Polymath

Bloomberg Terminal–style interface for Polymarket prediction markets — real-time data, cross-market analytics, and AI-assisted research. Built with Next.js 15 on Vercel.

Live: polymath-app.vercel.app

---

What's in it

Market browsing

• Dashboard — active events from Polymarket's Gamma API, sorted by volume, filterable by category (Politics, Sports, Crypto, Pop Culture, Business, Science).
• Market terminal (/market/[id]) — price history (1D/1W/1M), live orderbook with depth bars, top bid/ask/spread, outcome pricing, news ticker, and a per-market AI analysis panel.
• Event view (/event/[slug]) — groups all markets under a single event (e.g. "Who wins 2026 Champions League?") with shared metadata. Works from both numeric IDs and slugs.
• Keyboard-first navigation — / to search, J/K to navigate, Enter to select, P for portfolio, Esc to go back.

Cross-market analytics

• Arbitrage scanner (/arbitrage) — scans the top 80 multi-outcome events and flags those whose YES prices sum to more or less than 100¢ by ≥2%. Ranked by |arbPct| × √liquidity so tradeable edges surface above mid-priced noise.
• Correlation matrix (/correlation) — paste 2–6 market IDs, computes pairwise Pearson correlation on 1-week price history aligned by timestamp. Flags mispriced pairs where correlation ≥ 0.70 but the current price ratio has drifted ≥15% from the historical mean.
• Rolling correlation + decoupling detection — on the same page, sliding-window Pearson (1/3 of series, clamped 5–24 samples) rendered as per-pair sparklines. Pairs where the recent window has broken >0.30 below the overall correlation get a DECOUPLING badge — often a regime change or pair-trade revert opportunity.
• Resolution calendar (/resolutions) — events whose endDate falls within a 1-to-30-day horizon, bucketed into 24h / 3d / 7d / 14d / 30d bands, sorted by volume within each bucket. Shows where event risk is about to hit.
• Market movers (/movers) — largest 24h price movers, each with a one-sentence Gemini-generated "why" explanation grounded in scraped Google News headlines.

Per-market AI analysis

Two independent analysis modes plus a digest:

Mode	What it does
Technical Analysis	Gemini reads price history + orderbook, outputs direction (YES/NO/NEUTRAL), confidence, and detected signals (SMA crossovers, RSI, momentum, book imbalance).
Math Prediction	500-path Monte Carlo via Geometric Brownian Motion (Box-Muller), scenario analysis from −30¢ to +30¢, and a payoff curve for the user's position.
News Signal	Pulls recent headlines from Google News, asks Gemini to estimate the implied probability, compares to market price, and labels OVERPRICED / UNDERPRICED / FAIRLY_PRICED with per-headline sentiment. 10-min in-memory cache.
News Digest · 30d	Google News over a 30-day window, synthesized by Gemini into: narrative summary, net direction, weighted key themes, chronological timeline of significant headlines, and forward-looking catalysts. 30-min cache.

Trading tools on the market page

• Slippage calculator — takes a $ order size, walks the live orderbook, returns avg fill price, shares acquired, levels consumed, and slippage in bps. Warns on book exhaustion or ≥200bps slippage.
• Position builder — YES/NO, quantity, entry price → payoff curve and P&L math.
• Time-decay visualizer — theta-like decay curve for price-in-the-money positions as resolution approaches.

Portfolio

• Add YES/NO positions with custom quantity and entry price.
• Live price refresh on mount (can also be manually refreshed).
• Aggregate payoff curve across all positions.
• Real risk metrics — parametric Value at Risk for binary outcomes: VaR₉₅ = 1.645 × √(Σ qᵢ² · pᵢ · (1 − pᵢ)), plus net exposure and per-position P&L using the refreshed price.
• Close selected positions.
• Persisted in localStorage via Zustand.

LLM routing

Primary provider is Gemini 2.5 (gemini-2.5-flash-lite first, then gemini-2.5-flash on 429/503). On full Gemini failure, it falls back to Groq + Llama 3.3 70B. Timeouts (15s) and cache on every LLM-driven endpoint to protect free-tier quota.

---

Tech

• Framework: Next.js 15 (App Router, Turbopack, Fluid Compute on Vercel)
• Language: TypeScript (strict)
• Styling: Tailwind CSS + custom terminal/CRT theme
• State: Zustand with persist (localStorage)
• Charts: hand-rolled SVG + Recharts
• UI primitives: Radix UI + lucide-react
• Market data: Polymarket Gamma API + CLOB API (public, no auth)
• LLMs: Google Gemini 2.5 (primary) · Groq Llama 3.3 70B (fallback)
• News: Google News RSS (no key)
• Deployment: Vercel

No backend service — every route is a Next.js server function. The Polymarket APIs and LLM providers are called directly from server code.

---

Setup

npm install
cp .env.example .env.local   # fill in keys
npm run dev

.env.local:

GEMINI_API_KEY=...      # required for AI features; get one at aistudio.google.com/apikey
GROQ_API_KEY=...        # optional fallback; get at console.groq.com/keys

All market-data features work without keys — AI-driven features degrade gracefully (show raw headlines or disabled analysis) when both are absent.

---

API routes

Route	Purpose
GET /api/events	List active events (paginated, volume-sorted)
GET /api/events/[id]	Event details — accepts numeric ID or slug
GET /api/markets	List active markets
GET /api/markets/[id]	Market + orderbook + price history
POST /api/analysis	Technical analysis + Monte Carlo
POST /api/news-signal	Per-market news-vs-price gap signal
POST /api/news-digest	30-day narrative digest
POST /api/correlation	Pearson matrix + mispriced pairs + rolling series
GET /api/arbitrage	Cross-market arbitrage scanner
GET /api/resolutions?days=N	Events resolving within N days
GET /api/movers	Top 24h movers with AI-generated "why"
GET /api/news	Google News RSS for a query
POST /api/news/summarize	Summarize an article
GET /api/price-events	Significant historical price-event detection

---

Math reference

Monte Carlo (GBM)

S_{t+1} = S_t · (1 + μΔt + σ√Δt · Z)
Z  ~ N(0, 1)     via Box-Muller
μ  = annualized mean log-return from price history
σ  = annualized daily volatility
S  clamped to [0.01, 0.99]   (binary-outcome boundary)
500 paths, percentiles at 5/25/50/75/95

Portfolio VaR (parametric, binary)

VaR₉₅ = 1.645 · √( Σ qᵢ² · pᵢ · (1 − pᵢ) )

Where qᵢ is position size and pᵢ is current market price. Treats each market as an independent Bernoulli outcome. Real positions are not fully independent, but this is a defensible baseline for a multi-market book.

Pearson correlation

ρ(X,Y) = Σ(xᵢ − x̄)(yᵢ − ȳ) / √( Σ(xᵢ − x̄)² · Σ(yᵢ − ȳ)² )

Series aligned by timestamp first; falls back to positional alignment if <3 common timestamps.

Arbitrage detection

For any event with ≥3 mutually-exclusive YES markets:
  sum = Σ priceᵢ
  if |sum − 1.00| ≥ 0.02:  flag
  rank by |arb| · √liquidity

---

Keyboard shortcuts

Key	Action
/	Open search
J / K	Navigate list down/up
Enter	Select / open
P	Portfolio
Esc	Back / close modal

---

What this is not

• Not a trading client. Positions are local-only. No wallet integration, no order submission.
• Not financial advice. Backtests and AI outputs are exploratory tooling, not signals to act on.
• Not fully real-time. Prices are polled, not streamed; orderbook updates on page load.

License

MIT.