Algorithms that Won $1bn in Horserace Betting

Applications in Data Science, Training Models and Investing

If you know me, you know how much I love Ed Thorp, probability theory, decision trees, and poker. Bill Benter wrote an algorithm to win over a billion dollars at the Hong Kong horse-racing tracks. A few takeaways on how he did it:

  1. Go directly to the academic papers. Benter first bought every book on horse-betting, but wanted something more rigorous. He went to the university library, and found an academic paper: “Searching for Positive Returns at the Track: A Multinomial Logit Model for Handicapping Horse Races.”
  2. Teach yourself the tech tools. Benter taught himself advanced statistics and learned to write software on an early PC with a green-and-black screen.
  3. Backtest. Benter hired typists to key in stack of yearbooks containing the results of thousands of races.
  4. Start with small bets to experiment and correct the model.
  5. Understand the odds, the other players, and the scenarios. One such scenario is Gambler’s Ruin: if a player with limited funds keeps betting against an opponent with unlimited funds (that is, a casino, or the betting population of Hong Kong), he will eventually go broke, even if the game is fair. All lucky streaks come to an end, and losing runs are fatal. The same is true in the stock market: the market can stay wrong longer than you can stay solvent.
  6. Find solutions that have worked for similar problems. Texas physicist John Kelly Jr posited a scenario in which a horse-race better has an edge: a “private wire” of somewhat reliable, but not perfect tips. How should he bet? Wager too little, and the advantage is squandered. Too much, and ruin beckons. Kelly’s solution was to wager an amount in line with the his confidence in the tip.
  7. Some old rules still apply: many small bets to spread the risk.
  8. Start with all the variables, trim down using logic to the variables that matter. Benter’s model monitored only about 20 inputs — just a fraction of the infinite factors that influence a horse’s performance.
  9. Get the arcane historical data if the variable matters. Benter became convinced that horses raced differently according to temperature, and when he learned that British meteorologists kept an archive of Hong Kong weather data in southwest England, he traveled there by plane and rail. When he entered the data into his computers, found it had no effect whatsoever on race outcomes. Such was the scientific process. Turn over every stone to find the variables that drive prediction.
  10. Never be too proud: build on existing models. Sometimes the breakthrough is right infront of you, in plain sight. For Benter, it was the Jockey Club’s publicly available betting odds. Using the public odds as a starting point and refining them with his proprietary algorithm was dramatically more profitable than building his own set of odds from scratch.
  11. Know when to fold. Benter spent three summers developing a system to bet on baseball and only broke even — for him, a stinging professional defeat. America’s pastime was just too unpredictable. When competition eroded profits in Hong Kong, Benter started betting on horse races in the US. “There is a golden age for a particular market. When there aren’t many computer players, the guy with the best system can have a huge advantage.”

Partner @indbio (IndieBio.co) @sosv. Prev: @fusionfundvc @singularityu @suNordic @jpmorgan @GoldmanSachs @ucsf @UChicago @yale @LMU_Muenchen

Partner @indbio (IndieBio.co) @sosv. Prev: @fusionfundvc @singularityu @suNordic @jpmorgan @GoldmanSachs @ucsf @UChicago @yale @LMU_Muenchen