NBA Player Predictions — End-to-End (R + Python)

Link to Shiny App for final product: https://jmotta31.shinyapps.io/NBA_Prediction_Tool/

Portfolio project that builds a daily NBA player-level prediction dataset and a Shiny app to explore it.
Data is gathered and engineered in R (hoopR) and Python (nba_api + custom pipeline), modeled with a Bayesian-style neural network (MC Dropout, and scheduled on Windows via Task Scheduler.

Status: designed to run daily during the NBA season (Oct 21 – Jun 22) and skip the offseason.

---

Highlights

• Automated data refresh (R): pulls player box scores, schedule, and builds lineup/on-off features; guarded by a month/day season window.
• Feature & model pipeline (Python): scraping/assembly → feature engineering → preprocessing → BNN training.
• Daily predictions: generates predictions/nba_predictions_YYYY-MM-DD.parquet for upcoming games; Shiny app hot-reloads.
• Launchers for production: small wrapper scripts handle season windows, logs, and lock files to avoid overlapping runs.
• Shiny UI: dark theme, sortable/filterable tables, and a lightweight implied-probability calculator.

---

Repository layout

.
├─ R_Scripts/
│  ├─ update_nba_data.R              # Pulls hoopR data; builds lineup/on-off; writes parquet datasets/
│  ├─ launch_update_nba_data.R       # Season guard + logs + lock; calls update_nba_data.R (Task Scheduler)
|  ├─ www                            # DALL-E generated images for the Shiny App
│  └─ app.R                          # Shiny app; reactive polling for predictions + metrics (parquet)
│
├─ Python_Scripts/                   # (or project root, depending on your layout)
│  ├─ scraping_loading.py            # Scrapes advanced data, merges with hoopR gamelogs + injuries, light FE
│  ├─ feature_engineering.py         # Main feature engineering block
│  ├─ preprocessing.py               # Train/test split, imputers, encoding, pipelines, scalers
│  ├─ bnn.py                         # Bayesian-like NN (MC dropout for quantiles)
│  ├─ run_pipeline.py                # Orchestrates training end-to-end
│  ├─ nba_predictions.py             # Loads artifacts; produces daily predictions parquet
│  └─ launch_nba_predictions.py      # Season guard + logs + lock; calls nba_predictions.py (Task Scheduler)
│
├─ predictions/                      # Daily predictions parquet (nba_predictions_YYYY-MM-DD.parquet)
└─ README.md

---

How the pipeline runs (daily)

1. Data refresh (R)

   • R_Scripts/update_nba_data.R

     • Pulls player gamelogs via hoopR.
     • Builds upcoming schedule.
     • Derives lineup stints and on/off summaries (parallelized).
     • Writes parquet files in datasets/.

   • R_Scripts/launch_update_nba_data.R

     • Season window (month/day only), timestamped logs in logs/, and a lightweight lock file (prevents overlap).
     • Called by Windows Task Scheduler every morning.

2. Modeling & predictions (Python)

   • scraping_loading.py assembles a unified dataframe (nba_api + hoopR + injury flags).
   • feature_engineering.py derives features
   • preprocessing.py builds pipelines (imputers/encoders/scalers); bnn.py defines a Bayesian-style NN (MC dropout).
   • run_pipeline.py trains the model; artifacts saved under models/ and pipelines/.
   • nba_predictions.py loads artifacts, scores upcoming games, filters by available props, and saves predictions/nba_predictions_YYYY-MM-DD.parquet locally and pushes a separate nba_predictions.parquet to this repo to update the Shiny App's predictions reactable.
   • launch_nba_predictions.py wraps the run for scheduling (season window, logs, lock).

3. Shiny app

   • app.R reads:
     • predictions/nba_predictions.parquet via GitHub URL.
     • datasets/Evaluation_Metrics.parquet via GitHub URL.
   • No republish needed day-to-day; the app picks up the latest parquet.

---

Quickstart

This project uses both R and Python. Any recent R (≥4.3) and Python (≥3.9) is fine.

R setup

# From an R console:
install.packages(c("dplyr","arrow","hoopR","lubridate","tidyr","data.table","purrr","stringr","tibble","parallel"))

Run once at the project root to populate datasets/:

Rscript R_Scripts/update_nba_data.R

Python setup

Create/activate an environment, then:

pip install numpy pandas pyarrow joblib scikit-learn torch unidecode nba_api
# (plus any others used in your scripts)

Train models:

python Python_Scripts/run_pipeline.py

Generate daily predictions (writes to predictions/):

python Python_Scripts/nba_predictions.py

Shiny app

Run locally:

# from project root
shiny::runApp("app.R")

Environment overrides (optional):

• PREDICTIONS_DIR (default: predictions)
• METRICS_PATH (default: datasets/Evaluation_Metrics.parquet)

---

Windows Task Scheduler (daily, in-season only)

Create two daily tasks (one for R data refresh, one for Python predictions). Each task:

• Trigger: Daily at a morning time you prefer.
• Action → Start a program:
  • Program/script: "<path to Rscript.exe>" or "<path to python.exe>"
  • Add arguments: "<full path to launcher>"
    • R: "...\R_Scripts\launch_update_nba_data.R"
    • Py: "...\Python_Scripts\launch_nba_predictions.py"
  • Start in: "<project root>" (e.g., C:\Users\...\NBA_Prediction_Tool)
• Run whether user is logged on or not
• Run with highest privileges (if needed for paths)

The launchers handle:

• Season window: Oct 21–Dec 31 and Jan 1–Jun 22 (month/day only).
• Lock file: avoids overlapping runs.
• Logs: timestamped files under logs/ plus a data_refresh_latest.log or similar copy.

---

Modeling notes (BNN)

• The network in bnn.py uses MC dropout at inference to approximate predictive uncertainty.
• Quantile outputs (e.g., 10th/50th/90th) enable prediction intervals per target (Points, Assists, Rebounds, 3PTM, Steals, Blocks).

---

Shiny App

• Link: https://jmotta31.shinyapps.io/NBA_Prediction_Tool/
• Predictions table is driven by a reactivePoll keyed on path + mtime + size so the UI refreshes when a new-day parquet lands or the current file is rewritten.
• Metrics use reactiveFileReader and refresh daily.
• No redeploy needed; app reads the freshest files from disk.

---

Disclaimers

• This project is for portfolio and educational purposes.
• The predictions and any betting-related outputs are not financial advice.
• Use responsibly and in accordance with local laws.

---