NBA Predictions 🏀

A Streamlit-powered data analytics platform for NBA game predictions and DraftKings Pick 6 analysis.

Features

• Game Predictions: Win probabilities, predicted spreads & totals for every NBA game
• Pick 6 Analysis: Player prop modeling with over/under probabilities and entry building tools
• Team Stats: Interactive team dashboards with rolling averages, rankings, and comparisons
• Player Stats: Individual player analysis, game logs, splits, and trend charts
• Model Performance: Accuracy tracking, calibration plots, and backtesting results

Data Sources

• nba_api — Primary data source for game data, box scores, player stats, and league standings
• Basketball Reference — Advanced metrics, four factors, and historical data
• The Odds API — Market odds for model calibration and edge detection
• ESPN / Rotowire — Injury reports and lineup confirmations

Tech Stack

Category	Tools
App Framework	Streamlit
Data	nba_api, BeautifulSoup4, requests
Analysis	pandas, NumPy, SciPy, statsmodels
ML Models	scikit-learn, XGBoost, LightGBM
Visualization	Plotly, Matplotlib, Seaborn

Roadmap

See the docs/ folder for detailed planning:

• Data Sources — nba_api endpoints, scraping targets, odds APIs
• Features — Team, player, matchup, and Pick 6 feature engineering
• Layout — Streamlit page structure and UI design
• Models — ML model selection, training, and evaluation
• Predictions — Game outcome and Pick 6 prediction methodology

Latest updates (2026-03-24)

• Referee assignments now use nba_official as primary source with ESPN fallback; eliminated duplicates (e.g. J.T. Orr / JT Orr mismatch), and dropped ESPN_GAME_ID from schema.
• Added workflow .github/workflows/referee-assignments.yml scheduled at 0 16 * * * (11 AM ET) and manual workflow_dispatch.
• Implemented scripts/preload_cache.py to prewarm parquet cache + run predictions; integrated into .github/workflows/nightly-pipeline.yml.
• Added get_today_predictions() and run_and_cache_predictions() in utils/data_fetcher.py.
• Added disk cache for get_standings() with same-day freshness guard in utils/data_fetcher.py.
• Fixed injury report team field parsing in get_injury_report() (team_entry.displayName instead of nested team.displayName).
• Moved injury report section to top-level in pages/1_Game_Predictions.py so it renders independent of prediction success.
• Verified and corrected column mapping in Streamlit UI:
  • pages/3_Team_Stats.py: standings columns HOME, ROAD, L10, strCurrentStreak (real API fields), plus humanized labels.
  • pages/4_Player_Stats.py: renamed game log columns with confirmed fields (GAME_DATE, MATCHUP, WL, FG_PCT, etc.).
  • pages/1_Game_Predictions.py: injury report title-case columns and include team correctly.
• Added a robust data verification process to inspect actual parquet schema before renaming columns (avoid assumptions).