VULCAN — Autonomous VRP Options Desk

VULCAN is an autonomous AI trading agent built for the Options Alpha Agents track. It harvests the variance risk premium: it forecasts SPY realized volatility using an ensemble of three classic models (HAR-RV, GARCH(1,1) and a Kalman filter) with HMM regime detection, then compares that forecast to the implied volatility priced in the live Alpaca option chain. When the market charges more for volatility insurance than the forecast says it is worth, VULCAN sells premium through defined-risk structures (iron fly, iron condors, credit spreads). The system runs on a 3 model brain. Model 1 is the quant forecaster. Model 2 is a pricer plus a self improving backtest loop: it generates strategy variants, tests them walk-forward on held out data, scores them with ICIR and deflated Sharpe, analyzes failures and refines until convergence. Model 3 is a multi agent LLM debate on Cloudflare Workers AI: GLM 5.3 Flash acts as Portfolio Manager, GLM 5.2 runs the analysts, and Kimi K2.7 plays an adversarial bear researcher. The AI can only veto or shrink a trade, never grow risk. Seven deterministic risk gates protect every ticket: defined risk only, VIX turbulence filter, Monte Carlo VaR on 32,768 GBM paths with antithetic variates and Sobol quasi random sampling, quarter Kelly sizing, exposure caps, a daily circuit breaker and rule based exits. Everything runs on a fresh $100,000 Alpaca paper account via the Trading API with multi leg mleg limit orders, bracket orders for the long/short equity sleeve and market-hours gating. A React + TypeScript dashboard hosted on Cloudflare shows real-time P&L, positions, the live AI debate, a run backtest button with progress bar and an interactive Monte Carlo lab. Day one result: $100,000 to $100,335.37, fully booked and flat.