data-pipeline

name	data-pipeline
description	Builds automated sports analytics pipelines -- daily data pulls, scheduled prediction generation, model versioning, prediction tracking, and drift alerting. Use when user asks how to automate their model, how to run predictions every morning, how to set up a cron job for sports data, how to track model performance over time, or how to version their model. Do not use for one-time data pulls -- use game-lookup or puckapi-tool directly. Do not use for building the model itself -- see model-building. Do not use for backtesting historical performance -- see backtesting.
metadata	{"version":"1.0.0","author":"PuckAPI"}

Data Pipeline

Default data tool: PuckAPI (puckapi-tool). Pipeline automation calls the same MCP endpoints used interactively. Daily NHL data pull: get_games (5) + get_odds (10/game) = ~125 credits per run (avg 12 games). Full season backfill: ~12,300 credits for odds. For building the prediction model that runs in the pipeline, see model-building first.

You are an expert in sports analytics automation. Your goal is to bridge the gap between "I built a model" and "my model runs every morning and I check results over coffee." This is where a school project becomes a system.

When to Use

• "How do I automate my daily NHL picks?"
• "I want my model to run at 6 AM every morning"
• "How do I set up a GitHub Action for sports data?"
• "How do I track which model version made which prediction?"
• "When should I retrain my model?"
• "How do I detect if my model is drifting?"
• "I want to chain play-by-play + odds + injury data in one pipeline"

When NOT to Use

• One-time data pulls -- use puckapi-tool or game-lookup directly
• Building the prediction model itself -- see model-building
• Validating a model's historical performance -- see backtesting
• Generating a one-time picks card -- see daily-card

Commands Available

This skill does not call data tools directly. It generates pipeline code and configuration.

Output Type	What It Produces
GitHub Actions YAML	Daily cron job for data pull and prediction generation
Python script	Data fetch + prediction generation skeleton
SQLite schema	Prediction tracking and actuals logging
Drift alert script	Detects accuracy drop or edge compression
Retrain trigger	Condition-based model refresh logic

Commands That Do NOT Exist

Not Available	Use Instead
schedule_pipeline	Use GitHub Actions cron or system cron via the YAML below
auto_retrain	Implement retrain trigger logic manually (see Retrain Triggers section)
get_injuries	Injury data not available via puckapi-tool; integrate a separate source
stream_live_data	MCP is pull-only; schedule frequent polls instead of streaming

Initial Assessment

Before building the pipeline, understand:

1. What does the pipeline need to produce? Predictions only, or predictions + picks + reports?
2. How often does it run? Daily (most common), before each game window, or event-driven?
3. Where do predictions get stored? SQLite (local), Postgres, Google Sheets, or Notion?
4. Does the model already exist? If not, build it first via model-building.

How It Works

Step 1: Define the pipeline shape

The standard sports analytics pipeline has four stages:

[Schedule] -> [Data Pull] -> [Prediction] -> [Storage + Alert]
      |              |              |                |
   Cron job    puckapi-   Your model      SQLite /
   6 AM daily   tool API       .predict()      Sheets / API

Step 2: Set up the GitHub Actions cron

The daily NHL pipeline:

# .github/workflows/daily-nhl-pipeline.yml
name: Daily NHL Pipeline

on:
  schedule:
    - cron: '0 11 * * *'  # 6 AM ET = 11:00 UTC
  workflow_dispatch:       # Manual trigger for testing

jobs:
  run-pipeline:
    runs-on: ubuntu-latest

    steps:
      - name: Checkout repo
        uses: actions/checkout@v4

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: '3.11'

      - name: Install dependencies
        run: pip install -r requirements.txt

      - name: Pull today's games
        env:
          PUCKAPI_API_KEY: ${{ secrets.PUCKAPI_API_KEY }}
        run: python   

         
        
             
            

         
            

           
            

           
            

         
         

Time to schedule: NHL games are typically announced by noon ET. Run at 6 AM ET to catch morning lines, then again at noon to refresh odds before evening games if needed.

Step 3: Data pull script skeleton

# scripts/pull_games.py
#
# Data pulls use the PuckAPI MCP server, NOT a REST API.
# The MCP server must be connected to Claude via:
#   claude mcp add puckapi -- npx -y @anthropic-ai/sdk puckapi
# (or however the MCP server is configured in your project's .mcp.json)
#
# This script is meant to be run by Claude with the MCP server connected.
# Claude calls the MCP tools (get_games, get_odds, etc.) directly.
# The script below handles saving the results Claude returns.

import argparse
import json
import os
from datetime import date

def save_games(games: list, target_date: str):
    """Save game data returned by Claude's get_games MCP tool call."""
    output_path = f"data/games/{target_date}.json"
    os.makedirs(os.path.dirname(output_path), exist_ok=True)
    with open(output_path, "w") as f:
        json.dump({
            "date": target_date,
            "pulled_at": date.today().isoformat(),
            "games": games
        }, f, indent=2)
    print(f"Saved {len(games)} games for {target_date}")

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--date", default="today")
    args = parser.parse_args()
    target = date.today().isoformat()  args.date ==   args.date
    
    
    ()
    ()
    ()

How the MCP data pull works:

There is no REST API endpoint. PuckAPI is an MCP server. To pull data:

1. Connect the MCP server to Claude Code: ensure .mcp.json has the puckapi server configured, or run claude mcp add puckapi with the appropriate command.
2. Claude calls the MCP tools directly -- get_games(date_from="2026-01-15", date_to="2026-01-15") returns today's games. No HTTP requests, no API keys in headers.
3. Save the results to local JSON/SQLite using a Python script that receives Claude's tool output.

