Thoroughly Testing Tipsters, aka, Parsing the Proficiency of Pundits
I like the movie “Let It Ride” starring Richard Dreyfus. It’s a fun movie based around a nice man with a gambling problem, and a heart of gold. At one point in the movie, there is probably the best example of Contrarian betting I’ve ever seen. Contrarian betting is essentially betting ‘contrary’ to popular opinion. I always thought of this as assuming that the experts don’t know anything.
So, when I found a dataset of Horse betting Tipsters (Kaggle dataset), I thought that I’d explore the data to see just how useful/useless they might be.
It is to be noted that this dataset has been kept up by it’s creators very, very well. There were no null or NaN values to contend with, the metadata was extensive and easily readable, and the imported columns were all formatted:
I reformatted a few columns and changed the Odds from Decimal format to Percentage for ease of use. The resulting dataframe looks like this:
[Terminology Note: An Each Way bet consists of both a bet that your horse will Win, and a bet that your horse will Place (winning places differ depending on event)]
Once I got my dataframe where I wanted it, I had to decide how I would analyze the data. How could I examine the expertise of my Tipsters? Well, the probabilities involved with any organized game of chance will always be set up such that, over time, the house will make a profit. This includes horse racing. Given a large number of instances, the mean of the wins will be low enough to guarantee that the track makes a profit. Unfortunately, I wasn’t able to find an open-source dataset that I could use to estimate that mean value. In lieu of that, I did what math students have been doing in introductory statistics and probability classes for centuries.
1) Comparing the Tipsters to choices made at random
Using the Racecard information (Alltheraces.com)for the UK for 06/27/2019 I found that the average number of horses in any single race is approx. 10 (exact value is 10.22, but I chose to round down to the super convenient option). I then grouped my data by Tipster, and used Seaborn to visualize the mean of Odds (%) vs the mean of Result.
As we can see, Tipsters V and O don’t even make the random chance cutoff. And while Tipsters H, T and W are right at the edge, they are indeed better than the average Joe off the street at betting on horse races. Now let’s set a slightly higher bar, the Tipsters themselves.
Comparing the Tipsters based on Results
After establishing that the majority of our tipsters are at least better that the average Joe, I decided to compare the Tipsters to each other, using the same values (means of Odds and Results). I hoped that this would help me further parse out the less proficient tipsters. I used Seaborn’s regplot here, with a bootstrap confidence interval of 95%.
From the graph, we see that regression line has a positive slope. This is reasonable. We would expect that results would be better as the odds of success increase. We also see that our barely and below average Tipsters H, T, W, O, & V all fall below the regression line, which was expected. The graph also lets us see that Tipster A has the highest mean Result value, and is over 300% better than random choice. IE their win-loss ratio was the best of our group. If that is how we choose to measure expertise, then we’ve Parsed our Pundits. However, there is another way to measure the expertise of a gambler.
Comparing the Tipsters based on (potential) Winnings
There was absolutely no data about winnings, or money that was bet in my dataset. To add this column, I used a standard of a $100 bet on the tip given.
[Terminology note: The amount you win on a bet can always be calculated using the odds, but the formula is different for every odds format. In this case the amount you would win is (Bet x 1/Odds%)]
By doing this I was able to investigate another measure of expertise as well. How well do the Tipsters do when making bets with poor odds of success? The worse the odds the better the payouts. With a standard $100 bet, I hoped that the data would show me a new picture.
And it has. The negative slope of the Regression line tells us that winnings get smaller as the odds get better, which is expected. We also see that Tipster G is far and above the most Expert tipster when measured in this way. While no longer top Tipster, Tipster A does have the best winnings for Odds (%) greater that 20, and is in the top 5 overall.
I’m very glad that this project allowed me to take a scene from a movie I love and look at it through a different lens. I’m also glad that I had the chance to work from top to bottom with a dataset that interested me, and analyze it in different ways. I was able to visualize how using different metrics to measure a qualitative feature can yield different results. And of course, see that not all experts are useless.