Inspired by the story of BillBenter, a gambler whodeveloped a computer model that made him close to a billion dollars1betting on horse races in the Hong Kong Jockey Club (HKJC), I set out tosee if I could use machine learning to identify inefficiencies in horseracing wagering.
- William Benter Pittsburgh
- Bill Benter Gambler
- William Benter Horse Racing Software
- Bill Benter Book
- William Benter Horse Racing Software
American gambler William Benter developed arguably the most successful computer software in the world to attack the giant Hong Kong horse racing market, where the pools routinely reach into the hundreds of millions of dollars. During the 2010 racing season alone the betting turnover reached HK$71.65 billion, on just two racetracks! William Benter (Bill Benter) William Benter is the legendary horse-racing gambler who used computer software to play the Hong Kong Horse Racing market and win millions of dollars. Gamproadmin - March 8, 2012 Amercian William 'Bill' Benter completes our fantastic Horse Racing Trifecta, including Alan Woods and Zeljko Ranogajec. After starting in partnership with Alan Woods, Bill Benter left to form his own operation after the pair had what some describe as a 'spectacular falling out'. To build a good horse racing model, teams rely on workers with the skills of hedge fund technicians. Rumor has it that one of the teams has wooed programmers from Fortune 500 companies. 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 system. This book examines the elements necessary for a practical and successful computerized.
Data
The Hong Kong Jockey Club publishes all of the results for each race ontheir website. Icreated a script to scrape result cards from all of the historicalavailable races. After running the script I was left with a dataset of938 races spanning 14 months.
Result card from a HKJC race.
Feature Engineering
Going into this project, I had no industry knowledge about horse racing.Since not much information is provided with the race result cards, muchwork must be done in engineering and selecting features in order to givea model more predictive power. Listed below are the features being used.
Draw: Which gate the horse starts in. This is randomly assignedbefore the race. Horses starting closer to the inside of the track (draw1) generally perform slightly better.
Horse Win Percent: Horse's win percent over the past 5 races.
Jockey Win Percent: Jockey's win percent over the past 5 races.
Trainer Win Percent: Trainer's win percent over the past 5 races. Jose canseco celebrity boxing record.
Actual Weight: How much weight the horse is carrying (jockey +equipment).
Declared Weight: Weight of the horse.
Days Since Last Race: How many days it has been since the horse haslast raced. A horse that had been injured in its last race may have notraced recently.
Mean Beyer Speed Figure: Originally introduced in Andrew Beyer'sPicking Winners2, the Beyer Speed Figure is system for rating ahorse's performance in a race that is comparable across differenttracks, distances, and going (track conditions). This provides a way tocompare horses that have not raced under the same circumstances. Afterreading Beyer's book, I implemented his rating system on my data. Forthis feature, I calculated the mean speed figure over the horse's past 5race the race.
Last Figure: Speed figure of the last race the horse was in.
Best Figure at Distance: Best speed figure the horse has gotten atthe distance of the current race.
Best Figure at Going: Best speed figure the horse has gotten at thetrack conditions of the current race.
William Benter Pittsburgh
Best Figure at Track: Best speed figure the horse has gotten at thetrack of the current race.
Engineering more features may yield better results; Benter's model3included many different types of features from many data sources.
Model
Before creating the model, it is important to understand the goal of themodel. In order to not lose money at the race track, one must have anadvantage over the gambling public. To do this we need a way ofproducing odds that are more accurate than public odds. For example,imagine the payout of horse is 5 to 1, and we have a model thatindicates the horse's probability of winning is 0.2, or odds of 4 to 1.Assuming our model is faithful, we would have an edge in this case, aswe would be getting an expected return of $(0.2 * (5 + 1)) - 1 = 0.2$ or20%. How do we create such a model?
Let's first assume the existence of a function $R$ that provides arating $R_h$ of a horse $h$, given input features $x_h in R^m$:
[R_h = R(x_h)]Assuming a horse with a higher rating has a higher probability ofwinning, we can compute the estimated probability of horse $h$ winning,$hat{p}_h$, given the ratings of all of the horses in the race:
[hat{p}_h = frac{exp(R_h)}{sum_i exp(R_i)}]Here we use the softmax
function, as its outputs will always sum to 1,and maintain the same order as the input.
