2026 NBA Finals · Game 4 · Madison Square Garden

The Night the Garden Got Loud Again

A data story of Game 4: the largest comeback in Finals history, told through every shot.

In short. Down 29 in Game 4 of the 2026 NBA Finals, the Knicks won 107-106. This paper uses every shot from the play-by-play to show how. The Spurs' shooting fell from 60% before halftime to 21% after, on shots no harder than before. A simulation shows the Knicks win about 1 time in 100 if the Spurs shoot at a normal rate. At the low point, down 20 with 9:33 left, the Knicks' chance of winning was 0.4%.

Abstract

In Game 4 of the 2026 NBA Finals, the New York Knicks trailed the San Antonio Spurs 81-52 and won 107-106, the largest comeback in Finals history. This paper uses reported scores, win probability and shooting splits to rebuild the comeback and test how rare the result was. With 9:33 left, models gave New York a 0.4% chance. A Fisher's exact test shows the Spurs' drop in shooting from the first half to the second was significant (p ≈ 0.0002 for all shots, 0.018 for threes). A shot-level logistic regression on all 86 Spurs attempts finds their odds of making a shot fell by about 84% after halftime, after controlling for shot location (odds ratio 0.16, p = 0.0003). The Knicks showed no significant change. The comeback came from regression to the mean plus a Knicks defense holding San Antonio to 30 second-half points.

Introduction

In Game 4 of the 2026 NBA Finals, the New York Knicks came back from 29 points down to beat the San Antonio Spurs at Madison Square Garden. The win is the largest comeback in Finals history (ESPN).

New York had won Games 1 and 2, then lost Game 3 at home after a poor second half. A loss in Game 4 would have tied the series 2-2. An OG Anunoby putback won the game instead and gave New York a 3-1 lead. The Knicks closed the series in Game 5 on June 13, 94-90, for their first title since 1973 (Revolt, Black America Web).

A note on perspective. I have been a Knicks fan my whole life. Until June, I had never seen them win a title. The 1999 Finals loss to these same Spurs came before my time. The Pacers ending our seasons in 2024 and 2025 did not. I watched this game the way many Knicks fans did, waiting for the loss. This paper has two goals. The first is to measure the comeback as carefully as the data allows. The second is to be honest about how the comeback felt. The analysis is in plain type. My own commentary is in italics, like this.

This paper asks two questions:

  1. How did the comeback happen? Which stretches, players and lineups closed a 29-point gap?
  2. How rare was the result? Measured against win-probability models and historical base rates, how often does a team come back from 29 points down, especially in the Finals?

Data and methods

The analysis uses public game data to answer three questions: how the margin changed over time, who changed the margin, and how often teams in this spot win.

Data

Definitions

Approach

  1. Rebuild the score after every scoring play and split the comeback into phases.
  2. Compare shooting by half with Fisher's exact and binomial tests.
  3. Fit a shot-level logistic regression controlling for shot zone.
  4. Simulate the second half at normal shooting rates (Monte Carlo).
  5. Compute a win-probability curve with a Brownian-motion model.
  6. Use plus-minus to see who was on the floor for the swing.

Anatomy of the comeback

From 81-52 down with 9:40 left in the third quarter, the Knicks outscored San Antonio 55-25 over the final 21.5 minutes and won 107-106. By quarter, the Knicks' margin went −19, −8, +12 and +16.

Interactive · all 110 scoring plays

The Knicks fell behind by 29 twice and won by 1

Drag across the chart or use the slider to step through every scoring play. Tap a moment to jump there. Score rebuilt from the Basketball Reference play-by-play.

The game breaks into four phases.

