Probability
Fourth-Quarter Outcome Probability
Pre-Game Close vs. Q4 Tail Risk
Eagles LIX baseline (14.6% hold) · Broncos 33-pt Q4 vs. Giants (0.2% WP) · Knicks 20+ pt playoff comeback streak (0.1%–2.9% WP).
Analyst · probability & live-pricing exposure · NFL & NBA
A sportsbook trading desk prices two completely different regimes: the pre-game closing line (where point-differential models and public narratives collide, like Super Bowl LIX) and the 4th-quarter live-betting tail (where historical play-by-play models spit out 99.8% or 99.9% win probabilities just before a team scores 33 points in a quarter or erases four 20-point playoff deficits).
Desk · 3 Regimes
Close & Tail
Risk / pricing analog — not betting advice. Evaluates how pre-game rating models (SRS) and in-play win-probability models behave under normal conditions vs. extreme 4th-quarter variance. Built from public NFL/NBA play-by-play checkpoints, Pro-Football-Reference, ESPN Analytics / Next Gen Stats, and Nevada Gaming Control Board figures.
Interactive desk board · 3 projection regimes
Switch between the three case studies below to inspect the exact checkpoint data from data/sports-win-prob-comebacks.csv: the standard pre-game model vs. closing line (Eagles Super Bowl LIX), the NFL 4th-quarter live-liability whip-saw (Broncos 33–32 Giants), and the NBA 3-point variance tail (Knicks 20+ point playoff comeback streak).
Regime 1 (Standard Pre-Game Pricing): When a simple point-differential model (SRS: PHI −3.5) and the closing line (KC −1) diverge by ~4.8 points, the story is public dynasty bias and sportsbook hold.
Case 1 · The baseline: Eagles vs. Chiefs (Super Bowl LIX)
When an independent pre-game forecast and the market close disagree by nearly 5 points, is the posted price stale — and what happens to realized hold?
In a standard pre-game environment, a trading desk starts with team rating differentials, posts an opening number, and shades the line as tickets and money come in. On Super Bowl LIX (Feb 9, 2025), 2024 Simple Rating System (SRS) numbers rated Philadelphia at 7.7 (+160 point differential) and Kansas City at 4.2 (+59 point differential) — a neutral-field projection of PHI −3.5 (60.5% win probability at 3% per point).
Instead, the market closed Chiefs −1 to −1.5 (moneyline KC −120 / PHI +100, de-vigged to 47.8% Eagles). Philadelphia controlled every quarter (24–0 at halftime, 34–0 in the 3rd, 40–22 final). Because the public loaded up on Kansas City parlays and three-peat futures, Nevada sportsbooks held 14.6% ($22.13M on $151.62M handle) — nearly four times the 3.7% hold from Super Bowl LVIII, and more than triple the 4.55% theoretical moneyline vig.
| Metric | Philadelphia Eagles | Kansas City Chiefs |
|---|---|---|
| 2024 regular season (PFR) | 14–3 · PF 463 · PA 303 · PD +160 | 15–2 · PF 385 · PA 326 · PD +59 |
| Simple Rating System (SRS) | 7.7 | 4.2 |
| SRS neutral-site projection | PHI −3.5 · Model P(PHI) = 60.5% | |
| Market consensus close | KC −1 to −1.5 · ML +100 / −120 · De-vig P(PHI) = 47.8% | |
| Model vs. close residual | ~4.8 points · +12.7 percentage points toward Philadelphia | |
| Final & realized state hold | Eagles 40–22 · Vegas hold 14.6% · NY hold 30.6% · PA hold −6.4% (home bias) | |
Case 2 · The NFL 4th-quarter whip-saw: Broncos 33, Giants 32 (Oct 19, 2025)
What happens to an in-play projection when a 7.5-point favorite is shut out for 45 minutes — and why did the pre-game model still lose the spread while winning the game?
