We’re now 4.0625 weeks into the NFL’s 2015 regular season. (If you don’t know what the NFL is, you should probably stop reading now.) That’s about one-quarter of the whole 256-game shebang, enough to start taking stock of preseason predictions. So I got to wondering: Which teams have been the biggest surprises so far?
To get one answer to this question, I downloaded game results from Pro-Football-Reference.com (here) and compared them to the central tendencies of my preseason predictive simulations (here). The mean error of the predictions for each team so far is plotted below. The error in this case is the difference between the number of points by which the team was expected to win or lose each game and the number of points by which it actually won or lost. For example, my simulations had the Colts, on average, winning this week’s Thursday-night game against the Texans by 4, but they actually won by 7. That’s an error of +3 for the Colts and -3 for Houston. The mean error is the average of those errors across all games played so far. So, a positive mean error (blue dots) means the team is over-performing relative to the preseason predictions, while a negative mean error (red dots) means it’s under-performing.
Most of those results won’t surprise regular NFL watchers. The New York Football Jets finished 4–12 last year and ranked near the bottom in my preseason wiki survey, but they’re off to a 3–1 start this year. The Falcons, who went 6–10 in 2014 and garnered a low-middle score in the wiki survey, are undefeated after four weeks. At the other end of the scale, the Dolphins got a high-middle score in the preseason survey, but they have stumbled to a 1–3 start.
It’s also interesting (to me, anyway) to note how the team-specific errors are only loosely correlated with differences between predicted and observed records. For example, the Giants are only 2–2 so far this year, but they show up as one of the biggest over-performers of the first four weeks. That’s partly because both of those two losses were close games that could easily have flipped the other way. The Giants were expected to be on the bad side of mediocre, but they’ve been competitive in every game so far. Ditto for the Ravens, who only show up as mild under-performers but have a 1–3 record (sob). At least three of those four games were expected to be close, and all of them turned on late scores; unfortunately, only one of those four late turns broke in Baltimore’s favor.
This exercise is only interesting if the preseason predictions on which we’re basing the calls about over– or under-performance are sound. So far, they look pretty solid. After four weeks, the root mean squared error for the predicted net scores is 12.8, and the mean squared error is 165. Those look large, but I think they’re in line with other preseason score forecasts. If we convert the predicted net scores to binary predicted outcomes, the model is 40–23 after four weeks, or 41–23 if we include last night’s Colts-Texans game. That’s not exactly clairvoyant, but it beats eight of ESPN’s 13 experts and matches one more, and they make their predictions each week with updated information.