BIG_GAME // SUPER_BOWL_LI // 2016_SEASON

28–3: SUPER BOWL LI AS A WIN-PROBABILITY CHART

Everybody knows the number. Almost nobody has seen the shape of it. Here is Atlanta's collapse plotted play by play — the plateau at the top of the mountain, and the cliff on the other side.

BY PROFESSOR CHEDDARHEAD · 2026-07-05 · GAME: 2017-02-05 · DATA: NFLVERSE / CHEDDARHEAD_ANALYTICS

New England Patriots

WIN (OT)

—

ATL

Atlanta Falcons

LOSS

ATL PEAK WIN PROB

99.4%

3rd qtr, up 28–3

NE LOW WIN PROB

0.6%

at its worst

NE DRIVE EPA

+14.6

on offense

ATL DRIVE EPA

+6.9

on offense

Atlanta's win probability across all 169 plays of Super Bowl LI. It climbs steadily through the first half, plateaus above 95% for most of the third and fourth quarters — peaking at 99.4% — then falls off a cliff in the final minutes and into overtime, ending near zero. — **ATLANTA'S WIN PROBABILITY, EVERY PLAY.** The line touches 99.4% and stays in the high 90s for roughly a quarter and a half. Then the cliff. Win probability is a model, not a promise — but a model this confident being this wrong is the whole story.
[ OPEN THIS GAME IN DRIVE_TRACKER → ]

THE PEAK, AND THE CLIFF

At its highest point, our win-probability model gave Atlanta a 99.4% chance of winning Super Bowl LI. That is early in the third quarter, moments after Tevin Coleman's touchdown made it 28–3. The number is not a typo and it is not hindsight. It is what the math said at the time, given the score, the time remaining, and the field position. New England's win probability bottomed out at 0.6%. One team in six-hundred-and-some.

What the chart shows that the memory blurs is the plateau. This was not a lead that evaporated in one bad snap. Atlanta sat above 95% for the better part of a quarter and a half. There was time. There were possessions. The collapse needed all of them — which is exactly why it still doesn't feel real. A fumble you can explain. A ninety-minute slide from near-certain champion to overtime loser is harder to file away.

It's worth sitting with what 99.4% actually means, because the number gets thrown around like a verdict. It is not a guarantee; it is a frequency. If you could replay this exact situation — 25-point lead, that much clock, that field position — a thousand times, the team in Atlanta's shoes wins about 994 of them. This was one of the six. That is the honest reading, and it is also why "the model was wrong" misses the point. The model wasn't wrong. Atlanta was that far ahead. The rare thing simply happened, on the biggest stage there is, to the team that could least afford it.

HOW ATLANTA BUILT A 25-POINT LEAD

Expected points added per drive for both teams in Super Bowl LI. Atlanta's early drives are tall positive green bars — two touchdown drives worth about +5.9 and +5.3 EPA, plus a +6.9 EPA touchdown early in the third quarter. New England has a deep negative bar on drive 9, the drive that ended in a pick-six against them. — **EXPECTED POINTS ADDED, PER DRIVE.** Atlanta (green) piled up the tall positive bars early — touchdown drives worth **+5.9**, **+5.3**, and **+6.9** expected points. New England's deepest hole is drive 9: Brady's interception returned for a touchdown, an **−8.2 EPA** possession that made it 21–0.

For a half and change, Atlanta earned every point. The expected points bars aren't close: the Falcons turned drives into touchdowns while New England turned drives into punts and, once, into a pick-six going the other way. Robert Alford's 82-yard interception return is the tall red bar — a single New England possession worth −8.2 EPA, because it didn't just end a drive, it handed Atlanta seven points. By the time Coleman scored to make it 28–3, the Falcons had out-gained, out-schemed, and out-EPA'd the greatest quarterback-coach pairing of the era for forty minutes.

A strip showing the outcome of all 23 drives in sequence. Atlanta's drives are frequent green touchdowns early, then a run of orange and gray punts late. New England's drives shift from punts and a turnover early to a dense cluster of touchdowns and field goals at the end. — **DRIVE OUTCOMES, LEFT TO RIGHT.** Read it like a timeline. Atlanta's green (touchdowns) is front-loaded; New England's is stacked at the very end. The two halves of this game were played by two different teams wearing the same jerseys.

