Worked example
How much does this lineup change matter?
A lineup decision has two versions: the projected team and the team after a change. This example compares them over the same 100 simulated games, then shows the inputs and limits behind the difference.
Archived example from Sun 25 Oct, 09:21 UTC. This game has been played or the example expired; there is no current hockey example right now. It is kept so the walk-through still makes sense; a new one appears here when the next one is made.
New York Rangers v Los Angeles Kings
Kickoff
Scenario engine (approximate): not calibrated on held-out matches
1The decision being tested
Out: Adam Fox (D, 1.2 points per game (goals + assists)) · In: Sean Durzi (D, 0.3 points per game (goals + assists))
Out: Alexis Lafrenière (LW, 1.2 points per game (goals + assists)) · In: Oliver Bjorkstrand (RW, 0.5 points per game (goals + assists))
Out: Mika Zibanejad (C, 0.8 points per game (goals + assists)) · In: J.T. Miller (C, 0.7 points per game (goals + assists))
For New York Rangers.
The rule: The home team's three top-scoring starters are taken out and the best available player at the same position comes in. Everything else is unchanged. We fixed it before running anything and ran it for 12 upcoming games. This is the next game on the schedule, shown whatever it moved; in a typical game it moved 2.5 points.
2 · What the simulation showed
Taking out Adam Fox, Alexis Lafrenière and Mika Zibanejad (Sean Durzi, Oliver Bjorkstrand and J.T. Miller in): these runs do not establish a clear change in the simulated result.
The biggest shift was New York Rangers win, from 52% to 49%; the 95% interval for that change runs from −7.0 to +1.0 points, which includes zero. That is not proof of no effect, only that 100 runs cannot tell.
Approximate scenario result. The engine has not been calibrated on held-out matches. These percentages describe its runs, not verified match-outcome probabilities. The 95% intervals measure simulation sampling variability; they exclude model and player-input error.
| Outcome | Before | After | Change, 95% interval | Reading |
|---|---|---|---|---|
| New York Rangers win | 52% | 49% (−3) | −7.0 to +1.0 pts | not clear |
| Los Angeles Kings win | 48% | 51% (+3) | −1.0 to +7.0 pts | not clear |
3Inspect the paired runs
The opponent and random seeds are unchanged; only the listed lineup swaps differ.
Projected lineup · 100 games
- New York Rangers win 52
- Los Angeles Kings win 48
Changed lineup · the same 100 games
- New York Rangers win 49
- Los Angeles Kings win 51
Every game we ran
- New York Rangers v Los Angeles Kings: New York Rangers win 52% → 49% (−3)
- Calgary Flames v Toronto Maple Leafs: Calgary Flames win 49% → 45% (−4)
- Edmonton Oilers v Detroit Red Wings: Edmonton Oilers win 55% → 52% (−3)
- Philadelphia Flyers v Columbus Blue Jackets: Philadelphia Flyers win 49% → 49% (±0)
- Carolina Hurricanes v Boston Bruins: Carolina Hurricanes win 52% → 51% (−1)
- Chicago Blackhawks v Los Angeles Kings: Chicago Blackhawks win 48% → 46% (−2)
- St. Louis Blues v Montreal Canadiens: St. Louis Blues win 52% → 50% (−2)
- Nashville Predators v New York Islanders: Nashville Predators win 50% → 49% (−1)
- Dallas Stars v New Jersey Devils: Dallas Stars win 55% → 49% (−6)
- Winnipeg Jets v Florida Panthers: Winnipeg Jets win 55% → 49% (−6)
- Utah Mammoth v Minnesota Wild: Utah Mammoth win 49% → 46% (−3)
- Vegas Golden Knights v Ottawa Senators: Vegas Golden Knights win 52% → 52% (±0)
What the comparison can tell you
Match forecast when this example was saved (Live logit model, Sun 25 Oct, 09:13 UTC): New York Rangers 59.1% · Los Angeles Kings 40.9%.
The match forecast is the kind of number the sport pages and the forecast record show: a statistical model of final scores, fitted and checked against games played before kickoff. The one above is a saved snapshot; the forecast on the sport page is the current one and moves with new results, so the two can differ. See the current forecast.
The scenario engine (approximate) above is a different model, built to compare one lineup against another. It plays the game out with the chosen players and has not been calibrated on held-out matches, so its percentages can sit far from the forecast (here New York Rangers 52% against 59.1%). Use it to explore how the model responds to the listed change. It does not establish the true effect of a player, a recruitment decision or the chance of a result.
How this example was made
- The game
- New York Rangers v Los Angeles Kings: of the 12 upcoming games from the published schedule that we ran, the first one on the schedule, kept whatever the change did to it.
- Paired runs
- Game i in each lineup uses the same random seed. The 95% change intervals use Newcombe’s paired score interval with continuity correction. They describe sampling variability within this simulation.
- The model
- Saved-player profile score sampler (approximation). Approximate score sampling; NBA/NHL ties resolved.
- What it is not
- 100 fast approximate scenarios, not 100 full animated matches. Same saved players; no automatic substitutions. Sampling intervals do not include model error. Not bookmaker prices.
- Squad data
- Published roster / depth chart: projected, not confirmed starters. Fetched Sun 25 Oct, 09:20 UTC.
- Computed
- Sun 25 Oct, 09:21 UTC. Shown until Mon 26 Oct, 23:00 UTC, then hidden until a new one is built.
- Starting lineups
- New York Rangers
- Mika Zibanejad
- Will Cuylle
- Alexis Lafrenière
- Vladislav Gavrikov
- Adam Fox
- Dylan Garand
Los Angeles Kings- Alex Laferriere
- Artemi Panarin
- Adrian Kempe
- Erik Gustafsson
- Drew Doughty
- Darcy Kuemper
Honest About Uncertainty
A lineup comparison is an estimate from a model, never a guarantee or betting advice. The scenario engine has not been calibrated on held-out matches, and its intervals cover sampling only.
- The same opponent and random seeds in both lineups
- The selection rule fixed before anything was run
- The match forecast is checked separately, on the forecast record
Try your own change
Editing and saving a lineup is free. Running it uses simulations from a plan, and what you get back reads like the example above.
Get a Second Opinion Before Game Time
Compare every sportsbook’s price, simulate matchups and check our forecast record. Start free and upgrade when you need more. Every number is a model estimate, never a guarantee or betting advice.