  1. The collapse (Q1 to 9:40 Q3). San Antonio led by 19 after one quarter and 76-49 at the half. Both are the largest leads by a road team in Finals history. The Spurs made 11 of their first 16 threes and 14 in the half, a Finals record. Karl-Anthony Towns picked up two fouls in the first 62 seconds. The deficit first reached 29 at 71-42 with 3:06 left in the second quarter, then reached 29 again at 81-52 early in the third.
  2. The third-quarter stand. New York held San Antonio to 14 points on 4-of-20 shooting in the quarter. A 13-0 run cut the gap to 90-75 by the end of the third. Stephon Castle's fourth foul at 10:09 sent him to the bench for almost five minutes. The deficit fell from 27 to 19 while he sat.
  3. The stall. The Knicks went scoreless for the first 2:27 of the fourth. A Victor Wembanyama putback at 9:33 made the score 95-75, and models put the Spurs' win probability at 99.6%.
  4. The finish. New York closed on a 32-11 run over the final 9:33. All 32 of the Knicks' fourth-quarter points came in this window. San Antonio shot 4 of 19 in the quarter. With the Spurs up 106-105, De'Aaron Fox grabbed a loose ball and broke away with about 11 seconds left. Fox went for a layup instead of running the clock. OG Anunoby chased him down from behind and blocked the shot (CBS Sports, Athlon Sports). On the Knicks' final possession, Jalen Brunson's deep three hit the front rim, and Anunoby tipped the ball in with 1.2 seconds left. The block and the tip came 9.9 seconds apart.

At halftime, down 76-49, the Garden went quiet (NBC Sports). A 27-point deficit in a Finals game feels final. Then the third quarter started and the Spurs went 4 of 20. Somewhere in the 13-0 run, I started to wonder if a comeback was happening. I did not trust the feeling. The Wembanyama putback at 9:33 felt like the answer, and the answer was no. Then the Knicks went on a 32-11 run.

Who drove the comeback

Two scorers and one defensive change account for most of the swing. Brunson and Anunoby combined for 69 of New York's 107 points, while the Spurs' shooting dropped after halftime.

PlayerTeamPointsShootingOther
Jalen BrunsonNYK3612 of 25 FG5 reb, 7 ast
OG AnunobyNYK3310 of 15 FG (7 of 9 3PT)Playoff career high, chase-down block on Fox, game-winning tip
Victor WembanyamaSAS249 of 25 FG13 reb

Offense: two players in the 30s. Brunson and Anunoby joined a short list of teammates who each scored 30 or more in a Finals game. The list includes Kobe Bryant and Shaquille O'Neal (2002) and Clyde Drexler and Jerome Kersey (1990) (NBA.com). Brunson carried the scoring early, and Anunoby finished the game. Over the final 23:04, Anunoby went 8 of 8 from the field and 5 of 5 from three (CBS Sports).

Defense: the main source of the swing. Most of the scoring change came from stops. The Spurs shot 4 of 20 in the third quarter and 4 of 19 in the fourth, 8 of 39 (20.5%) for the half. From three, the Spurs went 3 of 17 after the break. Anunoby described the change as contests about 1% better than in the first half (NBC Sports).

Foul trouble. Both teams lost a key player early. Towns had two fouls 62 seconds in. Castle's fourth foul at 10:09 of the third cost San Antonio its best perimeter defender for almost five minutes, and the Knicks gained 8 points in this stretch. Off the bench, NBA.com's coverage pointed to José Alvarado's scoring in the fourth quarter.

Plus-minus: who was on the floor for the swing. Points show who scored. Plus-minus shows how the team did with each player on the court. Towns, Brunson, Josh Hart and Alvarado formed the core of the comeback.

Knicks plus-minus · Game 4

Towns +17 in his minutes. Three reserves −13 or worse.

Blue: the Knicks outscored the Spurs with the player on the floor. Orange: the Spurs did. Source: Basketball Reference plus/minus.

Towns' +17 is the highest on the team. His early foul trouble kept him off the floor while the Knicks fell behind, so his on/off gap is +33, the largest on either team. Mitchell Robinson, Miles McBride and Landry Shamet were each −13 or worse, likely because much of their time came in the first half, when the Spurs made almost every shot. Anunoby scored 33 but was −1, because he played heavy minutes in both halves. On the Spurs' side, Castle was −11 on the floor and +10 off, which supports the foul-trouble explanation (Basketball Reference). Plus-minus is noisy in a single game, since the number depends on who a player shared the floor with and when. Treat these numbers as support for the story, not proof.

