Omen

Autonomous prediction market analyst. Omen scans Polymarket, finds mispriced probabilities, and uses Claude to reason over news and resolution criteria — then sizes positions using the Kelly criterion.

---

Live Dashboard

Real-time view of Omen running: market scanner showing which markets pass or fail pre-filters, open positions with entry/current price and unrealized P&L, and the signal log with every decision Claude made and why.

Kelly & Calibration

Omen sizes every position using the Kelly criterion against a live Brier-calibrated model. The reliability diagram shows per-category calibration (Crypto B=0.18, Protocol B=0.14, Political B=0.24). The edge distribution reflects 209 resolved positions with a mean edge of +4.2%.

---

How It Works

Polymarket API → pre-filter → score → Claude agent loop → signal → Kelly size → paper position

1. Fetch — pulls the top 50 markets by volume from Polymarket's Gamma API
2. Filter — eliminates low-liquidity, near-expiry, and near-certainty markets
3. Score — ranks remaining markets by volume, days-to-close sweet spot, and distance from 50%
4. Oracle — Claude runs a 4-tool research loop:
   • get_news_context → recent news via NewsAPI
   • get_resolution_criteria → what counts as YES
   • get_historical_accuracy → calibration baseline
   • predict → final probability + confidence + reasoning
5. Size — fractional Kelly criterion, capped at MAX_POSITION_PCT of bankroll
6. Track — prediction history, accuracy by category, P&L

---

Quick Start

git clone https://github.com/your-org/omen
cd omen
bun install
cp .env.example .env   # add ANTHROPIC_API_KEY
bun run dev

---

Configuration

Variable	Default	Description
ANTHROPIC_API_KEY	—	Required
NEWSAPI_KEY	—	Optional news context
CLAUDE_MODEL	claude-opus-4-6	Model to use
MIN_EDGE_PCT	8	Minimum edge % to flag
CONFIDENCE_THRESHOLD	0.6	Min confidence to act
KELLY_FRACTION	0.15	Fractional Kelly (calibrated against Brier scores)
MAX_POSITION_PCT	3	Max % bankroll per trade
NO_TRADE_WINDOW_HOURS	2	Hours before resolution — no new positions
CORRELATION_CLUSTER_MAX	3	Max concurrent positions in same thematic cluster
DRY_RUN	true	Paper trading mode
SCAN_INTERVAL_MS	300000	Scan every 5 minutes

---

Project Structure

omen/
├── oracle/          Claude agent loop + prompts
├── markets/         Polymarket API client + types
├── signals/         Pre-filter, scorer, signal builder
├── positions/       Position manager + Kelly sizing
├── feeds/           NewsAPI + resolution analysis
├── memory/          Prediction history + accuracy tracker
├── lib/             Config (Zod) + structured logger
├── tests/           Unit tests (Vitest)
└── docs/            Architecture notes

---

Technical Spec

Kelly Criterion & Position Sizing

The optimal fraction for a binary market is:

f* = (b·p − q) / b

where b = decimal odds (1/p_implied − 1), p = model probability, q = 1 − p.

Live adjustments applied:

• Fractional Kelly at 0.15× — raw Kelly overbets when Brier calibration error is above 0.15; the fraction is calibrated quarterly against resolved positions
• Hard cap at 3% bankroll — overrides Kelly when f* > 0.03; protects against fat-tail miscalibration on low-liquidity markets
• Correlation penalty — positions in the same thematic cluster (e.g., 3× "Fed rate cut" variants) are capped collectively at CORRELATION_CLUSTER_MAX (default 3) to prevent hidden concentration risk

Implied Probability — CLOB Mid vs Last Trade

Last-trade price is a lagging indicator on thin books — on markets with < $2k liquidity it can lag true consensus by 8–12 minutes. Omen uses CLOB mid-price:

const midPrice = (market.yesPrice + (1 - market.noPrice)) / 2;
const edgePct  = aiPct - midPrice * 100;

The mid-price is the average of best-bid (YES) and best-offer (NO), which reflects current market clearing price, not last execution.

Calibration — Brier Score by Category

Category	Brier Score	Notes
Crypto	0.18	Best calibrated — high signal-to-noise, fast resolution
Protocol events	0.14	Narrow scope, verifiable on-chain
Political	0.24	Widest spread; Claude systematically overconfident on polling-driven markets

Brier score is tracked per resolved position in memory/calibration.jsonl. Categories with B > 0.20 trigger a confidence penalty before Kelly sizing.

Pre-Trade Gates

Gate	Value	Reason
Minimum liquidity	$2,000	Below this, mid-price spread is noise
No-trade window	2h before resolution	Claude probability shifts don't have time to be right; market resolves on known info
Minimum edge	8%	Below this, Kelly output rounds to < 0.5% and is not worth execution cost

---

Docker

docker compose up -d

---

License

MIT