Variance97

Data science investigation of Connor McDavid's performance in high-stakes hockey across the NHL Stanley Cup Playoffs (2021–22 through 2024–25), the 2025 Four Nations Face-Off, and the 2026 Winter Olympics.

The question, narrowed

The popular narrative is that Connor McDavid can't win the big one. The data tells a more specific story:

• McDavid won the 2025 Four Nations Face-Off, scoring the OT winner himself in the final.
• McDavid set the Olympic scoring record with 13 points in 6 games at the 2026 Milan Cortina Games.
• McDavid's individual Stanley Cup Finals production drops about 0.28 pts/game vs his regular season — but Nathan MacKinnon's drops twice as much (0.54 pts/game) and MacKinnon won the Cup in 2022.

So the working thesis isn't "McDavid underperforms in championship-level games." It's narrower: his teams keep losing deep playoff runs even when his individual production isn't unusually low for an elite forward. The project tests where the predictive signal actually lives, against three hypotheses:

1. H1 — Individual: McDavid's personal output collapses in championship-level games.
2. H2 — Team Construction: Edmonton's supporting cast fails around him.
3. H3 — Matchup-Specific: elite goaltender + elite defensive system specifically suppresses him (Bobrovsky/FLA, Hellebuyck/USA).

What each phase actually does

Phase 1 — Exploratory Data Analysis (01_data_loading_and_exploration.ipynb)

Three acts (Stanley Cup Playoffs / Four Nations / Olympics) plus a synthesis section that adds a peer comparison against Nathan MacKinnon and explicitly documents the two structural confounds in the dataset (every Edmonton Stanley Cup Finals appearance is vs Florida; the Hellebuyck sample is n=3 across one tournament window). The takeaway: H1 is weakened by peer comparison, H3 is the best-fitting hypothesis but cannot be statistically isolated from confounds without more data.

Phase 2 — Statistical Validation (02_statistical_validation.ipynb)

Four formal tests with effect sizes (Cohen's d) alongside p-values and Bonferroni correction (k=4 → α=0.0125). Includes a peer-comparison test (McDavid SCF vs MacKinnon SCF) that the original version was missing. None of the tests reach significance — the dataset is structurally underpowered (n=3 to n=13 for the playoff/championship contexts) — and we no longer use "trending toward significance" framing. The peer-comparison non-result is itself informative: McDavid's SCF output is higher than MacKinnon's, directly contradicting the popular thesis.

Phase 3 — Feature Attribution (03_ml_model.ipynb)

Reframed from "logistic regression predicting pointless games" to Ridge regression on points/game with real gameplay features: opp_ga_per_game, rolling_pts_5, rest_days, is_back_to_back. Scoped to NHL games only. The result that matters: when game_context_stanley_cup_finals has to compete against gameplay features instead of standing alone, its coefficient drops from +0.67 (original) to roughly −0.06. The variance the original model attributed to "Stanley Cup Finals" reroutes to game_number (late-series fatigue) and opp_ga_per_game (opponent quality). The "Stanley Cup Finals effect" was largely a late-series + tough-defense effect masquerading as a context label.

Phase 4 — NHL API Pipeline (04_nhl_api_pipeline.ipynb)

Self-updating dataset off api-web.nhle.com. data/build/update_all.py orchestrates: cursor-based incremental fetch of McDavid + MacKinnon gameLog, per-new-game boxscore enrichment (so result/team_score/opp_score are populated), standings refresh into opponent_team_stats.csv, concat of the manual international_games.csv, and a full re-run of apply_features.py so is_elimination_game, rolling_pts_5, rest_days, is_back_to_back, and opp_ga_per_game stay consistent with the latest rows. Idempotent — reruns with no new games report +0 and exit cleanly, so it's safe on a daily cron.

Phase 5 — Interactive Dashboard (app/)

Multi-page Streamlit app built around the reframed thesis, not a point-prediction toy. Home.py opens with the headline (Four Nations win, Olympic record, McDavid's SCF drop is smaller than MacKinnon's). Pages: 1_Three_Acts (Playoffs / Four Nations / Olympics, interactive), 2_Peer_Comparison (the strongest finding, McDavid vs MacKinnon by context), 3_Feature_Contributions (per-game Ridge coefficient × standardized feature decomposition — explicitly not a "will-he-score-tonight" predictor), 4_Limitations (Florida confound, Hellebuyck n=3), and 5_Pipeline_Status (latest game date, row count, CSV mtime). The app only reads the clean CSVs — no API calls happen from the app itself; Phase 4 owns all external I/O.

