Title: iWinrNFL: A Simple and Well-Calibrated In-Game NFL Win Probability Model

Abstract: During the last sports season a lot of discussion has been generated for the several, high-profile, "comebacks" that were observed in almost all sports. The Cavaliers won the championship after being down 3-1 in the NBA final's series, which was exactly the case for Chicago Cubs and the World Series. The Patriots won the Super Bowl even though they were trailing by 25 points late in the third quarter, while more recently FC Barcelona in the top-16 round of Champions League scored 3 goals during the last 7 minutes of the game (including stoppage time) against PSG to advance in the tournament. This has brought the robustness and accuracy of the various probabilistic prediction models under high scrutiny. In this paper, we discuss some of the issues that have been brought up in the wake of last year's events and we also develop our own in-game win probability model for NFL (iWinrNFL). In particular, we build a simple logistic regression model that utilizes a set of 10 variables to predict the running win probability of the home team. In order to account for non-linearities that exist at the final minutes of the game we enhance the logistic regression model with a radial basis Support Vector Machine model for the last 3 minutes of the regulation. We train our model using detailed play-by-play data from the last 8 NFL seasons obtained through the league's API. Our results indicate that in 74% of the cases iWinrNFL provides an accurate winner projection, while most importantly the probabilities are well-calibrated.