Overview

IceChaser estimates NHL playoff probabilities using Elo-rated Monte Carlo simulation. Every remaining regular-season game is simulated 100,000 times using win probabilities derived from each team's Elo rating. The fraction of simulations in which a team qualifies for the playoffs is their playoff probability.

Data Sources

All data comes from the NHL public API (api-web.nhle.com) in real time:

• Standings: Current points, wins, regulation wins, games played, clinch status for all 32 teams
• Schedule: Remaining regular-season games (date, home, away)
• Game results: Every completed regular-season game this season, used to build Elo ratings
• Live games: Current game states (FUT/PRE/LIVE/CRIT/OFF/FINAL) — completed games are never re-simulated

Elo Rating System

Each team carries an Elo rating tracking their strength over the season. Ratings start at 1500 and update after every game.

Calibrated Parameters

Parameter	Value	Meaning
K-factor	10	How much a single game moves ratings. Low = stable, high = reactive.
Home bonus	100	Elo points added to home team before computing win probability.
OT discount	0.50	OT/SO wins move ratings at 50% of a regulation win.
Initial rating	1500	All teams start here at season open.

Rating Update Formula

expected = 1 / (1 + 10^((opponent_elo - team_elo - home_bonus) / 400))
k = K_FACTOR × (OT_DISCOUNT if overtime else 1.0)
new_elo = old_elo + k × (actual_result - expected)

Why These Values

• K=10: NHL teams are very stable week-to-week. A low K prevents one fluky game from distorting ratings. Outperformed K=20, 30, and 40 across all test seasons.
• Home bonus=100: Translates to roughly 64% expected win rate for equally-rated teams at home.
• OT discount=0.50: Overtime outcomes are near coin-flips regardless of team quality. Counting at half weight prevents random OT results from polluting true strength estimates.

Win Probability Model

P(home wins) = 1 / (1 + 10^((away_elo - home_elo - 100) / 400))
# Clipped to [0.25, 0.75] to prevent extreme probabilities.

Overtime

Each game has a 24% probability of going to overtime (NHL historical average). If OT occurs, the winner gets 2 points and the loser gets 1 point. The OT winner is determined by coin flip — no significant home advantage in OT empirically.

Outcome	Home pts	Away pts
Home wins regulation	2	0
Away wins regulation	0	2
Home wins OT/SO	2	1
Away wins OT/SO	1	2

Playoff Qualification Rules

After simulating all remaining games, playoff qualification follows the official NHL format:

1. Top 3 teams per division qualify (ranked by points, tiebroken by regulation wins)
2. Next 2 teams per conference qualify as wildcards
3. 8 per conference, 16 total

Simulation Architecture

Single-Pass Design

One 100,000-simulation pass produces everything:

• Playoff odds per team
• Tonight's game scenarios (conditional odds given each possible outcome)
• Best/worst case tonight (joint forced outcomes across all same-conference games)
• What If finish table (playoff odds grouped by W-L-OTL record)
• Tomorrow's scenarios (same analysis for upcoming games)

There is no separate base sim, no Rust binary, no redundant computation. One pass, all outputs.

Performance

Component	Time
NHL API fetch	~2s
Elo rating update	~1s
100k vectorized sim + scenarios + What If	~30–45s
Tomorrow's scenarios	~15s
Total end-to-end	< 60 seconds

Calibration Results

Tested against 3 historical seasons at 5 checkpoints each (30, 20, 15, 10, 5 games remaining). Brier Score: 0.061 (0.0 = perfect, 0.25 = random coin flip).

By Games Remaining

Checkpoint	Brier (all)	Brier (bubble 10–90%)
~5 games left	0.035	0.145
~10 games left	0.037	0.139
~15 games left	0.053	0.171
~20 games left	0.078	0.210
~30 games left	0.113	0.230

Known Limitations

1. No game-context factors: Injuries, back-to-backs, goalie matchups, and motivation are not modeled.
2. Fixed OT rate: 24% applied uniformly. Defensive matchups go to OT more often.
3. Independence assumption: Each game simulated independently. No fatigue/momentum effects.
4. Noise at extremes: Teams below 0.5% should be read as "essentially eliminated," not precisely calibrated.
5. No pre-season carryover: Elo starts fresh at 1500 each October.

Comparison

Model	Basis	Calibrated	Brier
IceChaser v2	Elo (K=10, HB=100)	✅ 3 seasons	0.061
IceChaser v1	Points pace ratio	❌	0.063
Random baseline	50/50	N/A	0.250

Future Improvements

• Goal differential integration — Pythagorean expectation alongside Elo
• Pre-season Elo carryover — regress last season's ratings toward 1500
• Schedule strength adjustment — weight remaining opponents
• Team-specific OT rates — defensive teams go to OT more often