The Gambler Who Cracked the Horse-Racing Code
Bill Benter did the impossible: He wrote an algorithm that couldn’t lose at the track. Close to a billion dollars later, he tells his story for the first time.

Horse racing is something like a religion in Hong Kong, whose citizens bet more than anyone else on Earth. Their cathedral is Happy Valley Racecourse, whose grassy oval track and floodlit stands are ringed at night by one of the sport’s grandest views: neon skyscrapers and neat stacks of high-rises, a constellation of illuminated windows, and beyond them, lush hills silhouetted in darkness.

On the evening of Nov. 6, 2001, all of Hong Kong was talking about the biggest jackpot the city had ever seen: at least HK$100 million (then about $13 million) for the winner of a single bet called the Triple Trio. The wager is a little like a trifecta of trifectas; it requires players to predict the top three horses, in any order, in three different heats. More than 10 million combinations are possible. When no one picks correctly, the prize money rolls over to the next set of races. That balmy November night, the pot had gone unclaimed six times over. About a million people placed a bet—equivalent to 1 in 7 city residents.

At Happy Valley’s ground level, young women in beer tents passed foamy pitchers to laughing expats, while the local Chinese, for whom gambling is a more serious affair, clutched racing newspapers and leaned over the handrails. At the crack of the starter’s pistol, the announcer’s voice rang out over loudspeakers: “Last leg of the Triple Trio,” he shouted in Australian-accented English, “and away they go!”

As the pack thundered around the final bend, two horses muscled ahead. “It’s Mascot Treasure a length in front, but Bobo Duck is gunning him down,” said the announcer, voice rising. “Bobo Duck in front. Mascot fighting back!” The crowd roared as the riders raced across the finish line. Bobo Duck edged Mascot Treasure, and Frat Rat came in third.

Across the road from Happy Valley, 27 floors up, two Americans sat in a plush office, ignoring a live feed of the action that played mutely on a TV screen. The only sound was the hum of a dozen computers. Bill Benter and an associate named Paul Coladonato had their eyes fixed on a bank of three monitors, which displayed a matrix of bets their algorithm had made on the race—51,381 in all.

Benter and Coladonato watched as a software script filtered out the losing bets, one at a time, until there were 36 lines left on the screens. Thirty-five of their bets had correctly called the finishers in two of the races, qualifying for a consolation prize. And one wager had correctly predicted all nine horses.

“F---,” Benter said. “We hit it.”

DR., has the model been backtested? If so, what did you find?

Yes. I looked at the results of a couple hundred races on six or seven different tracks. 32% winners with the combined estimate.

I think it's better than Prime Power for the reasons I give here: https://handicapwizard.com/hw-versus-prime-power/

Fun story, good long-read:

https://www.bloomberg.com/news/features ... acing-code




https://www.bloomberg.com/news/features ... acing-code

DR, so are you going to use the software and risk your own money? What is the play here? Hedge fund?

I risk my own money all the time using the software, but I am not a big gambler. I mostly play Friday afternoons and Saturdays. Not on Sundays.

I wouldn't encourage anyone to gamble. But if you are going to partake in the great sport that is horse racing, I would advise having a good system like this one.

Your website is an awful design btw. Nobody will take you seriously.

You are the chief jerk of the board.

Your website is an awful design btw. Nobody will take you seriously.

You need to add somebody with some user interface background, both for the Excel and the website. There are little things, like you need a more professional username for your photobucket uploads.

The model sounds interesting in that you can run it in Excel. I would have thought it would run in Python or R or something else which could crunch large scale data quickly. I've used logit and probit models in the past, our company does a lot of forecasting work for individuals, it's not the group I'm in but I learned it in graduate school.

A few questions - hint, you should start building an FAQ for your web site.
- I didn't see anything in the interface or the materials about confidence intervals. If horse A has an point estimate chance of winning of 18% and B is 14%, are you showing anything about your confidence that A actually > B?
- How much does the underlying data cost?
- How often are you planning to update the model to provide new coefficients? I can envision that while the general structure of the model is similar from track to track, the parameters will change over time, that is to say for example that post position may be a big deal at Track A, and less of a factor at Track B.
- Will it run on a tablet version of Excel? Or does it need VBA and all the bells and whistles of a full Windows Excel install?

Let’s use the software for the Arlington Outing.....

Cool. I didn't mean post position specifically, just used it as an example of what you may have statistically evaluated in the past.

$1/card is easy. When I go to Keeneland a couple of times a year I end up paying $7 for the DRF.

DRF is 19th Century technology. I don't know any serious players who use just that. The laptop ban for the Belmont is a disaster.

True, but when I go to the track just twice a year these days, i'd rather not just bet the horse with the funniest name. And carrying a laptop around the general admission part of Keeneland is rather shitty.