Washington Post: Redskins Analysis for Week 6

Today I've got another post up at the Washington Post's Redskins Insider site. This time I looked at a number of different items.

• Find out where McNabb and the rest of the offense stands in terms of one of the most predictive stats--yards per attempt.
• Read about possibly the nuttiest 4th down analysis I've ever made.
• Part of the reason for Washington's success so far this season has been thanks to the deep ball.
• How big was Clay Matthews' injury midway through the game? Huge.
• The 'Skins shouldn't fear the Colts this Sunday.

Team Rankings: Success Rate

Most readers are familiar with the weekly efficiency rankings, but that's not the only way to get a sense of team success. Another way is by looking at Success Rate (SR). SR is only counting up successes and failures, so it excludes the magnitude of play results. The more random types of outcomes, such as turnovers and very long plays are counted only as a single success or failure, no different than a 6 yard gain on first down or a stop on 3rd down. It also considers a successful red zone play no differently than one at midfield.

As I've been playing around with SR, I've noticed a few things. First, it correlates well with winning. And second, it correlates well with itself, meaning it is relatively stable throughout the season. These are the two attributes we want in a stat for it to be predictive of future outcomes. I'll have a future article that goes into this in more depth.

Game Probabilities - Week 6

Weekly game probabilities are available now at the nytimes.com Fifth Down. This week I also lead-in with a short discussion of how the Giants defense has outshined that of the Jets. I also shamelessly tout the prediction model's performance from last week.

Efficiency Rankings - Week 6

Two of the top three teams lost but still kept their spots. San Diego, in particular, lost to a far weaker team, but did so thanks to special teams play. One criticism the system receives is that it doesn't include special teams.

It's not that special teams don't matter or can't affect game outcomes. The reason they're excluded from the model is that they are wildly inconsistent from week to week. So while their explanatory power can be significant, they are not predictive. And because special teams performance is not predictive, is it really a mark of good or bad team? Or is it just randomness, subject to a handful of non-repeating plays in each game, of which there are only 16 in a season?

Todd Haley: Advanced NFL Stats Coach of the Week

On the road against the heavily favored Colts, Chiefs head coach Todd Haley began the game with some unconventional calls and received some heavy criticism. He started the game with a surprise onside kick and went for it on 4th down on the Chiefs first offensive drive.

The surprise onside kick is, statistically, almost never a bad idea. At the start of the game, we can just use expected points to analyze the decision. Surprise onside kicks are recovered 60% of the time. Last fall, I wrote (forgive the self quote):

What Is the Break-Even Run Success Rate?

I've been looking at run Success Rate (SR) lately, and it appears to be a fairly important indicator of team success. But if I said that a particular team's run SR was 45%, how would you know if that's any good? It's less than 50%, so does that mean it's bad? So I decided to plot Expected Points Added (EPA) against SR to find out where the break-even point is.

When I plot run EPA per play vs. run SR (for 2000 through 2009), the break-even point is where the best-fit line crosses EPA/P axis--just above 41%.

How Accurate Is The Prediction Model?

A frequent question around here is 'What's your prediction model's accuracy?' The answer is that it varies from year to year, depending on how the schedule works out, with some random chance thrown in. In some years, there just happens to be more evenly-matched games, and in others there are more mismatches.

What's more important, at least in terms of accuracy, is what's called calibration. For example, when the model says a game is a 60/40 game, I actually want the model to be "wrong" 40% of the time. I want an accurate estimation of the game odds more than favoring the eventual winner.

Fortunately, reader 'K Rich' tracked the model's performance since 2007 and sent me a thorough spreadsheet, and the chart below illustrates the model's calibration results.

2010 Play-by-Play Data Posted

Play-by-play data from weeks 1 through 4 is now available. Happy crunching.