Welcome! I intend this to be an ongoing project of predicting NFL game outcomes, point spreads, and final scores. My hope is that my models will improve over time as they become more sophisticated and use better data. I will try to regularly publish predictions to keep myself accountable, and will make observations along the way about what is working and what isn't. See below for latest updates. Enjoy!
_________________________________________________________________________________
Recap:
7 - 8 - 1. With an expected predicted value of 10.2, actual result was 7.5 (the tie is "half" correct). Not going so well thus far. Some definite surprises as Lions beat Cardinals and Bengals tied Eagles. I expected Seahawks to win against my model's prediction and I wasn't surprised to see the other missed predictions as most were close probabilistically. That's the way it is some times. Hopefully things improve next week.
_________________________________________________________________________________
Week 3:
Week 3 brings a retrained logistic model. As the game has changed somewhat since 2017, a model that relies on features and weighting that was predictive in 2017 may not be as effective in predicting games in 2020. Hence, the need for retraining.
The result of the retraining is a logistic model that predicts with training and test accuracy of about 73% (11 or 12 predictions of 16 per week). In practice, the accuracy has been much lower, suggesting that I am overfitting in some way still (and will need to do some investigation as to what needs changing). Important features in order of importance include:
- Rolling win ratio of team over past 16 games
- Team prior wins in season
- Team season win ratio
- Rolling win ratio of team over past 32 games
- Opponent's years since playoffs
- Opponent's season win ratio
- Opponent's years since playoffs with the same coach
- Opponent's years with same coach
- Opponent's top 1 passer playing in the game
- Team's top 1 passer player in the game
- Team's years since playoffs with the same coach
A lot of these variables are collinear and a future to-do item will be to test for and remove collinear variables.
Oddities and Bold Predictions:
- Dolphins vs. Jaguars
- 538 picks Jaguars but but my model chooses Dolphins by quite a bit. Dolphins are 0-2 against Patriots and Bills and did fairly well. Jaguars 1-1 against Colts and Titans. I am sort of inclined to go with my model on this one.
- Bills vs. Rams
- Essentially a toss up. Rounded to three decimal places the model predicts 0.500 victory for Rams. 538 goes for Bills.
- Browns vs. Washington
- My model has Browns losing, but 538 has them winning. Both are 1-1 but Washington played (IMHO) better teams. I'd choose Washington for this one.
- Cowboys vs. Seahawks
- My model picks against Seahawks again. But Seahawks are home and are 2-0 while Dallas is 1-1. I'd definitely go with Seahawks, as does 538. I don't know why my model doesn't like the Seahawks.
- Packers vs. Saints
- The model picks the Saints to win, but I don't know. Packers are 2-0 and Saints are 1-1 although they will be playing at home. 538 sides with Saints, and though it's a close call, I think I'd still pick Saints.
Good luck in week 3!
Here are the predictions for Week 3:
No comments:
Post a Comment