the company’s blog and publishing studies in full-season predictions. Last year, over the course of 200 NHL games, the swarm was able to pick winners at a 61 percent clip. Narrowing the field to the 20 games that represented the swarm’s best of the week, it connected at 85 percent. Betting 15 percent of an initial $100 bankroll on the pick of the week yielded a $270 profit. Simply choosing sportsbook moneyline favorites in those games would have ended in a $28 loss.
Unanimous isn’t the only AI company experimenting with sports betting, either. There are multiple European companies that are going full-steam ahead with offering AI-assisted picks to the public. London-based Stratapro offers AI-driven soccer picks for a monthly subscription fee. Poland’s AIbet covers everything from Japanese baseball to European handball. Based in Wellington, Washington, Sportsflare purports to be “a system to convert sports betting into a viable investment vehicle.” (Which guys have been saying to their wives about their weekly trips to the bookie since about 1921.)
But this hardly means sportsbooks are going to stand around and be outsmarted by would-be cyborg bettors using smart software to get an edge. On the other side of the counter, casinos are gearing up to employ artificial intelligence as well, to protect and even increase their sportsbook profits. For casinos, the next AI-powered sports-betting trend is the advent of “in-play” wagers — that is, a constant cycle of new bets and odds generated in real time as the live game happens.
“Not the next football season, but the following football season, in-play (wagers) will be widely recognized,” says John English, managing director of sports betting and technology at Global Market Advisors, a gaming consulting firm. He was previously with American Wagering Inc., the first company to develop a mobile sports betting app that received regulatory approval from Gaming Control.
“When we sold (American Wagering) to (United Kingdom-based) William Hill, I wanted them to see the sportsbook in full motion on a Sunday afternoon. They’re looking at the boards, going, ‘What the hell are you guys betting on? There’s nothing to bet on. This is all you got?’”
That’s because the Brits from William Hill were used to more sophisticated betting systems. Similar to what bettors can currently find on apps like those offered by Station Casinos, the Superbook and others, European bettors are accustomed to a dazzling array of wagers — all constantly updated by AI-backed algorithms — on live, in-play action during the course of a game. Operators trust AI to set all the lines, and adjust them on the fly.
With in-play betting, the menu of available bets can run a dozen or more, constantly updated as the clock ticks and plays develop. For example, in the third quarter of a college football game with the score at 24-16, you might get options on a point spread of 7.5, a point total of 61.5, game moneyline bets, the margin of victory, the first team to score 30 points, what type of play the next score will come on, and so on. As the seconds slip by and a drive for the team that’s trailing stalls, that total slides down to 58.5. The spread jumps to 9.5. Bets are frozen as plays develop and odds shift. If regular betting is knocking in 15-foot putts, in-play wagering is hitting your driver from a surfboard during storm season. English estimates that the advent of AI-driven in-play wagering has expanded the U.K. market by 70 percent.
New bot with an old trick
Artificial intelligence isn’t the first technological breakthrough to alter the sports-betting landscape. In some ways, AI is a new dog with an old trick. Which is why some veteran bookmakers aren’t fazed if squares or sharps have a new tool at their disposal.
Jay Kornegay, vice president of race and sportsbook operations at Westgate Resorts, was working at the Imperial Palace when the infamous Computer Group (famous sports bettor Billy Walters, now in prison for insider trading, was a partner) was revolutionizing betting by harnessing the power of algorithms in the ’80s and ’90s, long before the public caught on to the advantages of using burgeoning processing power to beat the books.
“These programs, what I found out over the years, is sometimes they go on these streaks,” Kornegay says. “They don’t seem to last. The game catches up to them, the lines adjust to certain trends or information. When we look at sharp play, they’re always going to get respect from the bookmakers, meaning when they do play something, that line will be aggressively moved. Once that line goes from 4 to 5, that perceived value that’s there may no longer exist after that first play.”
In other words, if too much money in a betting pool is on one side or another, sportsbooks can adjust the point spread, narrowing it to encourage more action to flow to the favorite, or widening it to nudge bettors toward the underdog. Sophisticated bettors, or sharps, tend to look for “soft” lines where they think the books are vulnerable. If the Steelers are favored by three, and a sharp has calculated they’ll win by seven, the money goes to Pittsburgh. If the sportsbooks adjust the line to eight, sharps will change their bet or stay away altogether.
There are two main weapons the books have to mitigate their exposure: adjusting moneylines and point spreads (to attract bets to both sides, thereby limiting their exposure to any one outcome), and limiting the action they’ll take on certain wagers or from certain customers. Paradoxically, as the sophistication of bettors increases through AI-generated information or other means, the more the books will have to rely on savvy, educated humans to plug leaks.
