(For a more detailed explanation on how these were derived: see also this.)
Well, nearly every game. You’ll have to look hard to find the missing games involving two D-I teams. Just head over to your favorite team’s page and click on the score of the game.
Since these are adjusted win probabilities, games involving non D-I teams are not included. While you might be looking for raw win probability, adjusted is the only way to go. Nobody was watching the opening moments of the Kansas/Alcorn State game with any notion that the game would be competitive, even when Alcorn State somehow scored the first four points. From a strategy standpoint, this is how the participants are (or should be) looking at the game.
If you’re a regular visitor, then you’ll be able to pick up on what’s going on with a little effort. The only new concept in the chart is leverage, which is a take-off of Tom Tango’s baseball version. It measures how much is at stake on a particular possession. The cut-offs for the five categories are fairly arbitrary at this point. You can think of it as a proxy for the watchability of a game at that point.
The colors range from blue, where win probability is largely unaffected by the potential outcome of a possession, to yellow, where the outcome of a possession can have significant impact on the win probability (more precisely, at least a 10% swing between a 2-point possession and zero points). Leverage is not based on what happened during the possession, but is the range of win probability based on what could have happened.
I’m comfortable that the probabilities are well-calibrated, although there’s a bit more work to be done. The limiting factors to their accuracy are the quality of the play-by-play data available and the algorithm I use to parse possessions. Therefore, there are games where the possession count is too low, and there may be long gaps between possessions. However, the most important part of the play-by-play for these purposes is time and score, and it’s also the most reliable piece of information in the play-by-play. I am using end of season ratings for to compute the initial win probability, which obviously has some limitations.
My goal is to have an easy reference for the evolution of a game in a way that goes beyond the final score to truly characterize the competitiveness of the contest. This will surely get some more tweaks, but I wanted to share it now since you may be bored tracking the latest in NCAA investigations and I won’t be able to put much more into it for the next few weeks. Rest assured, this kind of framework for describing a game opens the door for other avenues of research.