GridIronMine.com's Victory Forecast System
The victory forecast is the probability that a team wins from any point in a game. GridIronMine.com allows you to view victory forecasts in either graphical or tabular form. Below, we describe how victory forecasts are created.
We have collected every single play run in the NFL over the last 5 years (2002-2006), which totals over 150,000 plays. These plays are stored in our database.
In order to generate victory forecasts, we feed these plays into a logistic regression. A regression is simply the process of fitting a set of data points to a line or curve. A logistic regression is a special regression that allows us to predict a binary (i.e., win or loss) outcome from a set of variables. For victory forecast in football, those variables include things such as the current score of the game, the time remaining, and who has the ball. So the logistic regression provides a function, that, given these variables, produces the probability of winning the game given the situation.
The output of the logistic regression is smooth. That is, consider the following situations:
Team A is beating Team B by 10 points, with 11 minutes to go in the 3rd quarter. Team A has the ball at their own 20 yard line, 1st and 10.
Same situation, except there are 8 minutes to go in the 3rd quarter.
Same situation, except there are 5 minutes to go in the 3rd quarter.
Our victory forecast returns nearly identical probabilities---a little over 87% in all three cases---which is because they are quite similar situations.
However, if one looks at the actual NFL data over the last 5 years, in the first situation, teams won the game approximately 84% of the time. In the second, ths number was 65%, and in the third, the number was 76%. The reason for this is that these situations do not happen often in games. Specifically, one team had a 10 point lead and the ball near their own 20, with between 12 and 4 minutes to go in the 3rd quarter, only 66 times. This is not enough data points to get a good idea of the actual victory forecast. For example, in the range between 12 and 9 minutes, there were only 25 occurrences. In 21 of them Team A won, for a probability of 84%, which is near our prediction of victory forecast (87%). But just a small change to this---say that Team A just happened to lose two additional times---the probability changes to 76%, which is not close and seems to suggest that our model is inaccurate.
However, this is not the case. The regression has the benefit of having hundreds of thousands of data points, so the victory forecasts do not wildly flucuate because of small variances in small data sets. Note also that the 87% number corresponds a lot more with our intutition about the likelihood that a team with a 10 point lead in the 3rd quarter actually wins the game.
Victory Forecasts shown on the site are team-independent. That means that if New England is beating Houston at the half, 10-3, their Victory Forecast is the same as if Cleveland is beating San Diego 10-3. Clearly, New England has a better chance to win their game than Cleveland does theirs. Subscribers to our site can get team-dependent Victory Forecasts.
We also provide access to the WhetherStation so that you, the user, can investigate different scenarios.