Now that we know how we'll compute our probabilities, we must define ourrating function $R$. For this we will use a neural network that takes aninput vector of length $m$ (where $m$ is the number of features), andoutputs a single scalar value. The structure of this network consists oftwo Fully Connected
layers, each followed by a ReLU
,Batch Normalization
and Dropout
layer. Lastly, there is a finalfully connected layer to produce the single output.
Now we can visualize our model:
Bill Benter Gambler
Training
We have defined our model, but how do we train it? For each race, let'scall the winning horse $w$. If we had a perfect model, the predictedprobability of $w$ winning should be 1, that is $hat{p}_w = 1$. We canencourage the model to approach this value by defining a loss functionI'll call win-log-loss:
[L(hat{p}_w) = -log(hat{p}_w)]Win-log-loss will approaches 0 as the win-probability of the winnerapproaches 1, and approaches $infty$ as the win-probability of thewinner approaches 0. Now by minimizing win-log-loss via stochasticgradient descent, we can optimize the predictive ability of our model.
It is important to mention that this method is different than a binaryclassification. Since the ratings for each horse in a race arecalculated using a shared rating network and then converted toprobabilities with softmax, we simultaneously reward a high rating fromthe winner while penalizing high ratings from the losers. This techniqueis similar to a Siamese NeuralNetwork, which isoften used for facial recognition.
Betting
Now that we have predicted win probabilities for each horse in the racewe must come up with a method of placing bets on horses. We can computeour own private odds for each horse using $1/hat{p} - 1$. Now wecould just bet on every horse whose odds exceed our private odds, butthis may lead to betting on horses with a very low chance of winning. Toprevent this, we will only bet on horses whose odds exceed our privateodds, and whose odds are less then a certain threshold, which wewill find the optimal value of over on our validation set.
Results
We split the scraped race data chronologically into a training,validation, and test set, ensuring there would be no lookahead-bias. Wethen fit the horse-rating model to our training set, checking itsgeneralization to the validation set:
After fitting the model, we find the optimal betting threshold on thevalidation set, in this case odds of 4.9
. We can now display oursimulated results of betting 10 dollars on each horse that our modelindicates. To compare, we also show the results of betting 10 dollarsevery race on the horse with the best odds, and one of the beststrategies there is, not betting at all.
We can see that always betting on the horse with the best odds is asure-fire way to lose all your money. While the model was profitableover both the training and validation sets, it is hard to say for surehow reliable it is because of the very low frequency of its betting.
Benter claimed that his era was a 'Golden Age' for computer bettingsystems, where computer technology had only recently become affordableand powerful enough to implement such systems. Today, the betting markethas likely become much more efficient with large numbers of computerhandicappers and more detailed information available to the public.Nevertheless, developing a profitable model, especially with modernmachine learning methods, may be a feasible task.
Thank you for reading! For any questions regarding this post or others,feel free to reach out on twitter:@teddykoker.
Kit Chellel, The Gambler Who Cracked the Horse-Racing Code,2018,https://www.bloomberg.com/news/features/2018-05-03/the-gambler-who-cracked-the-horse-racing-code. ↩
Andrew Beyer, Picking Winners: A Horseplayer's Guide (HoughtonMifflin Harcourt, 1994). ↩
William Benter, 'Computer Based Horse Race Handicapping andWagering Systems: A Report,' in Efficiency of Racetrack BettingMarkets (World Scientific, 2008), 183–98. ↩
Born | William Benter 1957 (age 63–64) |
---|---|
Alma mater | Case Western Reserve University University of Bristol |
Occupation | professor Businessman Gambler |
Years active | 1984 –present |
Spouse(s) | (m. 2012) |
Children | 1 |
William Benter Horse Racing Software
William Benter (born 1957) is an American and Hong Kong professional gambler and philanthropist who focuses on horse betting. Benter earned nearly $1 billion through the development of one of the most successful analysis computer software programs in the horse racing market.[1]
Benter has served as president of Hong Kong Rotary Club,[2] founded the Benter Foundation, is chairman and International CEO of Acusis LLC, Pittsburgh, Pennsylvania, and occasionally lectures university students on subjects like statistics and mathematical probability.[3]
Benter is a philanthropist donating to charitable causes both in Hong Kong and the United States. [4][5][6]
Bill Benter Book
Early life and career[edit]