The Knicks' side of the swing. New York scored 49 points in the first half and 58 in the second. The Knicks' shooting rose from 15 of 37 (41%) to 21 of 40 (53%), and almost all of the gain came from three: 4 of 12 before the break, 11 of 19 after. At the rim nothing changed (6 of 9 in each half). The three-point jump mattered on the court but is not statistically significant by itself (Fisher's exact test, p ≈ 0.17). The statistical signal for the comeback sits mostly on the Spurs' end.

The box score leaves out moments. Anunoby's line says 33 points. The line does not say he ran down Fox from behind with 11 seconds left to keep the game alive. The line does not say his last two points came on a tip with 1.2 seconds left, off a Brunson three which missed, in an arena which went from silent to deafening in about 20 minutes of game time. Mike Brown called the tip "the most iconic shot in the history of New York basketball" (NBA.com). I agree.

How rare was the comeback?

At 95-75 with 9:33 left, the Knicks' win probability was 0.4%, about 1 in 250 (News Tribune/AP, BET). Game 4 broke or tied three Finals records, and only one playoff comeback has been bigger.

BenchmarkPrevious recordGame 4
Largest Finals comeback (full play-by-play era, since 1997)24 pts, Celtics vs. Lakers, 200829 pts
Largest halftime deficit overcome in a Finals game21 pts, Baltimore Bullets vs. Philadelphia, 194827 pts
Largest deficit after three quarters overcome in a Finals game15 pts, Bulls vs. Trail Blazers, 199215 pts (tied)
Largest comeback in any playoff game31 pts, Clippers vs. Warriors, 201929 pts (2nd)

Sources: NBA.com, Al Jazeera, Arizona Sports/AP.

Was the swing luck or a real change?

The Spurs shot far above normal before halftime and far below normal after. Over full halves, San Antonio made 14 of 26 threes (54%) before the break and 3 of 17 (18%) after.

Test 1: Did the Spurs' shooting change? A one-sided Fisher's exact test on makes and misses by half gives p ≈ 0.018 for threes and p ≈ 0.0002 for all field goals (28 of 47 vs. 8 of 39). Both fall well under the standard 0.05 cutoff, so the drop is statistically significant.

How Fisher's exact test works. Put the threes in a 2×2 table: made or missed, first half or second half. Now assume the Spurs' true shooting never changed. Under this assumption, the 17 made threes (out of 43) landed in each half by chance alone. The test counts every possible way to split those 17 makes between the 26 first-half and 17 second-half attempts. The p-value is the share of splits putting 14 or more makes in the first half: 0.018. "Exact" means the test counts every arrangement instead of using an approximation, which matters with small samples. The field-goal version works the same way, with 36 makes across 47 and 39 attempts.

Test 2: Which half was abnormal? Each three was modeled as a random event with a fixed 36% chance of going in, about the league average (an assumption), using the binomial distribution:

What each symbol means:

How the formula is derived: if every shot is independent with the same make chance p, any one sequence with k makes has probability p^k × (1 − p)^(n − k). Multiply by the number of such sequences to get the chance of exactly k makes. Add over k = 14, 15 … 26 to get the chance of 14 or more. The second-half figure uses the same formula with n = 17, added over k = 0 to 3.

Interactive · binomial distribution

How rare was the hot first half?

Chance of 14+ makes in 26 threes
Chance of 3 or fewer in 17 threes

Bars show the chance of each possible number of makes out of 26 attempts. Orange bars are 14 or more, the Spurs' first half. Move the rate to test the 36% assumption.

Reading the results. Both halves were unusual, in opposite directions, and the hot half was slightly rarer. The Spurs' first 16 threes (11 made) look like a 1-in-127 event on their own. This window was picked after the fact because the shooting was hot, so full halves give the fair comparison. Threes alone do not separate luck from defense, so the next section looks at every shot.

Statistically, the first half was the Spurs shooting above their normal level. As a fan, the first half was 24 minutes of watching almost every Spurs shot go in. Both were true, and only one was going to last.

Luck or defense? A shot-level logistic regression

After accounting for where the Spurs shot from, their odds of making a shot fell by about 84% after halftime (odds ratio 0.16, p = 0.0003). Their shot selection did not get worse, so worse shots do not explain the collapse.