It will require operators to have humans watching the machines to see where bets are coming from. Sharp play can be throttled down through limiting so that sportsbooks aren’t vulnerable to bad lines. But it takes a wise hand at the wheel to also know the difference between when the money is coming in from someone truly sharp, or is just coming from someone on a hot streak. (Ask anyone who’s lost their office fantasy football league to the 22-year-old receptionist who’s never watched a single down in her life.)
Bad lines on in-play bets do leave operators vulnerable to the truly sharp, but it’s a rich revenue stream by offering dozens of more options per game — and therefore dozens of more chances for the betting public to make mistakes.
“In in-play wagering, it’s happening so fast, people don’t have the time to handicap it or think it through,” English says. “It’s more of an impulse style of betting. It typically has a much higher hold.”
The future gets weird
American sportsbooks are only a couple of years away from being run like their AI-enhanced European counterparts, but it also won’t be long before AI becomes available on a mass consumer scale to sports bettors. That’s when things could get weird.
“As both sides have good information, it’s like you and I without information,” English says. “As much as their information increases, we need to make sure our technology from the bookmakers’ side of the equation keeps up with the technology. Artificial intelligence is constantly updating its systems to be the best it can be. As long as you’re doing that, for the most part it’s kind of an equalizer. It goes back to when nobody had any information. You just had a hunch. If you have great information and I have great information, let’s go at it and see whose is better.”
The magic number is 52.38 percent. That’s the hit rate a bettor must make in point-spread wagers like in football and basketball to break even against the built-in house edge that requires a $110 bet to win $100.
For Rosenberg’s version of artificial intelligence, a 52.38 percent accuracy rate is small potatoes. “There is a maximum limit,” he said. “I think it’s different in every sport. It really just depends on what kind of random events can happen. But in terms of accuracy, it’s probably somewhere north of 85 percent.”
It’s not a wild stretch to envision a day when bettors turn their AI systems loose to do battle with AI-enabled betting platforms, as alerts clue in operators that something is happening out there, and human adjustments might need to be made. If that kind of equilibrium is reached, English expects to see lines that stay more or less static over the course of a day or week before the game.
But those heights are still a way off. For a span from weeks nine to 12 in the NFL, Unanimous’ publicly listed predictions only identified 28 of 51 straight-up winners of the games (a 55 percent accuracy rate) and went just 16-26 against the point spread (38 percent accuracy).
“It’s hard to put your head around this, that something could hit 85 percent,” Kornegay says. “We would have to have a large sample of success for us to move (a spread) more than a point or two.”
If betting matured with the advent of the Computer Group into the cat-and-mouse between bookmakers and bettors, artificial intelligence may accelerate that process. But whether it can fundamentally change it is another story.
Dick Carson is cut from the cloth of the great old-time road gamblers. He was a pool hustler, a freewheeling bookie who got shaken down by Anthony Spilotro in the late ’70s, and a champion lowball poker player in the ’80s. He used to set his lines by pencil and paper. He had to learn to adjust through the years as bettors became more educated with the advent of more and better public information, including information based on computer-driven predictive models. That initial awakening of the betting public in the ’80s and ’90s may be the closest example of a fundamental shift in how bettors wager to help frame the coming age of AI.
Carson told the Thinking Poker podcast, “You get the price right, you put the line up and you make them lay 11-to-10. What’s the only thing that hasn’t gone up in 119 years? Everything’s gone up. Not 11-to-10 though. 11-to-10 must be okay. I don’t know what Abraham Lincoln was booking, but I think 11-to-10. If you have the right number, 11-to-10 will overcome everything.” In other words, the house will always get its cut.
English doesn’t think that laying 11-to-10 will ever change. Books trying to force bettors to lay 12-to-10 (that is, betting $120 to win $100) will force them to turn to illegal, offshore competition. The margins for the house in sports betting are small relative to the pits and the slots. A generation ago, though, 3-to-2 for a blackjack seemed as eternal as the pyramids. In a world of 6-to-5, we’re seeing that fundamental gambling maxims aren’t as fundamental as we thought.
The days of savvy handicappers matching wits directly against seen-it-all bookmakers armed with nothing but box scores and a feel for the game are long behind us. The era of traditional bets on a total or a side as the main wagers of choice in Las Vegas and other American markets may be drawing to a close. What won’t change — what keeps people coming back to sports betting, even in a theoretical future where the computers precisely predict which team will win every matchup — is exactly what will keep the house in business. The instruments might change, but there are only so many notes on the scale.
“People want to win. They want to get down,” English says. “There’s always been a guy that’s got a system who feels he’s going to win more than the other guy. In the end, as long as the house can keep that edge, you could have all the best information in the world, and you still can’t always 100 percent predict the outcome of a sporting event. Nothing’s going to stop (another player) from taking out Tom Brady’s leg.”
The machines are coming, but we random, capricious, unpredictable humans aren’t going anywhere anytime soon.