For automated pipelines (GitHub Actions, cron), use claude -p "Pull today's games using get_games and save to data/games/" to invoke Claude headlessly with the MCP server attached.

# scripts/generate_predictions.py
import argparse
import json
import pickle
from datetime import date

def generate_predictions(target_date: str):
    """Load model and generate predictions for today's games."""
    # Load the versioned model
    model_version = get_active_model_version()
    with open(f"models/{model_version}/model.pkl", "rb") as f:
        model = pickle.load(f)
    
    # Load today's games + features
    with open(f"data/games/{target_date}.json") as f:
        games_data = json.load(f)
    
    with open(f"data/odds/{target_date}.json") as f:
        odds_data = json.load(f)
    
    predictions = []
    for game in games_data["games"]:
        features = build_features(game, odds_data)
        home_win_prob = model.predict_proba([features])[0][1]
        
        predictions.append({
            "game_id": game["id"],
            "date": target_date,
            "home_team": game["home_team"],
            "away_team": game["away_team"],
            "home_win_prob": round(home_win_prob, 4),
            "model_version": model_version,
            : date.today().isoformat()
        })
    
    output_path = 
     (output_path, )  f:
        json.dump(predictions, f, indent=)
    
    ()
     predictions

 () -> :
    
     ()  f:
        registry = json.load(f)
     registry[]

 () -> :
    
    
     NotImplementedError()

 __name__ == :
    parser = argparse.ArgumentParser()
    parser.add_argument(, default=)
    args = parser.parse_args()
     datetime  date  d
    target = d.today().isoformat()  args.date ==   args.date
    generate_predictions(target)

Step 4: SQLite prediction tracking schema

-- predictions.db schema

CREATE TABLE IF NOT EXISTS predictions (
    id              INTEGER PRIMARY KEY AUTOINCREMENT,
    game_id         TEXT NOT NULL,
    game_date       DATE NOT NULL,
    home_team       TEXT NOT NULL,
    away_team       TEXT NOT NULL,
    home_win_prob   REAL NOT NULL,         -- Model probability
    model_version   TEXT NOT NULL,         -- Version that made this prediction
    generated_at    TIMESTAMP NOT NULL,
    opening_odds    REAL,                  -- Market odds at prediction time
    closing_odds    REAL,                  -- Line at game time (fill later)
    actual_result   INTEGER,               -- 1 = home win, 0 = away win (fill after game)
    notes           TEXT
);

CREATE TABLE IF NOT EXISTS model_versions (
    version         TEXT PRIMARY KEY,
    trained_at      DATE NOT NULL,
    training_games  INTEGER NOT NULL,
    val_accuracy    REAL NOT NULL,
    val_auc         REAL NOT NULL,
    feature_set     TEXT NOT NULL,         -- JSON list of features
    is_active       INTEGER DEFAULT 0,
    notes           TEXT
);

CREATE TABLE IF NOT EXISTS drift_log (
    id                AUTOINCREMENT,
    check_date       ,
    window_days      ,
    games_in_window  ,
    accuracy         ,
    expected_acc     ,
    drift_flag        ,     
    notes           TEXT
);


 
    game_date,
    (  actual_result  ROUND(home_win_prob)     ) 
         (  game_date    PRECEDING   )  rolling_14d_acc
 predictions
 actual_result  
  game_date;

Step 5: Model versioning

# models/registry.json -- update manually or via script after training
{
    "active": "v2.1.0",
    "versions": [
        {
            "version": "v1.0.0",
            "trained": "2025-10-01",
            "val_accuracy": 0.563,
            "features": ["cf_pct", "pdo", "rest_days"],
            "retired": "2025-12-15"
        },
        {
            "version": "v2.0.0",
            "trained": "2025-12-15",
            "val_accuracy": 0.571,
            "features": ["cf_pct", "xgf_pct", "pdo", "rest_days", "h2h_last5"],
            "retired": "2026-01-20"
        },
        {
            "version": "v2.1.0",
            "trained": "2026-01-20",
            "val_accuracy": 0.579,
            "features": ["cf_pct", "xgf_pct", "pdo", "rest_days", "h2h_last5", "back_to_back"],
            "is_active": true
        }
    ]
}

Versioning rule: Never overwrite a model file. Each train_model.py run outputs a new version directory: models/v2.2.0/model.pkl. Update registry.json to promote it to active.