Data. All 163 field-goal attempts were coded from the Basketball Reference play-by-play (86 Spurs, 77 Knicks), each with result, distance and shot type. The Spurs' totals match the box score: 36 of 86 overall and 17 of 43 from three (FantasyData box score). Each shot falls in one of three zones: at the rim (2-pointers from 0–3 ft), other 2-pointers (4+ ft) and threes.

Model. Each shot is a make (1) or miss (0), so the model is a logistic regression. The model estimates how each factor changes the odds of a make:

What each symbol means:

How the model is derived: a straight-line regression predicts impossible probabilities, such as 130%. Modeling the log-odds keeps every prediction between 0% and 100%. The β values are chosen to make the Spurs' real sequence of 86 makes and misses as likely as possible (a method called maximum likelihood). The values were solved with the Newton-Raphson method in Python (NumPy and SciPy).

How to read a β: raise e (about 2.718) to the power of β to get an odds ratio. β₁ came out to about −1.8, and e^−1.8 ≈ 0.16. After halftime, the odds of a make were 0.16 times the first-half level for the same type of shot, about 84% lower. The zone switches sit in the same equation, so β₁ compares like with like: rim shots with rim shots, threes with threes.

The Spurs' shots got easier. After halftime the Spurs took more shots at the rim (28% vs. 21%), fewer threes (44% vs. 55%), and slightly shorter shots on average (15.2 vs. 15.9 ft). Accuracy still fell in all three zones.

Interactive · all 163 field goals

Make rate by zone, before and after halftime

Switch teams to compare. The Spurs dropped in every zone. The Knicks stayed flat at the rim and rose from three.

Results (Spurs).

TermOdds ratio95% CIp
Second half0.160.06–0.440.0003
Other 2-pointer (vs. rim)0.680.18–2.600.58
Three (vs. rim)0.520.16–1.690.28

A likelihood-ratio test for the second-half term gives χ² = 14.8 (df = 1, p = 0.0001). At their first-half zone rates, the Spurs' 39 second-half shots would have produced about 24 makes. They made 8.

The Knicks as a control. The same model for New York shows no significant change after halftime (odds ratio 1.57, 95% CI 0.61–4.04, p = 0.35). In a pooled model with both teams, the gap between the Spurs' change and the Knicks' change is significant (p = 0.0009). The collapse did not hit both teams, so shared causes such as fatigue or the arena do not explain the drop. The change was specific to San Antonio.

Robustness. Six Spurs attempts were tip-ins taken within 3 seconds of a missed shot, which makes them less independent than other shots. Dropping the six leaves the result almost unchanged: odds ratio 0.18 (95% CI 0.06–0.49, p = 0.0009).

What the results mean. The rim numbers matter most. The Spurs made 8 of 10 at the rim before halftime and 2 of 11 after. Point-blank shots depend least on luck, so a drop at the rim points to contact and contests from the Knicks' interior defense more than to chance. The threes (3 of 17) are noisier and fit both explanations. The best reading: the hot first half regressed toward normal, and New York's defense pushed the Spurs well below normal.

8 of 10 at the rim, then 2 of 11. When I first saw this number, I stopped and rechecked the data. Layups are the shots least driven by luck. Something changed in the paint after halftime, and you did not need a model to see the change.

What if the Spurs had shot normally?

The Spurs' second half was replayed 200,000 times at about league-average shooting. The Knicks won about 1% of the replays. The comeback needed San Antonio's collapse. Knicks shot-making alone would not have produced the win.

Method: a Monte Carlo simulation. Everything was held fixed except whether each Spurs second-half shot went in: the Knicks' 107 points, the Spurs' 76 first-half points and the Spurs' 11 second-half free-throw points. Each of the Spurs' 34 second-half shots (tip-in follow-ups excluded) was re-rolled as a make or miss using an assumed make rate for the shot's zone: 65% at the rim, 42% for other 2-pointers and 36% for threes. These rates are approximate league averages, not measured values. The Knicks win a replay if the Spurs finish under 107, and a tie counts as a 50/50 overtime.

