MLB Intelligence

MLB betting intelligence system with ML-driven predictions, live execution on Kalshi, and transparent performance analytics.

Problem

Sports betting (especially MLB) has a few persistent problems:

• Too much data, too little time — schedule context, pitcher changes, weather, team form, and live odds all move quickly.
• Execution is the hard part — even good models fail if betting decisions aren’t consistent, timed correctly, and recorded.
• Post-hoc analysis is usually missing — without clean history, you can’t answer “did this strategy work?” with confidence.

Action

MLB Intelligence turns the day-to-day work into a repeatable system:

1. Ingest schedules, odds, weather, pitcher stats, and team stats
2. Engineer features (80+ game-level features)
3. Predict win probabilities (LR default; optional LightGBM/XGBoost/MLP/ensemble)
4. Compare to the market to identify edge
5. Execute on Kalshi (live or dry-run)
6. Settle + audit results and ROI in a single database

Solution

MLB Intelligence is an end-to-end stack for MLB trading:

• One orchestrator that runs the full workflow (run_pipeline.py)
• A dashboard (FastAPI) for transparent performance analytics + operator controls
• A Postgres-backed history for bets, balances, orders, and model artifacts
• A public modeling snapshot (data/master_mlb.csv) that can be updated periodically (not continuously)

What’s inside

• Dashboard — FastAPI app serving public analytics + private operator controls
• Pipeline — runs ahead of scheduled first pitch; predicts and places bets in parallel
• Model — calibrated classifiers + walk-forward validation; Kelly stake sizing
• DB — PostgreSQL source of truth for games, bets, and run history

Key components:

run_pipeline.py       # Main orchestrator
run_pipeline_v2.py    # Parallel V2 orchestrator (LightGBM; dry-run by default)
fetch/                # Data ingestion — odds, weather, stats
models/model_v1/train.py   # Model training + walk-forward evaluation
models/model_v1/predict.py # Inference — win probabilities + sizing
models/model_v2/    # V2 LightGBM predict + nightly deterministic eval (`eval.py`)
bet/place_bet.py      # Kalshi execution
settle_games.py       # Post-game settlement
dashboard/app.py      # Dashboard API
homelab.py            # SSH helper for app/db LXCs

Data model philosophy

• Postgres is the source of truth.
• Files under data/ are primarily caches/backups and may be stale.
• data/master_mlb.csv is intended as a public snapshot of modeling data and can be refreshed on a controlled cadence.

Docs

• Pipeline (end-to-end)
• LXC SSH helper

Contributing / local development

See CONTRIBUTING.md for local setup, testing, and dev commands.

Deployment

Push to GitHub → GitHub Actions runner auto-deploys to the app LXC.

Direct LXC SSH debugging via homelab.py:

python3 homelab.py app "systemctl status mlb-dashboard --no-pager -l | tail -40"
python3 homelab.py app "journalctl -u mlb-dashboard -n 100 --no-pager"

Data sources

• MLB Stats API
• The Odds API
• Kalshi
• Open-Meteo
• FanGraphs

License

BUSL-1.1 (Business Source License 1.1)

• See LICENSE
• See LICENSE-FAQ.md