Below are updated predictions based on the actual results of games played through Sunday, September 1 and predicted results of games not yet played. For background information this week, following the predictions there is some information on the win-loss-tie likelihood part of the prediction process.
Team NCAA RPI and Balanced RPI Ranks, Plus NCAA RPI and Balanced RPI Strength of Schedule Contributor Ranks
Conference NCAA RPI and Balanced RPI Ranks Plus NCAA RPI Strength of Schedule Contributor Ranks
Predicted NCAA Tournament Automatic Qualifiers, Disqualified Teams, and At Large Selection Status, All for the Top 57 Teams
Win-Loss-Tie Likelihoods
To predict the results of games not yet played, over the first part of the season my program bases them on assigned pre-season NCAA RPI ratings for teams. Those ratings are based on team ranks over the last 7 years. Using the predicted ratings, the program determines the rating difference between each set of opponents, as adjusted for home field advantage. For that rating difference, the program then assigns a win, loss, and tie likelihood for the game.
The win, loss, and tie likelihoods are based on a study of all games played since 2010 (over 40,000 games, with pre-2022 game results adjusted to treat games decided in overtime as ties). The study looks at the location-adjusted rating differences between teams and the actual game results. It produces a Result Probability Table that shows the win, loss, and tie likelihoods in relation to the location-adjusted rating differences between opponents. For each of this year's future games, the program determines the location-adjusted rating difference between the opponents and then extracts from the Result Probability Table the win, loss, and tie likelihoods for their game. The program then tallies teams' actual results for games already played and their win, loss, and tie result probabilities for games not yet played to produce teams' predicted end-of-season records. From there, the program computes teams' predicted RPI ratings and ranks.
The Result Probability Table is reliable, when applied to a large number of games. For the more than 40,000 games played since 2010, using the RPI bonus and penalty adjustment regime in effect in 2023, but with tie values changed from 1/2 a win to 1/3 of a win in the RPI computation process as is expected for this year, here is how the the actual results for the better rated team in each game compare to the predicted results using the Result Probability Table:
Higher rated team actually wins: 65.08%
Higher rated team actually ties: 21.12%
Higher rated team actually loses: 13.80%
......................................................................
Higher rated team predicted wins: 65.06%
Higher rated team predicted ties: 21.18%
Higher rated team predicted losses: 13.76%
This means that, from an overall perspective, when this year's predicted results miss the mark, it is not due to problems with the calculated result probabilities, but rather is due to the pre-season assigned ratings used in predicting game results turning out to not reflect actual team strength. When this happens, of course, it is not surprising since it is impossible to predict in advance with great accuracy what teams' true strength will be.
No comments:
Post a Comment