Bain Luck
Probability, not betting. The world's honest guess about what happens next.
Bain Luck aggregates prediction markets, sportsbook lines, and statistical models into one blended probability per question. You always see a probability like "60%", never a price like "-150" or "+3000". It started with sports and now covers politics, economics, entertainment, tech, geopolitics, and weather.
Live Site · Discover Feed · Calibration · API Docs
Why
At the 2026 Australian Open, Carlos Alcaraz beat Alexander Zverev in a five-set tennis semifinal. So why did his odds run from 98% up two sets to 14%, the brink of elimination, before he won?
That swing (an adductor injury, a near-collapse, a fifth-set comeback) is the story of the match, and it's invisible in the final score. Bain Luck gives every question one honest number and shows you how that number moved.
It isn't just sports. The same probability lines tracked who would stand in Taylor Swift's wedding party ahead of her July 2026 wedding, how many rate cuts the Fed delivers this year, and which movie tops Netflix this week. If the world is guessing about it, there's a number, and the way the number moves is the story.
What makes it different
- One number per question. Sources are blended into a single probability with a weighted median (staleness-decayed, share-capped) — not an average, because one heavy source must not be able to straddle the midpoint alone. Weights are hand-set priors today; a fitted-skill harness is specified in docs/aggregation-weighting-methodology.md and will only replace them if it beats them out-of-time. See
aggregation.py. When sources disagree, that's a data problem to fix, not a feature to display.
- Public calibration. The calibration page shows how often our probabilities were right, broken out by category, over every graded outcome back to 2021 —
curl -s https://api.bainluck.com/api/calibration | jq .total_outcomes for the current count. Each outcome is graded against a registered settlement authority. It's the scoreboard we hold ourselves to.
- No gambling formats. Prices like -140 or +3000 never appear anywhere. Every number is a probability.
- Honest charts. No smoothing and a fixed 0-100 axis. Real movement is the product.
Features
- Discover feed: ranked, story-grouped prediction cards across every category, with Today's Challenge
- Probability-first event pages: blended win probability charts, market maps, player props, championship paths
- Cross-source aggregation: Kalshi, Polymarket, ESPN, DataGolf, MLB Stats API, and a sportsbook consensus (12+ books via The Odds API), blended into one probability per question
- Championship grids: visual probability grids for NBA, NHL, MLB, NFL, and Golf
- Multi-platform: web (Next.js), iOS and macOS (shared SwiftUI), Apple Watch app (ships; the complication is not yet wired)
- Real-time: live game updates, ESPN win probability, score tracking
- Tested: thousands of backend (pytest), frontend (Jest), and native (XCTest) tests, run on every push. Exact counts move every day, so they aren't pinned here —
cd backend && pytest --collect-only -q | tail -1 is the current one
Under the hood
- Stories, not markets. The feed ranks stories: a golf tournament is one story with forty markets inside it, not forty competing cards. Ranking the story is what keeps a just-ended World Cup above a routine Tuesday slate.
- Blending, and the honest state of it. Sources combine by weighted median with staleness decay and a per-source weight cap. The weights themselves (
betting 3.0, espn 1.5, models 1.0, markets 0.8) are hand-set priors with no recorded empirical basis — that is the current truth, not the goal. docs/aggregation-weighting-methodology.md specifies how they get replaced by weights earned from measured per-category skill; that harness is staged, not shipped, and promoting it is gated on beating the hand-set weights out-of-time.
- Calibration as infrastructure. Graded outcomes are scored against a registered settlement authority, with provenance tiers and versioned populations, so the public accuracy page can survive a skeptical audit. The counts are deliberately not repeated here — the endpoint reports them (
total_outcomes, total_markets, total_winners on GET /api/calibration), and a number copied into a README is a number that stops matching the meter.
- Eval-gated changes. A fenced eval workshop turns hard-won definitions (what counts as "stale," what counts as "settled") into versioned fixture corpora that gate CI, and adversarial audit passes review the riskiest shipped diffs.
Tech Stack
| Component | Technology | Hosting |
|---|
| Backend API | FastAPI (Python 3.11+) | Heroku |
| Database | PostgreSQL | Heroku Postgres |
| Task Queue | Celery + Redis | Heroku Redis |
| Frontend | Next.js 14 (React) | Vercel |
| iOS/macOS | SwiftUI (shared codebase) | TestFlight |
| watchOS | SwiftUI (app ships; complication unwired) | Direct install |
Data Sources: The Odds API, Kalshi, Polymarket, ESPN, DataGolf, MLB Stats API
How it's built
This repo is developed by one person orchestrating multiple AI agent lanes running in parallel: a backend lane, a frontend/iOS lane, a read-only production verification lane, and a fenced adversarial audit lane, all coordinated through file-based work queues, with automated sentinels watching production and filing issues. Product judgment stays human. Taste, ranking, and calibration calls are made by the maintainer and recorded as standing rulings in docs/PRODUCT-BRAIN.md. The GitHub Issues board is the single source of priority.
Project Structure
bainluck/
├── backend/
│ ├── app/
│ │ ├── main.py # FastAPI entry point
│ │ ├── models/ # SQLAlchemy models
│ │ ├── routes/ # API endpoints
│ │ ├── services/ # External API clients
│ │ ├── tasks/ # Celery tasks (incl. sentinels)
│ │ └── utils/ # Pure logic
│ ├── alembic/ # Database migrations
│ └── tests/ # Backend test suite
├── frontend/
│ ├── app/ # Next.js app router
│ ├── components/ # React components
│ └── lib/ # API client, types
├── ios/Bain Luck/ # iOS + macOS + watchOS (SwiftUI)
├── scripts/ # Operational scripts
└── docs/ # Documentation
Documentation
API
The API is documented at api.bainluck.com/docs (Swagger UI).
Key endpoints:
GET /api/events lists events with probabilitiesGET /api/events/{id} returns event detail with full historyGET /api/feed serves the Discover feed (events + futures)GET /api/events/typeahead powers search suggestionsGET /api/calibration returns calibration data by category
License
MIT
