Autonomous analyst for Robinhood prediction markets. Omen scans event contracts, finds mispriced probabilities, and uses Claude to reason over news and resolution criteria — then sizes positions using the Kelly criterion.

Website

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.

Omen turns market setup, evidence quality, historical calibration, liquidity, and bankroll constraints into one conviction score before sizing a position. The engine keeps edge, confidence, Kelly sizing, and exit risk visible instead of treating Claude's probability as a blind trade signal.

Omen is designed around real Robinhood event-contract shapes, not invented meme prompts. Typical examples include:

• July CPI year-over-year above 3.4%
• July payrolls above 80k
• RBA holds rates on August 11

That matters because resolution logic, liquidity structure, and sizing discipline behave differently across:

• multi-outcome election markets
• range and target ladders
• date-bounded geopolitical event contracts

Robinhood event contracts → pre-filter → score → Claude agent loop → signal → Kelly size → paper position

1. Fetch — pulls active Robinhood event contracts and their latest market data
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

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

omen/
├── oracle/          Claude agent loop + prompts
├── markets/         Event-contract market 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

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

Last-trade price is a lagging indicator on thin markets. Omen uses the midpoint between current YES and NO pricing:

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.

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

docker compose up -d

MIT