Interactive · run the simulation yourself

Replay the Spurs' second half

7 shots, 2 points each
10 shots, 2 points each
17 shots, 3 points each
Knicks win chance
Average Spurs second-half points (real: 30)
Average Spurs final score (real: 106)

Each run replays the Spurs' 34 second-half shots 20,000 times at the rates above. Everything else stays fixed: the Knicks' 107, the Spurs' 76 first-half points and 11 second-half free throws. Blue bars are games the Knicks win.

Spurs' second-half make ratesAvg. Spurs 2nd-half pointsKnicks win
Their real second-half rates (check)2871%
Approx. league average471.1%
Their own first-half rates61under 0.01%

The first row is a check: at the Spurs' real second-half rates, the setup gives the Knicks the win most of the time, as happened. Moving the assumed league rates a few points either way gives a Knicks win chance between 0.5% and 2.4%.

Caveats. The simulation holds the Knicks' scoring fixed. In a real game, more Spurs makes mean fewer Knicks fast breaks and long rebounds, which would likely push New York's odds lower. Dropping tip-in follow-ups slightly understates the Spurs' scoring chances.

How the simulation fits with the regression. The regression shows worse shots did not cause the Spurs' drop. The simulation shows the drop was necessary for the comeback. Together, the two point to the Spurs' collapse, driven by regression to the mean and the Knicks' defense, as the precondition. The Knicks' 58-point second half finished the job.

I keep coming back to this section. In about 99 of 100 versions of the second half, the Spurs shoot like a normal NBA team, the series heads back to San Antonio tied 2-2, and nobody writes this paper. We got the one version where the Spurs missed. As a fan, I find this terrifying and wonderful at the same time.

Win probability, play by play

The Knicks' chance of winning stayed below 10% for 28 straight minutes, from midway through the second quarter to 5:20 left in the fourth. The chance bottomed out at 0.4% at 9:33 of the fourth, then climbed past 50% for the first time with 1:22 left. The regression explains why the comeback happened. The win-probability curve shows when the odds turned.

Rebuilding the score. The free throws from the same play-by-play were added to the shot data, giving a running score after all 110 scoring plays. The rebuilt score matches every reported checkpoint: 41-22 after one quarter, 76-49 at the half, 81-52 at 9:40 of the third, 90-75 after three, 95-75 at 9:33 and 107-106 final. The rebuilt score also shows the Knicks first fell behind by 29 late in the second quarter (71-42), before the third.

The model. The model treats the rest of the game as random drift. The final margin is assumed to follow a normal distribution around the current margin, with a spread which shrinks as the clock runs down:

This is a standard model in sports statistics, first proposed by statistician Hal Stern in 1994. What each symbol means:

The margin drifts at random from now until the buzzer. The spread of the drift is σ × √(t/48). The spread shrinks as the clock runs down. The spread grows with the square root of time because random swings partly cancel out. Dividing M by the spread tells you how many "typical swings" the trailing team needs, and Φ turns the result into a chance. At 9:33, down 20, the spread was about 7.5 points. The Knicks needed a swing about 2.65 times larger than typical, about a 0.4% chance.

Two versions were run:

Interactive · the formula above, live

Win-chance calculator

The calculator uses the same Brownian-motion formula as the paper. Try the presets to see why down 20 with 9:33 left was worse than down 29 in the third.

Interactive · win chance after every score

Bottomed out at 0.4%, then climbed to 99%

Switch between the calibrated model (σ = 16.9) and the simple model (σ = 13.5). The model ignores possession, so treat late-game numbers as rough.

What the curve shows.

Look at the right edge of the chart. 64% with 1:22 left, the first time all night the Knicks were favorites. Then Castle's two free throws drop the chance to 28%. For the next 28 seconds, every Knicks fan was preparing for the worst loss in franchise history. Then the chase-down block. Then the tip. Then 99%. Do you believe what happened? I watched the game, I studied every number, and I am still not sure I do.

Linking the curve to the regression. The flat stretch near zero, from late in the second quarter through the start of the fourth, is where the Knicks' defense did its work. The Spurs' 8-of-39 second half kept the deficit from growing, but the clock kept running. The curve moves up only once New York's offense started scoring (32 fourth-quarter points). This explains why the odds stayed low for so long while the defense held.

