-
@ Real Man Sports
2024-09-29 08:47:44I’m sure people are already doing this, but I know someone who runs a company that builds models, and I had some ideas. One of them is to divide the AI agents into two groups of bettors: (1) Situational and (2) Fundamentals. The former group would have parameters like days since the last game, travel distance, whether they are coming off a win or loss, closeness of the win or loss, difference in weather conditions from the norm, more than one consecutive road game, national game or not, etc. All the situational correlation stuff many consider when betting.
The latter would solely focus on yards per play, play success rate, YPP allowed, success rate allowed, big plays, big plays allowed, first downs, etc. It would be solely based on performance and ignore situation.
The idea would be to throw in 20 years of play data, weather, situational and ATS results and figure out which parameters are most predictive. Use those for each agent (maybe weight more or less on particular ones for each) and have them bet and compete against one another. The results should tell you what is more priced in — performance or situation?
You could also do things like have some of the agents for each ignore games where the QB was starting his first 3 games for either team, eliminating discontinuity and the noise of backups. There are hundreds of parameters and tweaks you could make, but the idea is to take the best performing of both and when they align on a game, you bet it, track the results.
I really should have been more of a tech guy — too much partying in college and afterwards. But maybe the tech will progress to a point where regular people will be able to create advanced models solely with verbal instructions.
My other big idea, and I’ve expressed it before, is only to look at outliers — games where the spread was covered by more than two TDs on either side. See what was it ahead of those games that teams (winning and losing) had in common. I’m sure turnovers would be big, but I would look at games where the surprise result happened in the performance metrics, not just on the scoreboard too. Build a model of extreme over or underperformance, see if there’s not something that could be captured, either via situation or metrics.
The holy grail would be to develop rules to analyze matchup anomalies. Extreme run-heavy vs extreme pass, different base defenses vs base offenses, team speed and size matchups, coaching styles etc. Maybe there are combinations where the styles portend anomalous results.
It’s too much data, and there are too many variables for a person to comb through. But that’s where AI can help.