35,492 NBA and 6,469 WNBA games

How big a lead is safe, and is it still safe?

Almost every broadcast eventually asks whether a lead is safe. We can answer that question by looking at the score at fixed moments of every regular-season game since the late 1990s. This includes 35,492 NBA games from 1996-97 to 2025-26 and 6,469 WNBA games from 1997 to 2026. We want to see how often the trailing team still wins, by lead and time left, and whether that has changed across six five-season eras. In the NBA, a lead of 10 to 14 entering the fourth quarter is lost more often. It happened 10.1% of the time in the last five seasons compared to 5.8% in the first five. Score margins also swing more than before. In the WNBA, the samples are too small to say either way.

Fourth-quarter comebacks

5.8% to 10.1%

NBA teams trailing by 10 to 14 entering the fourth quarter who won: 1996-97 to 2000-01 against 2021-22 to 2025-26.

The change

+4.3 points

95% interval +2.0 to +6.5; +1.0 to +7.6 after adjusting for all 12 comparisons.

Margin volatility

2.43 to 2.78

NBA sigma, in points per square-root minute, first era against last. The two 95% intervals do not overlap.

Last-minute leads

6.2 times

How much more often NBA leads of 7 to 10 are lost in the final minute than the random-walk model predicts (0.17% against 0.03%).

The question

Two ways to measure a safe lead

The study defines a lead as safe when the trailing team comes back to win 1% of the time or less. It reads the score every five seconds over the last two minutes of regulation and every 30 seconds before that. The study answers in two ways, each with its own claim.

One game read the way the study reads every game: New Orleans against Cleveland on March 21, 2026. Each dot is a checkpoint, where the study records who leads and by how much. New Orleans entered the fourth quarter up 12, inside the 10 to 14 band the era comparison uses, and Cleveland won, so this game counts as one comeback in that situation.

The first is a count. For each lead, time left, and era, it takes the share of games the trailing team went on to win, overtime included. These rates are smoothed so that a bigger lead is never less safe than a smaller one, and each comes with a 95% Wilson interval. To compare eras, the study uses six fixed game situations where the data are dense. An example is a 10 to 14 point lead at the start of the fourth quarter. It puts a Newcombe interval on each change. It also uses a Bonferroni adjustment for running 12 comparisons at once.

The second is a model. Hal Stern's 1994 Brownian-motion model treats the score margin as a random walk with one volatility, sigma, per era or season. The model turns any lead and time left into a win chance. It provides a smooth safe-lead curve. It makes the change in volatility one number with an interval.

Games are included only if their play-by-play final matches the league's game log. The running score must never jump by more than one basket on a single row. We leave out 55 NBA and 27 WNBA games because they fail these checks. Every table was computed twice, in Python and in R, and the two agree on every cell.

Then and now

A double-digit lead entering the fourth is less safe than it was

In the NBA, teams that trailed by 10 to 14 entering the fourth quarter won 5.8% of the time from 1996-97 to 2000-01 (62 of 1,064 games) and 10.1% from 2021-22 to 2025-26 (118 of 1,167). The change is 4.3 percentage points, with a 95% interval of 2.0 to 6.5. After the adjustment for multiple comparisons, the interval stays above zero, 1.0 to 7.6.

How often the team down 10 to 14 entering the fourth quarter wins, by era, with 95% Wilson intervals. The NBA sits near 7.7% for three eras, then steps up to about 10% from 2016-17 on; the WNBA intervals are wide and show no trend.

This rise was not gradual. The rate held between 7.6% and 7.8% for the three middle eras, and then it stepped from 7.8% to 10.2% into 2016-17 to 2020-21 and stayed there. Across the six NBA situations, three show a rise whose interval excludes zero: the start of the fourth quarter up 10 to 14, 6:00 left up 10 to 14, and 1:00 left up 5 or 6. After the adjustment, only the fourth-quarter one remains. This study does not test why the step happened.

The WNBA shows no clear change. None of its six situations has an interval that excludes zero. For example, entering the fourth up 10 to 14, the trailing team won 3.5% of the time in 1997 to 2001. It won 3.7% in 2022 to 2026. The change is +0.2 points. The interval is -4.1 to +3.9. However, the WNBA samples in these situations are 18% the size of the NBA's, so a change as large as the NBA's cannot be ruled out either.

Safe leads

About 20 points is safe entering an NBA fourth quarter

Counting comebacks directly, the 1% safe lead at the start of the fourth quarter was 21 points in 1996-97 to 2000-01. This had a Wilson bounds range of 17 to 21. It was 20 points in 2021-22 to 2025-26, with a range of 18 to 23. With 6:00 left, the lead was 15 (13 to 17) and then 17 (13 to 19).

How often the NBA team trailing in the fourth quarter goes on to win, by lead (rows) and minutes left (columns), smoothed so a bigger lead is never less safe. Above the red line, leads were lost 1% of the time or less. In the recent era the darker cells reach higher in most columns: the same lead is lost more often. The boxed cells are the 10 to 14 point leads entering the fourth.

The NBA's 1% safe lead by minutes left, first era against last. Dots are observed safe leads; dotted lines are the Brownian model's. The model's curve for the recent era sits above the early one at every time left.