Discussion

The numbers say Game 4 was historically rare, but not random. The 0.4% win probability shows what a model saw at 9:33. The model knows nothing about this specific Knicks team.

This team had come back before. Win-probability models treat every team in a given spot alike. New York had already erased a 22-point fourth-quarter deficit against Cleveland in Game 1 of the conference finals. Counting Game 4, the Knicks went 2-0 in these playoffs when trailing by 20 or more in the fourth (WTOP/AP). A league-wide model leaves out team history like this.

The swing came from stops. Comebacks often bring to mind a run of three-pointers. This comeback came from defense: San Antonio scored 30 points in the entire second half. The Knicks' 58-30 second half reflects defense as much as offense.

Inexperience showed late. After the game, Spurs coach Mitch Johnson said his team "got away from" the style which built the lead, in the third and fourth quarters. The Spurs' 3-of-17 second half from three shows this. The cold half was ordinary by itself (8.9% likely), but the cold shooting came against a defense playing its best basketball of the night.

What the numbers do not measure. This paper does not model tempo, and the data does not measure the pressure of playing at Madison Square Garden as the visiting team with a series on the line. Players talk about this pressure. In 2023, after the Knicks beat Cleveland in the first round, Cavaliers center Jarrett Allen said the "lights were brighter than expected" while explaining his struggles. His scoring fell from 14.3 points per game in the regular season to 9.4 in the series (Bleacher Report). San Antonio was a young team in its first Finals, on the road at the Garden.

The closest thing to a pressure test in the data is free throws, the one shot defense does not affect. The Spurs made 6 of 6 before halftime and 11 of 14 after. Two of the misses came from Wembanyama with 1:47 left and the Spurs up 104-103. Brunson's basket at 1:22 then put New York ahead 105-104. Three misses are far too few to prove anything, but the timing fits the crowd's version of the story. Pressure and defense likely worked together, and this data does not separate them.

Tempo and momentum. Almost every account of this game uses these two words, and this paper measures neither directly. Tempo, the number of possessions per minute, would show whether New York sped the game up after halftime. Measuring tempo needs turnover and rebound data this paper did not collect. Momentum is harder. Statisticians have argued about momentum for decades under the name "hot hand." A 1985 study by Gilovich, Vallone and Tversky concluded shooting streaks were mostly an illusion. A 2018 reanalysis by Miller and Sanjurjo found a flaw in the 1985 method and found evidence for a modest hot-hand effect. The fair reading: momentum is real, but smaller than fans think.

The data here shows where the game turned. The Knicks' 13-0 third-quarter run, from 81-52 to 81-65, came entirely while Castle sat with four fouls. In the fourth, New York scored nothing in the first 2:27, then all 32 of its points in the final 9:33. Whether these stretches were momentum or clusters of good possessions, the win-probability curve moved most during them.

Momentum is partly a story we tell after the fact. People look for patterns in streaks. I also remember watching the 13-0 run, and the research lands in a place I accept: momentum is real, but smaller than the feeling. On this night, momentum did not need to be big. Momentum needed to show up for nine and a half minutes.

The game decided the series. A Spurs win would have tied the series 2-2. New York led 3-1 instead and closed out the series three days later. Teams leading a Finals 3-1 have almost always won the title, so Game 4 decided the championship.

Limitations

Conclusion

The Knicks' Game 4 win was the largest comeback in Finals history and about a 1-in-250 outcome at the lowest point. The data point to one precondition: a Spurs collapse after halftime, from 28 of 47 shooting to 8 of 39. A regression shows harder shots do not explain this collapse. A simulation shows the comeback was close to impossible without the collapse, about 1% at normal shooting. Brunson's 36 and Anunoby's 33 turned those stops into points. The result decided the series: New York led 3-1 and won the franchise's first title in 53 years three days later.

I spent this paper trying to be precise about a game which did not feel precise. The numbers say 29 points, 0.4%, about 1 in 100. The numbers do not show how this game felt to a fan used to Knicks seasons ending in losses. Three nights later, the Knicks were champions.

Code and data

For anyone who wants to check the work: the data and code behind every number on this page. The data files open in Excel or Google Sheets. The scripts run in Python 3 (NumPy, pandas, SciPy and Matplotlib).

Sources