William Benter was born and raised in Pittsburgh, Pennsylvania.[2] As he grew up, he wanted to use his mathematical talents to make a profit so immediately after finishing a university physics degree in 1977,[3] he went to the blackjack tables in Las Vegas and used his skills to count cards. He came across the book, Beat the Dealer, by Edward O. Thorp, which helped him improve his methods.[7] Seven years later, he was banned from all of Vegas' strip's casinos.[2]
Benter then met with Alan Woods, a like-minded gambler whose expertise in horse racing complemented his own in computers. The two became racing partners and in 1984, moved to Hong Kong.[3] Starting with a mere US$150,000 (equivalent to US$369,138 in 2019), the pair relied on their mathematical skill to create a formula for choosing race winners.[2]
Using his statistical model, Benter identified factors that could lead to successful race predictions. He found that some came out as more important than others.[8] Benter later worked with Robert Moore.[9]
Benter is a visiting professor at the Southampton Management School[10] as part of the Centre for Risk Research and a fellow of the Royal Statistical Society.[11]
In 2007, Benter founded the Benter Foundation.[12]
Personal life[edit]
In March 2012, Benter married Hong Kong National Vivian Fung in a Tibetan Buddhist rite.[13] In 2015 they had their first child Henry.[14] Benter is currently residing in Pittsburgh.[15][16]
Philanthropy[edit]
Benter is a big contributor to charity and political groups. According to political campaign contribution records, in 2008 Benter and Acusis were listed as donors to Barack Obama's presidential campaign and the Democratic Party of Virginia. In 2010, The Advantage Trust donated to Israeli-based organization Rabbis for Human Rights.[17]The Atlantic reported in 2010 that Benter had raised and given at least US$800,000 in support to J Street.[18][19]
In 2012, Benter donated one million dollars to the University of Pittsburgh.[20] In 2013, Fox News reported that Benter donated thousands of dollars for pro-Hagel ads in Politico when he was nominated to be next Secretary of Defense by President Obama.[21]
Engineering more features may yield better results; Benter's model3included many different types of features from many data sources.
Model
Before creating the model, it is important to understand the goal of themodel. In order to not lose money at the race track, one must have anadvantage over the gambling public. To do this we need a way ofproducing odds that are more accurate than public odds. For example,imagine the payout of horse is 5 to 1, and we have a model thatindicates the horse's probability of winning is 0.2, or odds of 4 to 1.Assuming our model is faithful, we would have an edge in this case, aswe would be getting an expected return of $(0.2 * (5 + 1)) - 1 = 0.2$ or20%. How do we create such a model?
Let's first assume the existence of a function $R$ that provides arating $R_h$ of a horse $h$, given input features $x_h in R^m$:
[R_h = R(x_h)]Assuming a horse with a higher rating has a higher probability ofwinning, we can compute the estimated probability of horse $h$ winning,$hat{p}_h$, given the ratings of all of the horses in the race:
[hat{p}_h = frac{exp(R_h)}{sum_i exp(R_i)}]Here we use the softmax
function, as its outputs will always sum to 1,and maintain the same order as the input.
Now that we know how we'll compute our probabilities, we must define ourrating function $R$. For this we will use a neural network that takes aninput vector of length $m$ (where $m$ is the number of features), andoutputs a single scalar value. The structure of this network consists oftwo Fully Connected
layers, each followed by a ReLU
,Batch Normalization
and Dropout
layer. Lastly, there is a finalfully connected layer to produce the single output.
Now we can visualize our model:
Bill Benter Gambler
Training
We have defined our model, but how do we train it? For each race, let'scall the winning horse $w$. If we had a perfect model, the predictedprobability of $w$ winning should be 1, that is $hat{p}_w = 1$. We canencourage the model to approach this value by defining a loss functionI'll call win-log-loss:
[L(hat{p}_w) = -log(hat{p}_w)]Win-log-loss will approaches 0 as the win-probability of the winnerapproaches 1, and approaches $infty$ as the win-probability of thewinner approaches 0. Now by minimizing win-log-loss via stochasticgradient descent, we can optimize the predictive ability of our model.
It is important to mention that this method is different than a binaryclassification. Since the ratings for each horse in a race arecalculated using a shared rating network and then converted toprobabilities with softmax, we simultaneously reward a high rating fromthe winner while penalizing high ratings from the losers. This techniqueis similar to a Siamese NeuralNetwork, which isoften used for facial recognition.