THE PLAY THE MODEL NEVER FORGAVE

Comebacks have a hinge. On the win-probability chart, Atlanta's line holds near the top until one possession breaks its back: drive 17. Up 28–12 late in the fourth, with the ball and a chance to bleed the clock or add points, Matt Ryan was strip-sacked by Dont'a Hightower. Atlanta's offense generated −5.4 EPA on that possession — the worst drive of their night — and gave New England the ball in scoring range. The Patriots scored, converted the two-pointer, and the arithmetic that had felt academic all evening suddenly had a pulse.

THE UNDERRATED VILLAIN: everyone remembers the Julian Edelman catch that had no business being a catch — a 23-yard grab off three defenders' hands on the tying drive. Fair. But the model points a little earlier, at the Hightower sack. That is the play where a still-comfortable lead became a coin flip. The Edelman catch was New England spending the probability the sack had already handed them.

From there the chart does the thing that makes people who watched it live go quiet. New England's final regulation touchdown drive tied it at 28 and swung Atlanta's win probability by roughly 34 points on a single drive. Then the overtime coin toss, the drive nobody in Atlanta got to answer, and James White's two-yard run: a +6.1 EPA possession that ended it. Falcons' win probability, final: near zero. The mountain, and then the sea.

WHAT THE NUMBERS SAY vs WHAT THE STORY SAYS

// THE NUMBERS

New England was the better offense on the night by expected points — +14.6 to +6.9 in drive EPA — and it wasn't especially close down the stretch. The model was never wrong to give Atlanta 99.4%. It was reporting a real, enormous edge. Improbable is not impossible; 0.6% events happen 0.6% of the time.

// THE STORY

A lead that large isn't lost on one play — it's lost on a dozen small ones, each survivable alone. A sack here, a holding call there, an overtime rule that never let the 99.4% team touch the ball. The chart makes the collapse look like physics. Living it felt like a slow leak nobody could find.

THE DRIVE LOG

Every drive, in order, with the possessing team's expected points added and its net win-probability shift. This is the raw material behind the charts above — and behind every number in this article. All of it is computed from the same play-by-play; none of it is from memory.

Super Bowl LI drive log — NE 34, ATL 28 (OT). EPA and WP shift are from the possessing team's perspective.
#	TEAM	RESULT	DRIVE EPA	WP SHIFT
1	NE	PUNT	-0.17	-3.1%
2	ATL	PUNT	+0.87	-1.0%
3	NE	PUNT	+0.31	-0.7%
4	ATL	PUNT	-0.36	-2.3%
5	NE	TO	-1.56	-3.4%
6	ATL	TD	+5.85	+15.2%
7	NE	PUNT	-2.26	-4.4%
8	ATL	TD	+5.25	+7.3%
9	NE	DEF TD	-8.17	-6.0%
10	NE	FG	+2.08	-0.4%
11	ATL	PUNT	-2.57	-3.8%
12	NE	PUNT	-2.37	-3.9%
13	ATL	TD	+6.93	+4.5%
14	NE	TD	+6.15	+0.4%
15	ATL	PUNT	-3.31	-0.8%
16	NE	FG	+2.83	-0.7%
17	ATL	TO	-5.40	-3.5%
18	NE	TD	+3.87	+0.8%
19	ATL	PUNT	-0.21	+1.3%
20	NE	TD	+7.89	+33.9%
21	ATL	PUNT	-0.14	-14.9%
22	NE	END	-0.06	-2.6%
23	NE	TD	+6.07	+43.0%

THE PROFESSOR'S VERDICT

PROFESSOR'S VERDICT

The model gave Atlanta 99.4% and the model was right to. That is the uncomfortable lesson of Super Bowl LI: a number can be honest and still lose. New England was the better team over sixty-plus minutes by every EPA measure we have, and a lead built on forty good minutes was undone by twenty catastrophic ones. Win probability doesn't tell you what will happen. It tells you what's likely — and then, once in a very long while, it lets you watch the other 0.6% arrive in real time.

[ DRIVE_TRACKER ] [ NE_HUB ] [ ATL_HUB ] [ WHAT_IS_WIN_PROBABILITY? ] [ THE_28-3_NOBODY_REMEMBERS → ]