By the time most of you will be reading this, I’ll be on the road from Austin to Arlington for the Dallas Cowboys Wild Card Playoff Game against the Seattle Seahawks. This will be the fourth time the Cowboys have hosted a playoff game at AT&T Stadium since it opened in 2009, and this will be my third playoff game there (I was in New York after the 2014 season when the Cowboys hosted the Lions and won so I ended up selling my tickets). Definitely looking forward to the game although my win projections don’t really seem to agree with me too much there. But we’ll get there in a sec. First a recap…
Week 17 Recap
Week 17 ended up being my best week for predictions since I started publishing and making predictions in Week 13. I correctly predicted twelve of the sixteen games or 75%. Of course, I go by more than simply correcting who win’s a loses by scoring how confident I was in those predictions using the Brier Score. Even in that, my Brier Scores were the lowest of the season.
Last week, if you remember I had four different models plus an average of two of those models for my predictions. Here was a summary of those four models plus the average model:
- Original Model – Uses a Poisson Distribution to predict based on a teams average points scored at home and away
- Elo Model – Use code from FiveThirtyEight‘s Elo model with a slight tweak in the Home Field Advantage (HFA) metric that I use based on the actual HFA of the difference between the average points scored at home for all teams from the average points scored in away games for all teams.
- Original + FO + ELO – This model combines Model 1, Model 2, and adds four additional stats from Football Outsiders (Home and Away) for all NFL Teams
- Bayesian – Uses PyMC3 and Bayesian methods in simulating 1000 games for each match and calculates a probability that the home team will win based on those simulations
- Average of Models 3 and 4
Here were the Brier Scores in Week 17 for each of the models:
| Model | Brier Score |
|---|---|
| 1 | 0.1381 |
| 2 | 0.1193 |
| 3 | 0.1173 |
| 4 | 0.1327 |
| 5 | 0.1233 |
As it turned out the third model had the best score with the Elo Model a close second (remember that closer to zero is better).
Wild Card Win Predictions
For this week, I didn’t add any new models. I did once again tweak the HFA of the Elo model down to 55.8 (remember that FiveThirtyEight uses 65). I also decided to take an average of all four models, and those are the final and official numbers I’ll use as how I things will fare this weekend.
You’ll also notice I added a column for Cynthia Frelund‘s picks. She is NFL media’s first analytics expert and co-hosts a podcast called Game Theory and Money. She apparently likes to round her picks to whole numbers which is fine; on television I imagine it doesn’t matter how precise you are.
| # | Game | Model 1 (Original) | ELO Model 2 (HFA=55.8) | Model 3 (FO+ELO) | Model 4 (Bayesian) | Avg of All Models | Vegas | Cynthia Frelund Picks | Poisson Predicted Final Score |
Expected Tot. Points Scored |
Vegas Over/Under |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | IND | 28.86% | 45.66% | 53.22% | 30.65% | 39.60% | 48.08% | 54.00% | 25 | ||
| HOU | 66.14% | 54.34% | 41.46% | 69.35% | 57.82% | 54.35% | 46.00% | 26 | 51 | 48.5 | |
| 2 | SEA | 53.20% | 46.91% | 69.08% | 40.16% | 52.34% | 45.45% | 45.00% | 24 | ||
| DAL | 40.78% | 53.09% | 25.81% | 59.84% | 44.88% | 57.11% | 55.00% | 19 | 43 | 43 | |
| 3 | LAC | 27.21% | 41.55% | 35.44% | 50.28% | 38.62% | 42.37% | 42.00% | 20 | ||
| BAL | 67.31% | 58.45% | 58.51% | 49.72% | 58.50% | 60.17% | 58.00% | 21 | 41 | 41.5 | |
| 4 | PHI | 14.36% | 39.77% | 21.81% | 14.65% | 22.65% | 30.67% | 40.00% | 19 | ||
| CHI | 81.81% | 60.23% | 73.34% | 85.35% | 75.18% | 71.79% | 60.00% | 22 | 41 | 41.5 |
So what insights do we gain here?
For the first game on Saturday afternoon, three of my four models give Houston the advantage. Only Model 3 seems to think that Indianapolis has enough momentum to win in Houston. Cynthia also seems to favor the Colts over the Texans, but my (fake) money will be on Houston winning that one.
For the second Saturday game, the game I’ll be attending, there is a split among my four models with the average model giving the edge to Seattle. While I’m committed to that outcome on paper, in reality I hope I’m wrong.
For the early Sunday game, all but one of the models is picking the home team, the Baltimore Ravens to win against the Chargers. And the one model that has the Chargers winning is effectively a coin flip at 50.28%.
And finally the late Sunday game, is the only game where all the models confidently pick Chicago to win.
Granted, as a “big picture” guy, I would love it if Dallas won Saturday night and Philadelphia won Sunday afternoon. Then, in the Divisional Playoffs, Dallas would beat the Rams and Philly would beat the Saints so that Dallas could host the Conference Championship against the Eagles in two weeks!
Well, a guy could dream!
Enjoy the games this weekend. On Monday morning I’ll publish my recap of the College Bowl Games and give my predictions on who will win the National Championship Monday night!