How HedgeVision works — end to end.

Perplexity is chosen for real-time web access. The narrative contains current market conditions, recent news, and relevant context that grounds the debate.

Three simultaneous LLM calls in parallel — Bull, Bear, and Neutral. dissent_score is computed from the disagreement. 0.0 means unanimous consensus; 1.0 means maximum disagreement across agents.

Triggered when dissent > 0.7. Maximum 3 loops. The ResearchService performs a focused deep dive on the specific hypothesis causing disagreement, then injects findings back into the debate.

Kelly Criterion computes optimal position size. Four constraints are applied: correlation check, volatility targeting, sector limits, and ATR-based stops. Any constraint breach rejects the allocation.

An adversarial LLM (GPT-3.5-turbo chosen for instruction-following, not creativity). Its explicit job is to find every reason NOT to take the trade. kill_reason is a human-readable explanation.

The 5% price check compares LLM-cited prices against live market data. Hallucination detection catches contradictions with TimescaleDB records. Any failure is an unconditional veto.

Generates the IC-level deal memo with 10 sections. Includes debate summary, risk metrics, hedge recommendations. Routes through DecisionGateway for approval tier determination.

Runs before LLM deliberation (fast pre-check). Checks order book imbalance and wall detection. A hostile order wall can veto before any expensive model computation.

95th-percentile Conditional Value-at-Risk. Vetoes if the proposed position would breach portfolio risk budget or per-asset limits (5% max).

Kelly Criterion sizing with half-Kelly in production. Four constraints: correlation cap, sector limits, volatility targeting, ATR stops. Each can independently reject.

Monte Carlo simulation with jump-diffusion model. Not just historical stress — generates thousands of synthetic scenarios including fat-tailed jumps and crisis correlation spikes.

Applies four calibrated historical shocks: 2008 Financial Crisis, 2020 COVID Crash, 1987 Black Monday, 2010 Flash Crash. Position must survive all four.

Recommends protective instruments using Black-Scholes pricing. Advisory only — does not veto. Outputs are attached to the deal memo for human review.

A real adversarial LLM call, not a rules lookup. GPT-3.5-turbo is prompted to find every reason the trade should NOT be taken. Can veto with kill_reason.

Detects LLM hallucination by comparing cited prices to live data. A > 5% deviation is an unconditional veto. Also checks liquidity and structural consistency.

TimescaleDB

Hypertable with 7-day chunk intervals. Compression after 7 days. 3-year retention. Continuous aggregates for hourly/daily/weekly rollups. 35,000+ chunks under management.

Redis

1GB LRU cache with allkeys-lru eviction. Key patterns: market:{ticker}:latest (1hr TTL), sentiment:{ticker} (1hr TTL), whale:{ticker} (2hr TTL), regime:current (30min TTL). Pub/sub for cross-module events.

NAV Arbitrage Execution

For cross-broker arbitrage (e.g., SPY ETF vs ES futures, or BTC spot vs BTCUSDT perp):

• Entry: asyncio.gather(leg1, leg2) for simultaneous execution
• Entry trigger: |Z-score| > 2.0
• Exit trigger: |Z-score| < 0.3 (mean reversion)

Exchange Mapper

Internal tickers are translated to broker-specific symbols. BTC-USD becomes BTC/USDT on Binance. AAPL becomes AAPL STK SMART on IBKR. EURUSD=X becomes EURUSD on MT4.