What Was More Improbable: The Knicks’ Game 4 Comeback or the Patriots’ Super Bowl LI Rally from a 28-3 Deficit?

Using ESPN win probability models, the answer is surprisingly close, and probably closer than the exact decimals suggest.

The New York Knicks trailed the San Antonio Spurs by 29 points in the third quarter of Game 4 of the NBA Finals and somehow won 107-106, completing the largest comeback in NBA Finals history. The comeback felt impossible in real time, and the win probability numbers help explain why.

According to ESPN Analytics, San Antonio’s win probability reached 99.6% with 9:33 remaining in the fourth quarter, after the Spurs had pushed the lead back to 20. Put another way, the Knicks were sitting at 0.4%, which equates to about a 1-in-250 chance.

This is where win probability becomes useful. “They were down by a lot” is true, but it leaves out the shape of the game. A 29-point deficit matters, obviously, but the situation around that deficit matters too. A 29-point deficit early in the game is not the same as a 29-point deficit in the third quarter. A 20-point deficit with 9:33 left in an NBA Finals game is something else entirely. The scoreboard tells you what happened; win probability tries to estimate how often teams in similar situations actually come back.

In basketball, that usually means looking at variables like score differential, time remaining, possession, timeouts, fouls, pregame team strength and other contextual factors. Different models use different inputs and structures. Some are built with classical regression methods. Others use tree-based models or ensemble methods that can better capture nonlinear relationships. The basic idea is this: given the current state of the game, what are the possible paths from here, and how often does each team win from situations like this?

A version of that question is what helped launch modern probability theory. In the 1600s, Blaise Pascal and Pierre de Fermat worked on what became known as the “problem of points,” a puzzle about how to fairly divide the stakes of a gambling game that had been interrupted before it ended. Their key insight was that the fair answer did not depend only on the current score. It depended on the possible futures from that point forward.

That is not exactly the same thing as an NBA win probability model. Nobody in 1654 was trying to price the odds of OG Anunoby tipping in a missed three with 1.2 seconds left. But the underlying idea is familiar: the present state matters because it tells us something about the distribution of future outcomes.

Football models work similarly, but with football-specific game states. Score, time remaining, field position, down, distance, possession, timeouts and game context all matter. That is why the Patriots’ comeback from 28-3 against the Falcons in Super Bowl LI is such a useful comparison.

In that game, ESPN calculated that Atlanta’s win probability peaked at 99.8%. New England’s chance fell to 0.2%, or about 1-in-500. Other models have landed in slightly different places, which is a useful reminder that no win probability model is the ground truth. The model design, feature set and historical data all matter.

So, which comeback was more improbable?

If we use ESPN as the common source, the Patriots’ comeback was slightly more improbable at its lowest point. But this is where we should be careful with false precision. A 0.2% model estimate and a 0.4% model estimate are not the same number, but they are close enough that I would not treat the difference as definitive in the way we might treat a measured final score.

These are model outputs. They come with assumptions, uncertainty and modeling choices. The exact decimal is useful, but only up to a point.

The Patriots were in the 1-in-500 range. The Knicks were in the 1-in-250 range. Those are not identical estimates, but they are clearly in the same statistical neighborhood: outcomes that are still possible in the model, even after fans leave the arena, broadcasters start wrapping the story and everyone begins mentally filing the game away as over.

The Knicks’ comeback was absurd. The comparison just gives us a way to measure the absurdity.

That distinction matters. A 0.4% chance does not mean the Knicks had no chance. It means that in a large enough sample of similar game states, a team in that position still wins once in a while. Most of the time, they lose. Every so often, the unlikely path is the one the game actually takes.

This is also where we have to be careful with the model. Win probability models are descriptive tools, not truth machines. They do not know that Anunoby is about to crash the glass for a tip-in with 1.2 seconds left. They do not know that Jalen Brunson is going to keep pressing. They do not know exactly when a young Spurs team is going to tighten up, or when a crowd at Madison Square Garden is going to shift from anxious to fully believing.

The model can describe the situation going into those last nine minutes. The players still decide what happens inside them.

That is also why these numbers are useful beyond sports. Good probabilistic thinking keeps two ideas in view at the same time: unlikely things can still happen, and the thing that happened was not always destined to happen.

After the fact, the game is easier to narrate than it was to live. The Knicks had momentum, the Spurs looked young, the Garden got louder with every possession and Anunoby made the final play. Brunson kept New York close enough for it to matter. All of that is true. It is also a much easier story to tell once the comeback has already happened.

At 99.6%, nobody had that clean version of the story yet.

That is the value of the model. It brings us back to the uncertainty that existed before the ending became obvious. The outcome we saw was only one possible branch, and by the numbers, one of the least likely.

By ESPN’s numbers, the Patriots’ 28-3 comeback remains the more statistically improbable one, barely.

The better conclusion is that these were essentially equal-sized comebacks. Different sports, different clocks, different models, same basic lesson: when the probability gets that close to zero, the exact decimal matters less than the reminder that close to zero is not zero.

One in 250 is still possible. Watching it happen is what makes it feel impossible.

Mike Band NFL Next Gen Stats Research & Analytics Lander Analytics Contributor

---

Subscribe to our Substack and below to our monthly emails for practical AI strategies for your organization: what to build, what to avoid, and how to make systems reliable in the real world.

Work with us: If you want help identifying the right first workflow, building a permissioned knowledge base, or training your team to ship responsibly, reach out at info@landeranalytics.com.

About the author: Mike Band is the Sr. Manager of Research & Analytics at NFL Next Gen Stats and AI Researcher at Lander Analytics.