Playoff odds
NHL · how the odds on the playoff race board are made and how they were tested
What the numbers are
Each morning we play out the rest of the NHL regular season 10,000 times and count how often each team makes the playoffs and wins its division. Games already played count as they finished.
- Each remaining game is decided by our pregame win probability model, with every team’s rating frozen at that day: the simulation does not let ratings move with its own made-up results.
- Overtime and shootouts happen at the rates seen in past seasons (23% of games go past regulation; 43% of those reach a shootout), whoever wins.
- A rating is an estimate, so each of the 10,000 runs nudges every team’s strength up or down by a random amount (a standard deviation of 25 rating points) and keeps it for the whole run. Without this the odds months out are too sure. The size was chosen on 2013-14 to 2018-19, before the test seasons were scored.
- Final standings are ordered with the NHL’s tiebreakers; the top three in each division and two wild cards per conference make it.
Our simulation orders the standings with its own code, so on a close day its order can differ slightly from the league’s published standings, which the board shows.
How it was tested
The test and its pass rules were written down before any result was seen. We went back to Nov 15, Dec 15, Jan 15, Feb 15, Mar 15, Apr 1 in each of 2021-22 to 2025-26, ran the simulation as it would have run that day (2,000 runs, ratings frozen at that date), and checked the odds against which teams went on to play a playoff game: 960 predictions in all. Odds show on the site because it passed all four checks:
- ✓ Tiebreakers: on each season’s actual final records, our ordering picked exactly the playoff teams in 11 of 11 seasons (2013-14 to 2025-26, without the short 2019-20 and 2020-21 seasons). A tie on points at the cut line actually came up in 3: 2015-16, 2023-24, 2024-25.
- ✓ Better than the same simulation with no team strength (every game at the usual home win rate, 54.4%): Brier 0.0906 against 0.0988.
- ✓ Better than the standings alone (100% for a team in a playoff spot that day, 0% for one out): Brier 0.0906 against 0.1521.
- ✓ Calibration: in each of the 7 groups below, the share that made it was within the allowed distance of the average prediction (7 of 7).
The Brier score is the average squared gap between the predicted chance and what happened (1 or 0); lower is better. Always saying 50% scores 0.25.
Predicted against actual, 2021-22 to 2025-26
All 960 predictions grouped by predicted chance. A team appears at up to six dates a season, so each group holds fewer independent teams than its count suggests; the allowed gap accounts for that. Gap and allowed are in percentage points.
| Predicted range | Count | Predicted | Made it | Gap | Allowed |
|---|---|---|---|---|---|
| 0–5% | 217 | 0.8% | 1.8% | +1.1 | ±8.0 |
| 5–20% | 121 | 11.2% | 10.7% | −0.5 | ±18.1 |
| 20–40% | 100 | 29.2% | 28.0% | −1.2 | ±28.7 |
| 40–60% | 86 | 49.5% | 48.8% | −0.7 | ±34.0 |
| 60–80% | 90 | 69.9% | 64.4% | −5.5 | ±30.5 |
| 80–95% | 124 | 88.8% | 91.1% | +2.3 | ±17.8 |
| 95–100% | 222 | 99.0% | 100.0% | +1.0 | ±8.0 |
The middle groups are small, so their allowed gaps are wide and this check catches only large errors there.
Season by season
| Season | Model | No strength | Standings |
|---|---|---|---|
| 2021-22 | 0.0590 | 0.0696 | 0.1458 |
| 2022-23 | 0.0751 | 0.0866 | 0.0833 |
| 2023-24 | 0.0727 | 0.0980 | 0.1458 |
| 2024-25 | 0.0979 | 0.0934 | 0.1458 |
| 2025-26 | 0.1480 | 0.1463 | 0.2396 |
| All five | 0.0906 | 0.0988 | 0.1521 |
The model scored worse than the no-strength simulation in 2024-25 and 2025-26.
By date in the season
| Date | Model | No strength | Standings |
|---|---|---|---|
| Nov 15 | 0.1431 | 0.1770 | 0.2500 |
| Dec 15 | 0.1133 | 0.1207 | 0.1625 |
| Jan 15 | 0.1016 | 0.1039 | 0.1875 |
| Feb 15 | 0.0938 | 0.0971 | 0.1500 |
| Mar 15 | 0.0579 | 0.0614 | 0.1000 |
| Apr 1 | 0.0337 | 0.0326 | 0.0625 |
The gap is largest early, when the standings say least. At Apr 1 the model was slightly worse than no strength. The board shows odds from Nov 15 on, the first date the model beat the no-strength simulation; before that no test date covers it. This rule was set before the test ran.
What to keep in mind
- This is the second run of the test. The first failed the tiebreaker check: it used the current NHL tiebreak order for seasons before 2019-20, when the rules were different, and seated the wrong team in 2015-16. We fixed that one rule and wrote a new pre-registration before running again; it allowed no third attempt.
- The edge over the no-strength simulation is small and not certain: a Brier difference of −0.0082, with a 95% bootstrap interval of −0.0178 to +0.0019, which includes zero.
- The model lost to the no-strength simulation in 2024-25 and 2025-26, the two most recent test seasons.
- The game-level win probability model was fit on 2011-12 to 2025-26 that include all five test seasons, so the game predictions inside the simulation are in-sample. That favours the model over both baselines, which use no model.
- Beating the standings alone is a low bar; the comparison with the no-strength simulation carries more weight.
- The model knows nothing about injuries, trades or goaltending. The test covers past seasons; the real test is this one.
The model
Game model wp-nhl-pre-v2-20261112-9dd111f2 · strength spread 25 rating points (log loss 0.3049 on 2013-14 to 2018-19; 0.3086 with none) · past regulation 0.234, shootout share 0.427 (7,462 games) · test pre-registered 2026-11-13 · report sha256 a6a86c48a4c6. The board shows only runs made with this report.