NHL Total Model

Projects NHL game totals, flags overs worth a look, and paper-trades them live.

Live dashboard: https://prezbo8.github.io/NHL-total-model/

The rule

A game is flagged when:

1. the consensus line is 6 or 6.5 (never 5.5), and
2. both teams are projected at or above the high-scoring cutoff (top 40% of last season's team projections; about 3.09 goals in 2026-27).

Every game gets a call:

Call	When	Bet
SLAM	flagged + both teams good or great at 5-on-5	OVER, 1 unit
1U	flagged + one team great at 5-on-5	OVER, 1 unit
PASS	other flagged games	no bet
AVOID	not flagged, or both teams face strong goalies	no bet

Take it at the best over book shown on the dashboard. The paper-trading record uses closing information only: the call at the last run before puck drop, graded at the closing consensus line. P(7+) is the model's chance of 7+ total goals (wins an over 6.5 and an over 6; exactly 6 pushes an over 6). It comes from the projected total and how evenly matched the teams are; calls don't use it.

How the model works

Each team's goals are projected separately for:

• 5-on-5: expected goals + actual goals for/against per 60 (team attack × opponent defense × 5v5 minutes)
• Power play: PP efficiency × opponent's penalty kill × power-play time (penalties drawn × penalties taken)
• Other situations: 4-on-4, 3-on-3 overtime, empty nets (league average)
• Shorthanded goals + the shootout winner's goal: 0.22 a game, league average (counted in betting totals)

then adjusted for:

• Starting goalie: goals saved above expected, calibrated so ratings match how goalies actually play
• Back-to-backs: tired team −8% scoring, its opponent +6.5%
• Team skating speed (NHL EDGE, last season's 20+ mph bursts): +2% own scoring and −1.7% to the opponent per step above average
• Injuries and lineups: injured, suspended or scratched regulars replaced by a replacement-level player (effect sized 1.15×, from the backtest on who dressed)

Only data from the 2020-21 season onward is used.

Data sources

What	Source
Game stats (xG, goals, ice time by situation)	MoneyPuck
Goalie and skater stats	MoneyPuck, NHL stats API
Starting goalies	DailyFaceoff + Rotowire + GoaliePost, combined (Confirmed / Likely / Conflict); actual starters from the NHL box score
Injuries	ESPN + Rotowire
Projected lineups / scratches	DailyFaceoff line combinations vs the official NHL roster
Schedule, scores, rosters	NHL API
Team skating speed	NHL EDGE (NHL API)
Betting lines (open, current, 5 sportsbooks)	Action Network

Backtest (2021-26, 2020+ data only)

	Bets	Win %	ROI
At the opening line	580	53.7%	+2.6%
At the closing line	840	55.1%	+3.0%
Line shopping (best book)			about +2 pts more

The past edge is small (±3–4% uncertainty). The live paper-trading record decides whether it's real. Injury and lineup adjustments were backtested on who actually dressed (2021-26, research/studies/injury_backtest.py): they make projections more accurate and are about the right size.

Every idea tested so far (adopted, kept or rejected, with the numbers) is listed in research/README.md. Check it before testing something new.

Repo layout

.github/workflows/            daily.yml (the run) + keeper.yml (starts it on time) + tests.yml (checks) + monthly.yml (check-up)
src/                          everything the daily run uses
  model.py                      daily entry point: `today` projects, flags and logs; also the original model
  split_model.py                the 5v5 / power-play model used daily
  goalies.py  players.py        goalie and skater data + ratings
  speed.py                      team skating speed (NHL EDGE)
  injuries.py  lineups.py       ESPN injuries, DailyFaceoff projected lineups vs NHL roster
  confirm.py                    starting goalies from DailyFaceoff + Rotowire + GoaliePost combined; Rotowire injuries; actual starters (NHL box score)
  odds.py  books.py  grade.py   betting lines (consensus, opening, every sportsbook) and line math
  paper.py                      paper-trading log: log, settle, report
  health.py                     retries every data source until it works, re-runs in 5 min if one still fails, GitHub issue alerts
  retune.py                     monthly check-up: tier cutoffs, call record, calibration, goalie predictions (GitHub issue)
  store.py                      run history in Supabase (every run's projections; skipped without the key)
  rule_check.py                 one-time SLAM rule check at 75 settled picks (data/rule_check.json)
  trends.py                     last-10 and head-to-head context shown on the dashboard
  moves.py                      how each game's projection moved during the day, and why (data/moves/)
  dashboard.py                  builds the website in docs/ (slate, record, accuracy + P(7+) chart, rankings)
  backfill.py                   adds a past day from pre-game data only (marked retroactive)
  teams.py  paths.py            team names, file locations
research/                     backtests, not part of the daily run (index: research/README.md)
  lib/                          shared helpers: history.py (overs rule on every season), sbr.py (2007-22 odds), lineups_hist.py
  studies/                      one script per tested idea
data/                         CSV data
  paper_trades.csv              the paper-trading log (every logged game, picks, results)
  goalie_*.csv  skater_seasons.csv  opening_lineups.csv  team_speed.csv
  team_rankings*.csv            rankings tab (current + as of each game day)
  odds_*.csv  sbr_odds.csv      historical lines
  game_starts.csv  referees.csv research inputs (start times, referees)
  monthly/                      monthly check-up reports
  key_players/  moves/          per-day key players and projection-move snapshots (dashboard)
docs/                         the website (GitHub Pages)
supabase/schema.sql           Supabase tables (nhl_*) for the run history: paste once into the SQL Editor
tests/check_model.py          57 component checks (paper log, settling, goalies, model, injuries, odds, dashboard, projection moves)

Schedule

Runs at 12:01 AM and every hour from 11:17 AM to 10:17 PM Eastern, plus 15 minutes before each game, all year (daylight saving handled). The time of the last run is shown in the top-right corner of the dashboard.

• .github/workflows/keeper.yml stays running on GitHub and starts Daily NHL run at each slot. GitHub stops any run after 6 hours, so the keeper starts a fresh copy of itself every 5.5 hours; only one keeper runs at a time. (GitHub's own scheduler was too unreliable for this repo: on Oct 2, 2026 seven slots in a row never fired.)
• .github/workflows/daily.yml does the work: settles yesterday's games, re-downloads stats, goalies, injuries, lineups and lines, projects today's games, logs them, rebuilds the dashboard and commits everything back. Its own hourly :17 schedule stays as a fallback.
• Nothing runs on a local machine.

Run it now: Actions tab → Daily NHL run → Run workflow (or gh workflow run daily.yml). If the keeper ever stops (no "Keeper" run in progress on the Actions tab), restart it with Keeper → Run workflow (or gh workflow run keeper.yml).

Running locally

Before pushing: python3 -m pyflakes src/*.py research/lib/*.py research/studies/*.py tests/*.py and python3 tests/check_model.py (the Tests workflow runs both, plus a check that the backtest is unchanged).

pip install pandas==3.0.6 numpy==2.5.3   # Python 3.14
python3 src/model.py today [YYYY-MM-DD]   # projections + flags (only today's date is logged)
python3 src/paper.py report               # paper-trading record
python3 src/dashboard.py                  # rebuild the website
python3 research/lib/history.py           # backtest the overs rule
python3 research/studies/<name>.py        # any study (see research/README.md)
python3 tests/check_model.py              # 57 component checks (also run on GitHub: Tests workflow)

The first run downloads MoneyPuck's game file (~126 MB, not stored in the repo).

Paper trading only, not betting advice.