For my tennis fan readers and/or technology hobbyists…
I’ve been talking with the SportsPredict team about their Grand Slam Challenge, a free forecasting competition built around the US Open.
The idea is simple: you predict across the men’s and women’s singles draws, put probabilities behind your calls, and see how you do against everyone else on the leaderboard.
There are prizes for the top finishers, but the part I’m most curious about is the bot competition. While I’m conveniently skeptical that AI can replace writing at House of Strauss, I’m open to the idea that human-directed automation can succeed in the sports prognostication space. My suspicion is that the best way to attack something like this is to build a prediction bot, feed it the right data, and let it make forecasts throughout the tournament.
I do not currently know how to do that, but I’m the rare Substacker whose audience is generally smarter than its author. I’m guessing some of you actually understand how to pull this off.
So I want to try an experiment: if you know how to build a forecasting bot — or want to figure it out — enter the Challenge and tell me what you’re doing. What data are you using? What model? What are you learning? Is your bot actually any good?
I’m going to start playing around with this myself, partly because SportsPredict has a much bigger competition coming later and I want to be ready to build an HOS bot that can actually compete.
Consider this the preseason.
The Grand Slam Challenge covers the singles Open matches and hundreds of forecasting questions. You can compete manually or with a bot.
Prizes include an iPad Pro for first place, gift cards and cash-equivalent prizes for other top finishers, plus a separate prize for the top bot.
https://join.sportspredict.com/gsc-st
If you build something interesting, send it my way. I want to see how people approach this.
I'd love to see some bots as well, but I'm not optimistic we'll see any for a very simple reason. The number of people who will participate in this contest just isn't big enough for their to be overlap with HoS subscribers. There might be a few thousand optimistically in both categories, and that's just not enough for a chance of a meaningful overlap.
I won't be participating myself, but just in case anyone's curious I have built systems like this in the past. I don't think the new class of LLMs will change things much, other than they could make it easier to build "prediction slop". The state-of-the-art is probably still machine learning prediction models like XGBoost, Random Forests, Cat Boost, etc... These are models that are built off of historical data from prior statistics about the players and outcomes from previous matches and / or aggregated predictions from other sources like gambling websites. I'd expect that "and / or" to be the major differentiator between winning bots and also-rans, but I couldn't tell you what the right answer would be without actually doing the hard work.
The other wrinkle on what I mentioned above is that there are a new class of machine learning models called "pre-trained models" that would probably outperform the above mentioned "XGBoost, Random Forests, Cat Boost". These are kind of interesting because they've been fed information about tens of thousands of other predictions tasks, like "predict how much people will spend on music next year" "predict which hospital patients die" or "predict what the weather will be next week" and somehow they can use this apparently unrelated information to somehow get smarter on a new task, like "predict who will the US Open". If I had to guess somebody using one of those pre-trained models and absorbing player stats and / or betting odds from gambling sites will win.