Prizm

A full-stack NBA player-prop analytics engine. Prizm pulls live prop lines from sportsbooks every day, scores each one with an empirically-calibrated confidence model, grades its own past predictions against real box scores, and surfaces the highest-conviction picks and parlays through a Next.js dashboard.

Live: prizmproject.vercel.app

What it does

Every game day, on an automated schedule:

1. Ingests the day's player props (points, rebounds, assists, threes, PRA, steals, blocks) from The Odds API, and generates synthetic alternate lines around each book line to widen the search space.
2. Enriches every prop with a confidence score from a weighted model that blends 11 factors — matchup defense, home/away splits, recent-form trend, line value vs. the player's own distribution, pace, rest, blowout risk, injury news, and head-to-head history — plus empirical adjustments (per-player over/under bias, opponent stat leaks, minutes-uncertainty penalties).
3. Labels each prop LOCK / PLAY / LEAN against stat-specific thresholds tuned on historical hit rates.
4. Builds parlays — correlated, cross-game tickets (VALUE / PREMIUM / JACKPOT) with strict per-team and per-game caps.
5. Grades itself — after games finish, every scored prop is checked against ESPN box scores (hit / miss / DNP), and daily hit rates are tracked by confidence label so the model's real-world accuracy is measurable.

The result is a self-correcting loop: predictions are logged, graded, and fed back into per-player bias and opponent-leak tables that sharpen the next day's scores.

The confidence model

The engine (lib/confidence.ts) is a transparent, deterministic weighted model — not a black box. Each stat type gets its own weight set, and scores are built from:

• Matchup edge — opponent defensive rank for the specific stat
• Home/away & recent-form — the strongest signals once full-season data is present
• Line value — z-score of tonight's line vs. the player's rolling distribution
• Pace, rest, blowout risk, injury news, head-to-head — contextual modifiers
• Empirical corrections — over-bias adjustment (overs historically hit ~43% vs. unders ~50%), minutes-uncertainty penalties for bench players, and line/odds-movement tracking vs. the morning snapshot

Thresholds and weights are backtested against a library of historical prop lines and graded outcomes, not hand-waved.

Architecture

The Odds API ──> /api/props ──> Supabase (props + synthetic alts + snapshot)
                                      │
ESPN box scores ──> /api/gamelogs ────┤
                                      ▼
                              /api/enrich  ──>  confidence model v6  ──> scored props
                                      │
                              /api/feed (parlays)        /api/grade ──> hit/miss/DNP
                                      │                        │
                                      ▼                        ▼
                          Next.js dashboard          /api/results (daily hit rates)
                                                     /api/player-bias, /api/opponent-leaks

Stack: Next.js 16 (App Router) · React 19 · TypeScript · Supabase (PostgreSQL) · Tailwind + shadcn/ui · Recharts · Vercel (20 scheduled cron jobs) · Python (offline backtesting & weight optimization)

Key directories:

• app/ — dashboard pages: props board, feed (parlays), edge finder, performance, trends
• app/api/ — the ingest → enrich → grade pipeline, each route triggered by a Vercel cron
• lib/confidence.ts — the confidence engine
• lib/odds-api.ts / lib/espn-gamelogs.ts — live data adapters
• scripts/ — Python backtesting, calibration, weight optimization, and the Kalshi edge finder
• scripts/kalshi_edges/ — bridges Prizm's model probabilities to Kalshi NBA markets to find priced edges

How it's graded

Prizm keeps an honest scoreboard. Every prediction is snapshotted before tip-off and graded after the final whistle:

• prop_grades — individual results (hit / miss / DNP)
• prop_results — aggregated daily hit rates by confidence label
• player_line_bias / opponent_stat_leaks — systematic biases mined from the full game-log history and reapplied as score adjustments

This makes the model falsifiable — you can see whether a LOCK actually locks.

Running locally

npm install
cp .env.example .env.local   # add Supabase + The Odds API keys
npm run dev                  # http://localhost:3000
npm test                     # vitest

The scheduled pipeline runs on Vercel cron (see vercel.json); locally you can hit the /api/* routes directly to seed, enrich, and grade.

License

MIT