Data

The dataset is built and refreshed by the Phase 4 pipeline (data/build/). Sources of truth are the API-derived NHL logs and the manual international file; everything else is regenerated each pipeline run.

File	Source	Description
data/mcdavid_nhl_log.csv	NHL API	NHL-only McDavid game log. Rebuilt incrementally by the pipeline.
data/mackinnon_nhl_log.csv	NHL API	NHL-only MacKinnon game log (peer baseline). Rebuilt incrementally by the pipeline.
data/international_games.csv	Manual entry	Four Nations / Olympics rows — the NHL API doesn't cover these.
data/opponent_team_stats.csv	NHL API standings	Per-season GA/game for every team. Refreshed each pipeline run.
data/mcdavid_game_log_clean.csv	Pipeline output	Merged + featured (NHL + international). Consumed by Phases 1–3.
data/mackinnon_game_log_clean.csv	Pipeline output	Merged + featured (NHL only). Consumed by Phases 1–2.

is_elimination_game rule

A game is an elimination game if a single loss ends the run:

• NHL playoff series: True when the opponent already has 3 series wins entering the game.
• International knockout games (quarterfinals, semifinals, finals): always True.
• Regular season, group stage, exhibition: always False.

Rule logic lives in data/build/apply_features.py; results are materialized into the clean CSVs.

Refreshing the dataset

bash scripts/run_update.sh           # one-shot CLI wrapper
# or, equivalently:
python3 data/build/update_all.py

Idempotent — running with no new games reports +0 and exits cleanly. Safe to put on a daily cron during the season.

Live dashboard

Phase 5 ships an interactive Streamlit dashboard built around the reframed thesis. It does not predict whether McDavid will go pointless tonight — that would oversell what the data supports. Instead it surfaces the headline finding (peer comparison vs MacKinnon), the three-act narrative, per-game feature contributions from the Phase 3 model, and the limitations.

• Live URL: (https://variance97-s26u7gyazbqesy5jvmhggw.streamlit.app/)

Run locally

pip install -r requirements.txt
streamlit run app/Home.py

The app reads the clean CSVs from data/ directly — no API calls happen from the app itself. Phase 4's pipeline keeps those CSVs fresh.

Tech stack

• Python (pandas, numpy, scipy, scikit-learn)
• Jupyter (analysis notebooks)
• Streamlit + Plotly (dashboard)
• NHL public API (api-web.nhle.com)

Repository structure

data/
    build/
        fetch_player_log.py        # generic NHL gameLog fetcher
        fetch_boxscores.py         # adds result/team_score/opp_score
        fetch_team_stats.py        # NHL standings -> team GA/game
        apply_features.py          # is_elimination_game + ML features
        update_all.py              # pipeline orchestrator
    mcdavid_nhl_log.csv            # API source
    mackinnon_nhl_log.csv          # API source
    international_games.csv        # manual entry
    opponent_team_stats.csv        # team GA/game by season
    mcdavid_game_log_clean.csv     # pipeline output (analysis input)
    mackinnon_game_log_clean.csv   # pipeline output (analysis input)
notebooks/
    01_data_loading_and_exploration.ipynb
    02_statistical_validation.ipynb
    03_ml_model.ipynb
    04_nhl_api_pipeline.ipynb      # Phase 4 pipeline demo
app/
    Home.py                        # Streamlit entry point (headline)
    pages/
        1_Three_Acts.py
        2_Peer_Comparison.py
        3_Feature_Contributions.py
        4_Limitations.py
        5_Pipeline_Status.py
    components/                    # data loaders, charts, model, narrative
.streamlit/config.toml             # theme + server config
scripts/
    run_update.sh                  # CLI wrapper for cron / CI
requirements.txt
LIMITATIONS.md
PHASE4_PLAN.md
PHASE5_PLAN.md
README.md

Honest summary

The project's most interesting finding is the one that contradicts its own original framing: McDavid's individual Stanley Cup Finals production is not unusually low for an elite forward — MacKinnon, who actually won, dropped twice as much. Where the predictive signal does live, once real gameplay features are introduced, is late-in-series fatigue (game_number) and opponent defensive quality (opp_ga_per_game) — not the "championship" label. That is the narrower, defensible claim Phase 4's pipeline keeps fresh and Phase 5's dashboard puts in front of a reader inside 30 seconds.