These single cells are noisy. The 1% tail rests on a few hundred games per cell. This is why a lead can read safer at 6:00 than at 3:00 in some eras. It is also why the era comparison above uses the denser middle of the distribution instead of the safe lead itself.

The random-walk model

The margin swings more than it used to

The model provides a cleaner view of the same change. NBA sigma rose from 2.43 (95% interval 2.34 to 2.53) in 1996-97 to 2000-01. It reached 2.78 (2.68 to 2.88) in 2021-22 to 2025-26. The two intervals do not overlap. Most of that rise occurred at the same point as the comeback step: sigma was 2.55 in 2011-12 to 2015-16 and 2.81 in 2016-17 to 2020-21. The model's 1% safe lead entering the fourth quarter moved from 19.6 points (18.8 to 20.4) to 22.4 (21.6 to 23.2). In the WNBA, sigma went from 2.14 (1.96 to 2.34) to 2.44 (2.27 to 2.64), and those intervals overlap.

Score-margin volatility, sigma, fitted season by season with 95% intervals. NBA seasons from 2016-17 on sit at or above the top of the earlier range (the season intervals overlap), while the WNBA's seasons move up and down inside much wider intervals with no clear trend.

In this model, sigma links the two findings. With no drift, the model spreads the margin over the remaining minutes in proportion to sigma. A wider spread leaves more outcomes below zero at the buzzer. From a 12-point lead entering the fourth quarter, the model gives the trailing team a 7.7% chance with the first era's sigma and 10.6% with the last era's. The observed rates for leads of 10 to 14 moved from 5.8% to 10.1%, in the same direction.

The model's view of the fourth quarter from a 12-point NBA lead. The shaded fan is where the margin is likely to be as the quarter plays out; the curve at the buzzer is the margin's final spread, and its red tail is the share of games the trailing team wins. The recent era's sigma widens the fan (the dashed line is the first era's edge) and enlarges the red tail.

Where the model breaks

A random walk has no term for fouling or team strength

The model is a smooth summary, not a calculator for the endgame. In the NBA's last minute, leads of 7 to 10 are lost 0.17% of the time, against the model's 0.03%, or 6.2 times as often. The likely reasons are late-game fouling, three-point attempts, and timeouts, which add variance that a constant-rate random walk has no room for. The error runs the other way for big leads with time left. With 6 to 12 minutes to play, leads of 16 or more are lost 0.54% of the time. This is compared to 0.79% predicted. Big leads belong disproportionately to better teams, and the model has no term for team strength (an interpretation, not tested here). For the endgame, the observed counts are the better guide.

How far the model is off, by time left and lead (NBA, all eras pooled): observed comebacks divided by the model's prediction, on a log scale. On the dotted line at 1 the model is right. Points where the model expects fewer than five comebacks are left out.

Limits

What this does not show

These are regular-season games only; the playoffs are a different population. The term "safe" is unconditional: a strong team's 10-point lead is safer than these tables say, and a weak team's lead is less safe. The running score comes from the league's play-by-play score stamps, and a stamp wrong by exactly two or three points would pass both data checks. The eras are fixed five-season blocks chosen before any result was seen, not fitted break points, and the six comparison situations were picked after a first look at comeback rates by era, so they are reported with both plain and adjusted intervals, whatever they show. Finally, the study measures that the NBA's fourth-quarter leads got less safe; it does not test why.

Method, code, and verification

The full technical write-up (data provenance, model specification, validation gates, and stated limitations) is the README. The analysis is implemented independently in R and Python and reconciled to numeric tolerance; run_checks.sh replays every gate.

Key terms

95% interval: A range computed so that, over many repeated samples, about 95% of such ranges would contain the true value.
Wilson interval: A 95% interval for a rate, such as a comeback rate. Unlike the simplest formula, it stays reliable when a rate is near zero or a sample is small. Its two ends are the Wilson bounds.
Newcombe interval: A 95% interval for the difference between two rates, built from each rate's Wilson interval.
Bonferroni adjustment: A correction for making many comparisons at once. Each interval is widened so that the chance of any false alarm across all the comparisons stays near 5%.
Brownian-motion model: Hal Stern's 1994 model of a game's score margin. It treats the margin as a random walk, so it needs only the lead, the time left, and sigma to give a win chance.
random walk: A path in which each step is random and does not depend on earlier steps.
sigma: The model's measure of volatility, in points per square root of a minute. The margin's typical swing over T minutes is sigma times the square root of T.
drift: A steady tilt of the margin toward one team. With no drift, the model treats the margin as equally likely to move either way from here.
safe lead: The smallest lead at which that lead and every larger one are lost no more than a set share of the time. The 1% safe lead uses 1%. The model's version is the lead where its win chance for the trailing team falls to 1%.
smoothed: Adjusted so that a bigger lead never shows a higher comeback rate than a smaller one. The method, pool-adjacent-violators, averages neighboring rates that break this order, weighted by games.
percentage points: The plain difference between two percentages. A move from 5% to 7% is 2 percentage points.
checkpoint: A fixed moment in a game when the study reads who leads and by how much.
play-by-play: The league's event-by-event log of a game, including the running score.