We use the SpringRank algorithm to infer hierarchies in ATP Men's tennis tournament data. Using the SpringRank rankings, we can infer the "real" probability of one player beating some other player. If the implied probability of the money-line odds being offered for a particular match differ significantly from our real probabilities, it would make sense to lay a bet on the "undervalued" player (assuming our generated rankings are accurate).
Our program (atp_mens_testing.py in SpringRank/python dir) expects a set of historical data from which it can infer a SpringRank hierarchy. It also expects a test set of data. This could be a future match(es) in the form of an adjacency matrix:
player_i player_j 1 player_i2 player_j2 1 . . .
Tests of our program were run where the set of historical data was all ATP Men's Tennis Data from the year 2000 to the year being tested (exclusive) and the set of test data was the ATP Men's Tennis tournament results for the year being tested.
The data processing tool get_data.py in the tennis_prep dir was used to process the raw data into an adjacency matrix. It can be run like so:
python3 get_data.py [startYear] [endYear(exclusive)]
and with the optional arguments:
--surface | to include the type of surface played on in adjacency matrix --odds | to include the odds in the output data (only useful for the year being tested)
Our program is currently designed as a "testing environment." So, it decides whether or not to lay a bet based on a predefined divergence in our SpringRank calculated odds and the inferred money-line odds. The program has a predefined bankroll to simulate concrete real world results of using the SpringRank algorithm to calculate the "real" probabilities of players and sizing bets using the Kelly Criterion.
A write up of the program and the tests that were run can be found here:
Alternatively, you can look at the ATP_Mens_Tennis_Prediction.html file in the root of our project.
I plan on at the very least testing this on different data, like ATP women's tennis tournament data. I would also like to devise some better way of deciding whether or not to lay a bet. Currently, we bet if our "real" probability is .45 greater than the inferred money-line probability. For obvious reasons, this isn't a great solution. It's static and uncreative.