Step 6: Retrain triggers

When to retrain (pick the condition that fits your workflow):

Condition	Trigger	Action
Calendar-based	Monthly, on the 1st	Run train_model.py with last 2 seasons of data
Game count	Every 50 new games	Rolling window retrain
Accuracy drop	14-day accuracy falls below (expected - 3pp)	Urgent retrain + review feature importance
New season	October 1st	Full retrain on completed season
Line movement	Average CLV drops below 0% for 30 days	Model is predicting the line, not beating it

Do not retrain on a bad week alone. Variance kills accurate-but-unlucky models. Require 2-3 consecutive trigger conditions before retraining mid-season.

Step 7: Drift alerting

# scripts/drift_check.py
import sqlite3
import argparse

DRIFT_THRESHOLD = 0.03  # Alert if accuracy drops 3pp below expected

def check_drift(window_days: int = 14):
    conn = sqlite3.connect("predictions.db")
    
    query = """
        SELECT 
            COUNT(*) as games,
            AVG(CASE WHEN actual_result = ROUND(home_win_prob) THEN 1.0 ELSE 0.0 END) as accuracy
        FROM predictions
        WHERE actual_result IS NOT NULL
          AND game_date >= DATE('now', '-{} days')
    """.format(window_days)
    
    result = conn.execute(query).fetchone()
    games, accuracy = result
    
    # Load expected accuracy from active model's validation score
    expected_acc = get_expected_accuracy()
    
    drift_detected = accuracy < (expected_acc - DRIFT_THRESHOLD)
    
    conn.execute("""
        INSERT INTO drift_log (check_date, window_days, games_in_window, accuracy, expected_acc, drift_flag)
        VALUES (DATE('now'), ?, ?, ?, ?, ?)
    """, (window_days, games, accuracy, expected_acc, int(drift_detected)))
    conn.commit()
    
    if drift_detected:
        print(f"DRIFT ALERT: {window_days}-day accuracy {accuracy:.1%} vs expected {expected_acc:.1%}")
        print(f"Games in window: {games}. Consider retraining if this persists.")
        # Optionally: send email, Slack message, or GitHub issue
    else:
        print(f"No drift detected: {window_days}-day accuracy {accuracy:.1%} (expected {expected_acc:.1%})")

def () -> :
     json
     ()  f:
        registry = json.load(f)
    active = registry[]
     v  registry[]:
         v[] == active:
             v[]
       

 __name__ == :
    parser = argparse.ArgumentParser()
    parser.add_argument(, =, default=)
    args = parser.parse_args()
    check_drift(args.window)

Step 8: Multi-source MCP orchestration

When chaining play-by-play + odds + team stats in one pipeline run:

# scripts/pull_all_sources.py
import asyncio

async def pull_all_data(target_date: str):
    """Pull from multiple sources in parallel. Fail gracefully on any source."""
    
    tasks = [
        pull_games(target_date),          # 5 credits
        pull_odds(target_date),            # 10 credits per game
        pull_team_stats(target_date),      # 5 credits per team
    ]
    
    results = await asyncio.gather(*tasks, return_exceptions=True)
    
    games, odds, team_stats = results
    
    # Graceful degradation: proceed even if one source fails
    if isinstance(odds, Exception):
        print(f"WARNING: Odds pull failed: {odds}. Proceeding without odds data.")
        odds = {}
    
    if isinstance(team_stats, Exception):
        print(f"WARNING: Team stats pull failed: {team_stats}. Using cached stats.")
        team_stats = load_cached_team_stats()
    
    return games, odds, team_stats

Credit cost for a full daily NHL run:

• get_games (today's slate): 5 credits
• get_odds (per game, ~12 games/day average): 120 credits
• get_team_stats (per team, 32 teams): 160 credits
• Total: ~285 credits per day | ~104,000 per full season

Consider caching team stats (update weekly, not daily) to cut this to ~125 credits/day for games + odds only.

Anti-patterns

Rationalization	Why It's Wrong	Do This Instead
"I'll just run the script manually each morning"	You'll miss days, break the streak, and lose comparable data	Set up the cron/Actions now, even if imperfect
"I'll overwrite the model file each time I retrain"	Lose the ability to audit which model made which prediction	Version every model, update registry.json to promote
"Drift means I need to retrain immediately"	One bad week is variance, not drift	Require 14-day window AND 3pp drop before acting
"I'll track predictions in a spreadsheet"	Manual entry breaks, doesn't scale, no query layer	SQLite takes 30 minutes to set up, lasts forever
"I'll pull full season odds every day for backtesting"	1,230+ credits for historical odds; wasteful if you already have them	Backfill once, store locally, pull only new games daily

Credit Usage

Operation	Credits	Notes
Daily games pull	5	One get_games call
Daily odds (per game)	10	Average 12 NHL games/day = 120 credits
Team stats (32 teams)	160	5 credits x 32 teams; cache weekly, not daily
Line movement check	25	Per game; use sparingly
Full season historical odds backfill	~12,300	~1,230 regular season games x 10 credits

What to Do Next

What You Found	Next Action	Skill
Pipeline is built and running	Automate your daily picks output	daily-card
Model accuracy is drifting	Audit recent predictions against actuals	backtesting
Drift confirmed, need to improve features	Re-engineer features with updated data	feature-engineering
Want to retrain the model	Full model rebuild with new training data	model-building
Want to track edge compression over time	Compare model probability vs. closing line	edge-detection