Sports AI
Titans are favorites to win Super Bowl after model adjustments
Check previous articles for more model details and context: Pre-Wild Card Round, Pre-Divisional Round
The conference championships are quickly approaching after another entertaining weekend of football. While the Chiefs went on a historic 51–7 run to overcome a 24 point first quarter deficit against the Texans, perhaps even more surprising was a victory for the Titans against the heavily favored Ravens in Baltimore. That’s 2 weeks in a row that the Titans went on the road to beat teams that were predicted by V 4.0 (and other models) to have a sizable advantage. These unexpected Titans’ wins provoke the question: has V 4.0 been systematically underestimating Derrick Henry, Ryan Tannehill and the rest of the Tennessee crew? More on that topic in the Revising the Model section below, but first let’s check the original model predictions in comparison to ESPN’s FPI and FiveThirtyEights’s Elo & QB-adjusted Elo.
As measured by sum of winning chances of remaining teams, Model V 4.0 continues to rank high among other predictions. In the NFC, V 4.0 was more confident about the top 2 seeds than other models, which turned out to be an accurate prediction as the 49ers and the Packers will be playing for the conference title at Levi’s Stadium on Sunday. On a game by game basis though, model V4.0 has a disappointing record after 2 weeks of playoff action: 3 of 8. Albeit a small sample size, it’s a far cry from it’s 68.5% performance in the latter portion of the regular season. Additionally, the Titans were ranked dead last by V 4.0 in terms of their chances to win the AFC and eventually win the Super Bowl. After watching the games, it seemed obvious to me that more than just luck propelled Mike Vrabel’s squad over the Patriots and the Ravens. Thus, with the advice of some colleagues [1,2], I decided to revisit certain aspects of the model and generate a new set of predictions, which now suggest the Titans are the new favorites to win the Lombardi Trophy. The original V 4.0 predictions and revised (aka V 4.0TA) predictions for remaining games are immediately below, and you can read on for details on how the revised model differs from the original.
Explaining the Unexpected
I believe there are 3 main possibilities that can explain why V 4.0 was way off on the Titans:
- The Titans’ season-long aggregate statistics are misleading
- The Titans are better than average road team (and thus applying league average home field advantage to their games is misleading)
- The Titans have been luckier than other teams
I think it’s mostly 1 & 2 and maybe a little bit of 3.
On point 1: Those who have been following the Titans from the beginning of the season know that Mike Vrabel made a pivotal decision in week 7 by naming Ryan Tannehill as their starter. When Marcus Mariota was starting, the Titans were 2-4, but since Tannehill has taken over, the Titans are 9-3 (including the postseason). Since V 4.0 uses statistics aggregated across the entire season, those first six weeks are likely skewing the numbers and not representative of how they have been playing with Tannehill at QB. On point 2: As of late the Titans have been very good on the road. They are 5-1 in their last 6 games outside of Nashville’s Nissan Stadium, and V 4.0 only predicted they would have won a single game. On point 3: Sure the Titans have had some lucky breaks with higher than normal dropped passes from the Ravens and the Patriots and an injury to the Ravens’ starting running back Mark Ingram II, but they’ve also been playing excellent short-yardage defense and been consistently dominant rushing the ball.
Revising the Model
Based on points 1 & 2 above, I made two adjustments to the modeling approach and reran the simulations. First, I only used offensive statistics for the Titans after week 7 rather than season-long data. Second, I reduced the home field advantage of teams opposing the Titans. With this new model (V 4.0-Tannehill-Away-Adjusted aka V 4.0 TA), I reran simulations from the start of the playoffs (see table below), and for where things stand at the moment (see table above). It’s entirely possible that these adjustments are overcompensating based on only a couple data points (“overfitting” in data science parlance), and we don’t have a great way of testing these changes prior to Sunday’s matchup. (One related note is that these adjustments would only improve the “test set” performance. While this in itself does not rule out overfitting, it would be alarming if adjustments led to worse test set performance.)
The adjustments cause the Titans to jump from sixth to third in the AFC in terms of chances to play in and win the Super Bowl. From this perspective, their recent wins seem less surprising.
Who’s going to Miami?
While all five models favor San Francisco over Green Bay, only V 4.0TA thinks Tennessee is favored over Kansas City. Most seem to think the Titans were dealt a couple great hands and are in need of one more in order to win in Kansas City, but V 4.0 TA suggests they are outplaying their opponents on the flop, turn, and river and might do so once more. It won’t be easy as the Chiefs, led by sensational Patrick Mahomes, won’t go down quietly. Will the Titans hot steak continue? Will the 49ers successfully defend their turf? We’ll soon find out.
[1] Christopher Smith, PhD, Scientist and devout Titans fan
[2] Ali Bhanpuri, NFL expert who has been writing about quarterbacks all season
