The 2015 NFL season is over. It's time to look back on my predictions to see how I did over the course of the season. For those of you just tuning in or for those interested in a refresher, here are the links to my 2015 NFL season game predictions:
How Did I Do?
Over the whole season, including playoffs, I went 79-55 with a win ratio of 59%. Without the playoffs, I went 74-49 with a win ratio of 60%. Is this statistically significant? A two-tailed binomial test at 95% confidence on all the predictions gives me a p-value of 0.046 with a confidence interval of (.501,0.674). Just focusing on regular season play, I have a 0.030 p-value and a confidence interval of (.509,.689). So yes it is statistically significant, but barely. In other words, my predictions were better than random (50-50 chance), but barely.
Still, I have reason to hope. My trained and tested model now averages about 65% correct on past training and test data, which is in this confidence interval. I have improved the model over the course of the season, so earlier predictions come from a weaker model. Week 16 was likely just a fluke. And playoffs are much harder to predict and have fewer games. So for a first attempt, it's not great, but it's not terrible.
Going Forward
There is always next year! As this is true for NFL teams, it is true for my predictive models. What needs to improve? As always, it's better data and better models. In particular, I'd like to make the following improvements in the offseason:
- better data
- add splits, ranks of individual players, offensive and defensive ranks, playoff spot position throughout season, divisional game or not, coaching playoff success
- better models
- more testing, better feature selection, try new algorithms
We'll see what else comes along as I get ready for next season! See you then!
No comments:
Post a Comment