Betting
Now that we have predicted win probabilities for each horse in the racewe must come up with a method of placing bets on horses. We can computeour own private odds for each horse using $1/hat{p} - 1$. Now wecould just bet on every horse whose odds exceed our private odds, butthis may lead to betting on horses with a very low chance of winning. Toprevent this, we will only bet on horses whose odds exceed our privateodds, and whose odds are less then a certain threshold, which wewill find the optimal value of over on our validation set.
Results
We split the scraped race data chronologically into a training,validation, and test set, ensuring there would be no lookahead-bias. Wethen fit the horse-rating model to our training set, checking itsgeneralization to the validation set:
After fitting the model, we find the optimal betting threshold on thevalidation set, in this case odds of 4.9
. We can now display oursimulated results of betting 10 dollars on each horse that our modelindicates. To compare, we also show the results of betting 10 dollarsevery race on the horse with the best odds, and one of the beststrategies there is, not betting at all.
We can see that always betting on the horse with the best odds is asure-fire way to lose all your money. While the model was profitableover both the training and validation sets, it is hard to say for surehow reliable it is because of the very low frequency of its betting.
Benter claimed that his era was a 'Golden Age' for computer bettingsystems, where computer technology had only recently become affordableand powerful enough to implement such systems. Today, the betting markethas likely become much more efficient with large numbers of computerhandicappers and more detailed information available to the public.Nevertheless, developing a profitable model, especially with modernmachine learning methods, may be a feasible task.
Thank you for reading! For any questions regarding this post or others,feel free to reach out on twitter:@teddykoker.
Kit Chellel, The Gambler Who Cracked the Horse-Racing Code,2018,https://www.bloomberg.com/news/features/2018-05-03/the-gambler-who-cracked-the-horse-racing-code. ↩
Andrew Beyer, Picking Winners: A Horseplayer's Guide (HoughtonMifflin Harcourt, 1994). ↩
William Benter, 'Computer Based Horse Race Handicapping andWagering Systems: A Report,' in Efficiency of Racetrack BettingMarkets (World Scientific, 2008), 183–98. ↩
Born | William Benter 1957 (age 63–64) |
---|---|
Alma mater | Case Western Reserve University University of Bristol |
Occupation | professor Businessman Gambler |
Years active | 1984 –present |
Spouse(s) | (m. 2012) |
Children | 1 |
William Benter Horse Racing Software
William Benter (born 1957) is an American and Hong Kong professional gambler and philanthropist who focuses on horse betting. Benter earned nearly $1 billion through the development of one of the most successful analysis computer software programs in the horse racing market.[1]
Benter has served as president of Hong Kong Rotary Club,[2] founded the Benter Foundation, is chairman and International CEO of Acusis LLC, Pittsburgh, Pennsylvania, and occasionally lectures university students on subjects like statistics and mathematical probability.[3]
Benter is a philanthropist donating to charitable causes both in Hong Kong and the United States. [4][5][6]
Bill Benter Book
Early life and career[edit]
William Benter was born and raised in Pittsburgh, Pennsylvania.[2] As he grew up, he wanted to use his mathematical talents to make a profit so immediately after finishing a university physics degree in 1977,[3] he went to the blackjack tables in Las Vegas and used his skills to count cards. He came across the book, Beat the Dealer, by Edward O. Thorp, which helped him improve his methods.[7] Seven years later, he was banned from all of Vegas' strip's casinos.[2]
Benter then met with Alan Woods, a like-minded gambler whose expertise in horse racing complemented his own in computers. The two became racing partners and in 1984, moved to Hong Kong.[3] Starting with a mere US$150,000 (equivalent to US$369,138 in 2019), the pair relied on their mathematical skill to create a formula for choosing race winners.[2]
Using his statistical model, Benter identified factors that could lead to successful race predictions. He found that some came out as more important than others.[8] Benter later worked with Robert Moore.[9]
Benter is a visiting professor at the Southampton Management School[10] as part of the Centre for Risk Research and a fellow of the Royal Statistical Society.[11]
In 2007, Benter founded the Benter Foundation.[12]
Personal life[edit]
In March 2012, Benter married Hong Kong National Vivian Fung in a Tibetan Buddhist rite.[13] In 2015 they had their first child Henry.[14] Benter is currently residing in Pittsburgh.[15][16]
Philanthropy[edit]
Benter is a big contributor to charity and political groups. According to political campaign contribution records, in 2008 Benter and Acusis were listed as donors to Barack Obama's presidential campaign and the Democratic Party of Virginia. In 2010, The Advantage Trust donated to Israeli-based organization Rabbis for Human Rights.[17]The Atlantic reported in 2010 that Benter had raised and given at least US$800,000 in support to J Street.[18][19]
In 2012, Benter donated one million dollars to the University of Pittsburgh.[20] In 2013, Fox News reported that Benter donated thousands of dollars for pro-Hagel ads in Politico when he was nominated to be next Secretary of Defense by President Obama.[21]
In 2016, The Washington Post reported the Benter raised US$100,000 for A New Voice for Maryland, a pro-Joel Rubin group for Democratic nomination in Maryland's 8th Congressional District.[22]
William Benter Horse Racing Software
See also[edit]
References[edit]
- ^Colon, Nicholas G. (May 17, 2016). 'Inside The Blackjack Ball: An Exclusive Look At The Gathering Of The Smartest Gamblers In The World'. Forbes. Retrieved April 20, 2017.
