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!
Previous 2020 Predictions:
- Week 1: Kickoff
_________________________________________________________________________________
Recap:
8-8. Not a great start, especially with an expected predicted probability of 10.78 games correct. Some games went the way I personally expected contrary to the model (e.g., Bills, Buccaneers, Seahawks). The surprise was Saints losing to Raiders. But that's football, right? Maybe the model is a little nervous about its first performance of the year...
On to next week.
_________________________________________________________________________________
Week 2:
First week of predictions! I have my dusty code up and running, which had last been run over two years ago. After package updates, syntax changes, name changes (i.e., Washington), and other code adjustments I have a first successful pass at the predictions. So I begin with a model and code as good as I left it, which looking back on it now, isn't great. But it's a place to start.
A logistic regression model on the data yields a 72% accuracy on training data and a 72% accuracy on test data. AUC is 70% for both training and test. This is consistent with past usage of the model. Sensitivity and Specificity are roughly the same, and F1 is also about 72%. So the model is fairly balanced in predicting positives/wins and negatives/losses.
In practice, the model yielded about 61% accuracy in 2016 and 63% accuracy in 2017. Fivethirtyeight.com scored 64% in 2016 according to my calculations, and Elliot Harrison has a lifetime record of 65.5%. I can't remember where I read this, but I seem to remember reading an article that hypothesized something like a 70% upper bound prediction accuracy, calculating that 30% of a win is pure random luck instead of team skill or decision making (e.g., ball bounces this way vs. that way, wind direction, etc.). So a good goal for this season is to beat my own previous seasons, perhaps having an overall accuracy of 64%, with the realistic understanding that I am approaching the upper bound of what is possible in this field. However, I am in good company amongst these other predictors.
Oddities and Bold Predictions:
Not having rigorously followed NFL player movements or team trends in the past two years, I do not quite have the sense for which teams are "good" and "ought" to win. But here are some predictions that stand out to me.
- Browns vs. Bengals
- Browns win. I guess the Browns are good now? 538 agrees.
- Dolphins vs. Bills
- Dolphins beat Bills. Personally, I'd go the other way. 538 has Bills winning. We shall see.
- Eagles vs. Rams
- Eagles win. 538 has the Rams winning but barely (51%).
- Panthers vs. Buccaneers
- Panthers win? Probably not. My model doesn't truly grasp the meaning of "Tom Brady". 538 also picks Buccaneers to win.
- Patriots vs. Seahawks
- The model picks Patriots, but I'd personally go with Seahawks. With the Patriot's loss of Brady, and playing away, while Seahawks are at home and have lots of stability from the prior season, I'd think Seahawks are a much better pick. And I'm not just saying that as a Seahawks fan... 538 chooses Seahawks.
- Washington vs. Cardinals
- My model picks Washington with 79% and 538 picks Cardinals with 68%. Are the Cardinals good again? I know they beat SFO last week. I don't know what to make of this one.
So some oddities and perhaps bold/crazy predictions. I may get lucky, right? You never know... It's a new season! Good luck!
Here are the predictions for Week 2:
No comments:
Post a Comment