NFL simulations are far from an exact science. They attempt to mathematically project the future based on history and past performance, but they can’t account for everything. A stiff breeze, a tipped ball, a freak injury, a rolling fog bank, an ol’ coach’s return, or simply a change in player attitude can alter results in a large way. Instead, simulations give us a blurry view of a series of possibilities among an infinite number of potential realities. But they’re fun. If you believe in parallel or multiple universes, then one of these simulated results could possibly occur.
This week I aim to improve my simulation’s 1-1 record. After failing miserably to predict an Eagles’ win over the Kansas City Chiefs, the simulation successfully projected a Denver Broncos’ victory last week (though really, big woop). But the way in which they did it was spectacular. Peyton Manning and the Broncos managed to operate their offense at an incredible 88.7% rate of efficiency, almost ten percentage points above their season average, exceeding the maximum boundary of the simulation (87.3%). The Eagles, on the other hand, ran their offense at 75.4% efficiency, eight percentage points above their season average. At that rate of offensive efficiency, with no turnovers, playing away, the simulation predicted 19 points for the Eagles, one point shy of their actual.
This Sunday, the Eagles travel to East Rutherford to play the Giants. The simulation is based on each team’s offensive efficiency through the season (Giants = 62.5%, Eagles = 68.5%). After 10,000 runs, the Eagles win 63% of the time by an average score of three points (actually, in their 6,252 simulated wins, they win by an average score of 25-15; in their 3,748 simulated losses, they lose by an average score of 24-16). The shape of the scatter plot is interesting, much different than for the Broncos’ and Chiefs’ games. The trend seems to be for a lower scoring game, or more accurately, each team tends not to score a lot of points at the same time in a single game (which jives with Brent Cohen’s dislike for the over/under).
According to the model, The Eagles win the majority of the time thanks largely in part to the Giants’ poor offense. The Eagles have room for error. Even with one more giveaway than takeaway, the Eagles win 70% of the time by an average of two points. This is the first time I’ve seen an away team with an advantage in this scenario. The threshold for the Eagles seems to be the -2 turnover differential. If they do that, the Eagles lose 55% of the time by an average of two pints.
Ignoring the simulation for a moment, I think the Eagles enter the game with a HUGE "chip" on their shoulder and rout the G-men. If the Washington Redskins game is any indication (and I hope it is), Chip Kelly puts a lot of stick in division games. So let's see how it plays out!
*Simulation Details (because someone will ask)
The simulation is based on my home field advantage (HFA) research, which shows how there have been small but distinct and different offensive efficiency behaviors between home teams and away teams in the NFL. And not surprisingly, turnovers play a large role in equalizing the playing field. Offensive performances throughout the season were entered into a logistic regression formula born from the HFA research, and randomized according to standard error values and turnover differential.
Step 1: Calculate Offensive Efficiency (OE). I used Chip Kelly’s definition for this:
(Rushes + Completions) / (Total Off Plays + Offensive Penalties)
If you check out the HFA research, there’s a really strong correlation between offensive efficiency and team success.
Step 2: Calculate Win Probabilities using the logistic regression formula that correlated OE to team success. Here’s the formula:
Win Probability = 1 / (1 + e^-((A*OE+error value) + (B*Turnover Diff + error value) + C)), where A, B, and C are constants.
Step 3: Convert the results from Step 2 into points using a linear formula:
Points = A*(Win Probability for Eagles) + B*(Win probability for Chiefs) + C, where A, B, and C are constants.