- ^ abcdDelvecchio, Jerry. 'William Benter (Bill Benter), the richest and most successful gambler of all time? Hong Kong Horse Racing Legend – The Worlds Greatest Gamblers'. worlds-greatest-gamblers.com. Retrieved April 20, 2017.
- ^ abcChung, Yulanda (November 3, 2000). 'The Winning Edge'. CNN. Retrieved April 20, 2017.
- ^https://doi.org/10.1016/j.cis.2017.12.008
- ^https://www.causeiq.com/organizations/benter-foundation,208807953/
- ^https://www.bloomberg.com/news/features/2018-05-03/the-gambler-who-cracked-the-horse-racing-code
- ^'Bill Benter – One of the Wealthiest Gamblers in the World'. gamblingsites.org. Retrieved April 20, 2017.
- ^Kucharski, Adam (February 25, 2016). 'Are the Best Gamblers Skilled, or Just Lucky? From poker to horse racing, the statistics involved in coming out on top'. Discover Magazine. Retrieved April 20, 2017.
- ^https://www.themonthly.com.au/monthly-essays-tony-wilson-mr-huge-alan-woods-and-his-amazing-computer-nags-riches-story-149
- ^'Professor Bill Benter, University of Southampton'. University of Southampton. Retrieved April 20, 2017.
- ^'Who is Bill Benter?'. benterfoundation.org. Retrieved April 20, 2017.
- ^https://benterfoundation.org/board-staff/
- ^'Bill Benter – Pittsburghs Beautiful People'. pittsburgh-legends.com. Retrieved April 20, 2017.
- ^Bencivenga, Natalie; Bauknecht, Sara (May 6, 2016). 'Motherly wisdom from some new, and not so new, Pittsburgh moms'. Pittsburgh Post-Gazette. Retrieved April 21, 2017.
- ^'William Benter, an American expert horse racing gambler who found success using racing form analysis'. thegreattipoff.com. Retrieved April 20, 2017.
- ^'The Gambler Who Cracked the Horse-Racing Code'. Bloomberg L.P. May 3, 2018. Retrieved May 4, 2018.
- ^Steger, Isabella (September 30, 2010). 'Jewish Advocacy Group Donor Linked to Hong Kong'. The Wall Street Journal. Retrieved April 20, 2017.
- ^Good, Chris. 'J Street's Half-Truths and Non-Truths About Its Funding'. The Atlantic. Retrieved April 21, 2017.
- ^'Soros revealed as funder of liberal Jewish-American lobby'. The Washington Times. Retrieved April 21, 2017.
- ^'Bill Benter, the richest of them all, but who is he? – Gambler Profiles'. gambler-profiles.com. Retrieved April 20, 2017.
- ^'Media Matters funder bankrolls pro-Hagel campaign'. Fox News Channel. January 8, 2013. Retrieved April 21, 2017.
- ^Turque, Bill (April 4, 2016). 'Middle East peace activist wants to take fight to Republicans in Congress'. The Washington Post. Retrieved April 21, 2017.
Bibliography[edit]
- Hausch, Donald B.; Lo, Victor S. Y.; Ziemba, W. T. Efficiency of Racetrack Betting Markets. World Scientific. pp. 183–198. ISBN9789812819192. Retrieved April 21, 2017.
External links[edit]
- 'An Interview with Bill Benter'. macau.rotary3450.org. Rotary Club of Macau. Retrieved April 20, 2017.
- 'The Jockey Project'.