The Goto Chart#

All The Available Data#

Overall, I feel the most useful dashboard chart is this one:

This includes all games from 1996 (which is about ~36,000 games). If your team is down, this gives you a pretty good sense of how steep a mountain your team has to climb (and of course, on the flip side, how safe your lead is). Plus, I like being able to see the current record for a given time, knowing that if you are beyond the record it would be, well, a record if you came back. (And just noting that the Minnesota Timberwolves hold or tie the record for allowing the biggest comebacks at the 3, 6, 7, and 10 minute marks).

You can click on data points and see which game(s) hold the record. Going full screen makes it easier, and on mobile, you have to go full screen first. Once you go full screen, then you can click on any point to get the summary for that minute.

Please note, the ‘Record’ is calculated by getting the point margin exactly at that minute mark, so if a team was down 25 at the start of the 4th and then went down 26 and then back to 25 by the 11th minute, the 26 is not recorded here.

Of course, this is all games from 1996 with no other conditions (e.g., recent years, home vs. away, etc.). And this chart will change if you change the conditions.

How This Compares Versus ESPN’s Live Win Probability#

A useful way to judge the utility of this chart is to see how it compares to ESPN’s running win probabilities for a given game. Even though:

Despite these markedly different approaches, both models mostly tell a similar story about a team’s overall comeback chances. While requiring further analysis, this suggests the most correlated variable to comeback chances is points down versus time.

Timberwolves @ Bucks on 04/09/2025#

So let’s look at a recent run-of-the-mill Timberwolves game against the Bucks on 04/09/2025:

Here, we are again plotting on a normal probability plot instead of a linear y-axis so we can better examine the extreme probabilities. The dashboard probabilities are taken from the same ones shown in the goto chart at the top of this page. In fact, for any dashboard point, you can click on it and it will bring you to the interactive dashboard page and show the exact regression fit line used to calculate the probability for that point. (And if you click on the 10 minutes remaining point, you will see that the Timberwolves hold the record for losing a game when up 24 points with 10 minutes to go.)

But results vary from game to game. Let’s just look at a few more games I watched recently:

Some features are notable:

  • Clearly, ESPN’s BPI index is more heavily discounting certain teams independent of record.

  • The differences are larger until about the 4th quarter, where they tend to converge.

Adding Conditionals: Home Versus Away For The Modern Era#

Outside of team/player strength (which can have a very large effect), the conditional providing the greatest discrimination is probably adding whether the team coming back is at home or away. Then, limiting the seasons to the modern era. Adding those two conditions gives you these two plots, which will give you a more accurate probability:

Comparing To The ESPN Model#

Let’s re-compare to the ESPN model taking these conditionals into account:

Making the dashboard model account for home court advantage and increased chances of coming back in the modern era leads the fit to be a little different (worse for some cases and a little better in others).

Understanding how difficult it is to compare probability models, there still are some data points that stand out. For example, for the April 20 GSW @ HOU game, at 18 minutes remaining with Golden State having a 21-point lead, ESPN has them with a win probability of 98.4%.

Based just on all games from 1996 until now, the odds are about 97.8% – however, when we account for the fact that Golden State was the away team and limit our data to the modern era (where comebacks are slightly more likely), the dashboard model drops the win probability to 95.9%. Seeing how Houston was the 2nd seed in the West and Golden State got in through the play-in, I am curious what data ESPN has that pushes them above the historical average by a point. (Most likely, based on discussions of how prediction models work, the model is giving the seasoned veterans Steph Curry and Jimmy Butler some extra juice.)

But Sometimes You Still Need a Supplemental Chart#

Even though the first chart gets you most of the way there, sometimes a chart like this limited to our recent history is also useful:

Just to get a sense of what we’re capable of!