If Super Bowl LIX is the textbook pre-game exception report, Week 7’s Giants-at-Broncos game is the ultimate stress test for in-play win-probability models and live sportsbook liability:
- Pre-game baseline vs. 3-quarter reality: Denver closed as a −7.5 favorite (−424 moneyline, 76.5% de-vigged win probability; model projection 27–17). Through three quarters, Denver had 0 points and 111 yards across eight drives, trailing 19–0 entering the 4th.
- The 0.2% floor (7:12 Q4): After the Giants went up 26–8, Denver faced 3rd-and-12 with 7:12 left. ESPN Analytics dropped Denver’s win probability to 0.2% (and Next Gen Stats pegged 4th-and-3 at 6:38 at 0.7%). Live algorithms priced Giants moneyline near −10000 and Broncos around +3500.
- Three lead flips in 1:51 (the whip-saw): Bo Nix ran for a TD, Justin Strnad intercepted Jaxson Dart, and RJ Harvey caught a TD to make it 26–23 at 3:51 (19.1% DEN WP). Nix’s 18-yard TD sweep at 1:51 put Denver up 30–26, spiking Broncos win probability to 96.7% at 1:14. Then a pass interference penalty in the end zone set up Dart’s 1-yard TD run with 0:37 left (32–30 Giants; 89.3% NYG WP). Two completions later (29 yards to Marvin Mims, 23 to Courtland Sutton), Wil Lutz hit a 39-yard walk-off field goal for a 33–32 Broncos win — the most 4th-quarter points (33) in NFL history by a team shut out through three quarters.
Why the model broke at 0.2%
Standard NFL win-probability models treat possession count as fixed (~2.5 drives left in 7 minutes) and assume baseline offensive efficiency. They under-weight correlated state changes: hurry-up tempo, a rookie QB protecting a lead on the road (short fields via turnover), and missed extra points (which left 32–30 vulnerable to a field goal).
The sportsbook “middle”
Here is the desk twist: Denver pulled off a 0.2% miracle to win 33–32 — and still failed to cover −7.5. Pre-game Broncos −7.5 spread tickets lost; pre-game Giants +7.5 tickets cashed; pre-game Broncos −424 moneyline parlays survived; and live bettors who chased Giants −1600 in the 4th quarter were wiped out.
Case 3 · Structural tail risk: The Knicks’ 20-point 4th-quarter comeback streak
When the same team overcomes four 20-point deficits in two postseasons (including 0.1% and 0.4% 4th-quarter win probabilities), is the tail really 1-in-1,000 — or is NBA live pricing using a stale variance parameter?
A single 0.2% NFL comeback can be written off as a black swan. The 2025–2026 New York Knicks proved that in the modern NBA, 4th-quarter win-probability tails are systematically fatter than historical play-by-play buckets suggest:
- 2025 ECSF Games 1 & 2 at Boston (May 5 & May 7, 2025): Prior to this series, NBA playoff teams trailing by 20+ points were 27–830 overall, and no team had ever erased 20-point deficits in back-to-back playoff games. New York trailed 75–55 in Game 1 (winning 108–105 OT as Boston missed a playoff-record 45 threes) and trailed by 20 again in Game 2, facing a 2.9% win probability with 4:46 left before closing on a 30–17 4th quarter (holding Boston to 5-for-24 FG) to win 91–90.
- 2026 ECF Game 1 vs. Cleveland (May 19, 2026 — 0.1% WP): Down 93–71 with 7:49 left in the 4th quarter, ESPN Analytics gave Cleveland a 99.9% win probability (0.1% Knicks). New York outscored the Cavaliers 44–11 over the remainder of regulation and overtime to win 115–104 — the first team in the play-by-play era (since 1997–98) to trail by 20+ in the 4th quarter and win by double digits.
- 2026 NBA Finals Game 4 vs. San Antonio (June 10, 2026 — 0.4% WP): Trailing by 29 in the 3rd quarter and 95–78 with 8:59 left in the 4th (where Finals teams were previously 0–96 since 1970–71, and playoff teams trailing by 20+ in Q4 were 4–733, or 0.5%), the Knicks rallied to win 107–106 on OG Anunoby’s tip-in with 1.2 seconds remaining.
