Wednesday, 16 January 2013

...And Then There Were Four. The NFL Championship Games.

Firstly, I'd like to reassure any readers whom prefer their brand of football to come without points, that this blog is not turning by stealth into one based around American Football. However from 2001 to 2007 I posted almost exclusively about gridiron, so I couldn't resist the temptation to follow this years playoffs for just one more post. Or possibly two.

Sunday will see the NFC Championship game decided when Atlanta host San Francisco and three and a half hours later, New England entertain Baltimore for the AFC crown in a match that will finish around 3 am Monday, UK time.

As in soccer, the simplest metrics are often also the most powerful and the ability to efficiently move the ball, either on the ground or through the air in the NFL acts as an excellent proxy for team ability, especially if the unbalanced schedule is allowed for.

How one defense matches up with the opposing offense also has a big say in the long term outcome of NFL games. A potent passing attack can become merely pedestrian in the face of an excellent passing defense, whilst a run of the mill thrower can have a career day if faced by a poor secondary. Raw ability matters in the NFL, but it is the matchups that go some way to deciding the likely result and if you compile enough pregame matchups and compare them to actual results you can produce a passable predictive model.

Nate Silver was brave enough to put his predictive credentials gained in the field of politics and to a lesser degree baseball on the line on TV this weekend by plumping for a Seattle/New England Superbowl on February 3rd. That prediction is no longer an option following Seattle's demise in Atlanta and predictably Silver's reputation as a soothsayer is seen as tarnished in some quarters because his prediction was wrong. However, all any predictive model can do is produce estimates of how likely an event is to happen. An event may be predicted to happen 60% of the time, but that also means it won't happen 40% of the time. 40% is the minority event, but it is still going to happen fairly regularly.

A couple of statistically based models made Seattle narrow favourites in Atlanta, (I made them 52% favourites), but the numbers were close enough to call the game a coin toss. So if we called heads and saw a tail, we wouldn't be too surprised. Disappointed maybe, but not surprised. That's pretty much what occurred in Atlanta, a single trial on a coin flipped failed to go that extra half revolution in the last 25 seconds. A failed prediction on the day, but only time and numerous repeats of similar predictions can validate the model from which the single prediction originated.

Silver himself tweeted a version of Billy Beane's famous quote from Moneyball, "my sh*t doesn't work in the playoffs" at the end of the regular season. The truth is that no one's sh*t works that well in small sample sizes, which is what the playoffs are. Ten games and a (usually) neutral venue, inter conference showdown is all there is and random luck and a sudden death format is going to have a huge influence on individual games and the final destination of the Vince Lombardi Trophy.

I've looked back at every playoff prediction made by my corrected yardage based model from 2001 to 2007, excluding the actual Superbowl games which were all played on a neutral field. Some of the individual results make the model appear absurd. For example Pittsburgh to beat New England in the AFC Championship games of 2001 and 2004, both with Steeler chances of around 70% and the Patriots won both games. Baltimore (66%) to beat Tennessee in the 2003 AFC wildcard game, the Titans won and Dallas (74%) to see off the Giants in the 2007 NFC Divisional game, Manning junior won.

Individual games aren't a good advert for predictive models because they exaggerate their effectiveness if they do well over a run of half a dozen games or they appear useless if they continually pick favourites in the short term which then go on to lose. However, the same model that appeared to overrate the Steelers and discount the Pats also predicted 43 home wins from the 70 playoff games from 2001-2007 and the actual number was 44. The more times you run a predictive model, the more information you gather as to the effectiveness of that model. Predicting the result of an election is much easier than predicting the result of one NFL game, or even ten.

With that thinly disguised excuse over, let's move on to each game.

NFC Championship Game.

Team. Offensive Run Efficiency. Offensive Pass Efficiency. Defensive Run Efficiency. Defensive Pass Efficiency.
San Francisco @ 1.15 1.15 0.86 0.86
Atlanta. 0.86 1.05 1.12 1.04
Win % for Atlanta 33% Point Spread. +5

We have the same NFC west superiority on show here as last week. Offensive numbers in excess of 1.0 indicate that the team is superior to an average offense once the strength of opponents has been accounted for. The 49ers have faced some fairly poor run defenses, which have allowed 4.4 yards per carry, but the west coast side's 5.1 yards per carry average indicates that they are still well above average. They show a similar level of quality in their passing game.

The matchup with Atlanta's defense doesn't improve matters for the hosts. On this side of the ball numbers above 1.0 indicate that a defense is below average. Atlanta have allowed teams which averaged 4.3 yards per carry in their combined 250+ games to rush for 4.8 yards per carry when faced with the Falcon's defense. The Falcon's passing defense, while an improvement is also the wrong side of par.

...and a 67% Chance of a Superbowl Appearance in 2012.

After combining the paired statistics, the numbers make ugly reading for the hosts. They look likely to struggle moving the ball on offense, especially on the ground and face a San Francisco side which should be able to move the ball easily, particularly on the ground. Similar matchups from the past indicate that the visitors should win such a game about 67 times out of 100, by an average of 5 points. A bag, two red balls and one slightly off coloured red ball should suffice for a crude match simulation.

AFC Championship Game.

Team. Offensive Run Efficiency. Offensive Pass Efficiency. Defensive Run Efficiency. Defensive Pass Efficiency.
Baltimore @ 1.04 1.00 0.96 0.97
New England. 0.97 1.17 0.93 1.16
Win % for the Pats. 60% Point Spread. -3.5

Nothing much has changed about the Pats since they upset the Rams in 2001. They are rarely above average at running the ball, care very little about defensive vulnerability, even through the air and dare you to keep up with Tom Brady. Just as the Falcons of recent times have out played their stats by maintaining a, usually regression prone, high third down conversion rate (2nd best, 6th and 3rd since 2010), the Pats are also better than their raw stats. Outstandingly talented quarterbacks are often the reason, Manning performed the same trick for years at Indianapolis by continually taking a poor defense and limited running game deep into the post season, cementing each win by continuing to pass the football. It is no surprise to see New England favoured by just over a field goal, when experience tells you the margin should probably be at least twice that.

At quarterback for Baltimore, Flacco has spent his career trying to make the step up to elite, but remains in limbo just off top class. He's been average this season. Defensively Baltimore are above average, but someway off the frighteningly good unit of recent times, age is finally taking a toll.

Matchups indicate that both sides will be above average through the air, but struggle on the ground. Home field accounts for much of the advantage predicted by the model for the Patriots, but Belichick puts the ball in Brady's hands often and opponents find themselves having to deal with a greater than usual aerial assault. Baltimore will probably have to throw more and run less than they are accustomed just to keep up and the Patriots half of Silver's Superbowl prediction would appear to be a reasonable expectation come Monday morning in the UK.


  1. Does your model take into account turnovers?

  2. Hi Flash,
    It doesn't. They certainly have a major impact on match result, but they don't appear to be a very repeatable skill, especially recovering fumbles. Your more likely to see patterns which don't exist than skills that do persist over a season imo. I happy with how the model performed live from 2003-2007 compared to fairly reliable markers such as the Vegas lines. I don't mind adding other variables, but you often end up chasing noise.

  3. make that "you're" before the grammar police turn up :-)

  4. I agree. TOs are extremely important, considering each team only has around 10-14 possessions a match, but they are very difficult to predict.