Betting on ITF tennis tournaments using statistical models

[kuklachert]

Hi all,

in this thread I'll be sharing the results of my bets on ITF tournaments which I place using a model-based approach.

Why ITF?

ITF tournaments (lowest tier of professional tennis) seem to be an imperfect betting market where information is often scarce, so I assume bookies do not always know exactly what they're doing when they're offering odds. For Challenger/ATP/WTA tournaments my approach didn't work so well (albeit on a small sample) so I stopped betting there and concentrated on ITF. I bet on both men's and women's tournaments (no W60, W80, W100 because they attract a player field of a comparable level to ATP challengers which I find hugely unpredictable). The best bookies for ITF tournaments are bet365, Betclic. Bwin and Unibet occasionally also offer ok odds.

What does my model do?

It's a relatively simple Logistic Regression that predicts each player's win probabilities. Following main factors go into the model:

• ELO rating (several versions calculated with different formulae - some place more weight in recent form, others are more influenced by long-term results)
• Recent changes in ELO rating
• Players' form on the surface played
• Head-to-head (overall/on surface played)
• Home advantage

Additional factors I apply on top of the model as an "expert judgment" are:

• Are "rising talents" involved in a match? (they tend to have ELO rankings lower than their true level)
• Are players from different playing fields involved? (In case 2 players meet, one of which mainly plays in Challengers and the other in Futures, the ELO of the Challenger player tends to underestimate his relative level because he plays in a stronger field; similarly, players from the Asian fields tend to be weaker than their European/American counterparts with the same ELO)
• Has the model performed well with the players I'm betting on/against? (if I bet against Roger Federer twice this week and he wins both times, I'm not likely to bet another time even if the model recommends me to)
• Do players have sufficient match form? (I won't bet on a player who hasn't played for 6 months if the model recommends me to)
• Tiredness (I won't bet on a player who yesterday played a 4-hour match).

The probabilities generated by the model are then compared with to odds. I only bet on odds between 1.4 and 2.9. For higher odds, I found the model to be unreliable, for lower odds it just doesn't make sense as I live in Germany where all winnings on betting are taxed 5%. To account for model risks I apply an additional margin of conservatism to select my bets.

So if the model tells me that Player A wins with 60% probability, the lowest odds I want to take on Player A would be (1/0.6)*(1/0.95)*1,1=1.93. The (1/0.95) multiplier accounts for the betting tax, the 1.1 multiplier accounts for model error. In case I'm really convinced by the players' stats I might accept odds of 1.85-1.90.

The model is calibrated on an extensive history (ITF men's and women's matches starting 2001).

Final words

The 5% tax makes profitable betting in Germany pretty much impossible, so I only bet small amounts for fun. In this thread, I will show both profit before and after tax so people in other countries get a realistic view of the model performance. So far after ~950 bets I have achieved an ROI of around 7% before tax and 2% after tax which I find decent. I will always bet the same amount, let's say a symbolic 5 EUR on each match. Let's hope I can maintain my ROI after I start posting. I also hope this thread attracts some tennis betting enthusiasts, especially those who use statistical/ML methods and are willing to share their secrets because I truly believe a scientific approach can make betting profitable.

Edited February 5, 2020 by kuklachert

[kuklachert]

So let's start, my first picks are:

• Dusty H. Boyer v Simon Carr, M15 Cancun, 5 EUR, 2.10 with bet365
• Arlinda Rushiti v Victoria Kalaitzis, W15 Monastir, 5 EUR, 1.50 with bet365

Edited February 5, 2020 by kuklachert