Why historical NBA lookup tables fail
| Empirical bucket | Playoff teams down 20+ in Q4 were 4–733 (0.5%) since 1998; Finals teams down 17+ inside 9:00 were 0–96. |
|---|---|
| 3-point regime shift | When a leading team shoots 40+ threes (like Boston’s 25-for-100 across G1–G2), possession variance doubles compared to 1998–2014 data. |
| Clock-stop asymmetry | Offensive rebounds on missed 3s + transition threes create 8-to-10 point swings in under 75 seconds of game clock. |
What a live trading desk must adjust
| Fatten the tail | Apply a dynamic volatility multiplier \(\sigma_{\text{3PA}}\) when combined 3-point attempt rate exceeds 45% of field goals. |
|---|---|
| Cap extreme ML prices | Never auto-price a 17-to-22 point Q4 lead with 8 minutes left at 99.9% (−10000) when pace and 3PA variance are high. |
| Hedge cash-out liability | Early-payout promos and live cash-out algorithms bleed hold when both sides of a game cross 90% win probability in the same quarter. |
Interactive calculators · Pre-game de-vig & live tail converter
Use the first tool to de-vig any two-way moneyline and inspect posted hold. Use the second tool to compare what a naive historical win-probability model says vs. a fat-tailed live-pricing adjustment (with sportsbook vig applied):
1 · Two-way implied probability & de-vig hold
Type two American moneylines (or click a preset). Outputs raw implied probabilities, posted bookmaker overround, and de-vigged fair win probabilities.
2 · 4th-quarter tail risk & live-odds simulator
See why empirical play-by-play models underprice late comebacks when pace/3-point variance rises, and what fair vs. vigged live moneyline a trading desk should post.
So what · What I would ship on a trading or analytics desk
- Pre-game residual flag (Eagles LIX lesson): Track the gap between point-differential projections (SRS/Elo) and the closing spread alongside ticket vs. money splits. When the public bids a 4.2 SRS team past a 7.7 SRS team, the desk does not need to “pick” the game — it needs to monitor state-by-state liability so a 14.6% Nevada windfall isn’t offset by unhedged regional concentration.
- In-play regime switch (Broncos vs. Giants lesson): Separate pre-game spread liability from live moneyline pricing. When a game enters hurry-up/4-down mode in the final 7 minutes, replace static possession-count decay with tempo-adjusted drive probabilities so a single turnover + TD doesn’t whip-saw win probability from 0.2% to 96.7% to 10.7% to 100%.
- 3-point volatility floor (Knicks playoff lesson): Never price an NBA 4th-quarter deficit purely off 1998–2015 empirical records (like 0–96 or 4–733). Impose a fat-tail probability floor tied to live 3-point attempt rate and offensive rebound rate so live underdog prices and cash-out offers aren’t giving away free optionality.
Limitations & sources
- Case-study selection: Broncos–Giants (Oct 19, 2025) and the Knicks’ 2025–2026 playoff comebacks were selected specifically because they sit in the extreme 0.1%–2.9% tail. To calibrate a full live model, I can swap in a full-season play-by-play CSV to compute Brier scores and calibration curves across all 272 NFL or 1,230 NBA games.
- Sources: Pro-Football-Reference (2024 NFL SRS), Nevada Gaming Control Board (Super Bowl LIX & LVIII handle/hold), ESPN Analytics & Next Gen Stats (Week 7 Broncos–Giants 0.2% / 0.7% / 96.7% / 89.3% win probabilities; Knicks–Cavaliers 99.9% and Knicks–Celtics 97.1% win probabilities), NBA.com & Elias Sports Bureau historical play-by-play records.
- Not betting advice: Analytical portfolio case on probability calibration and bookmaker exposure.