AURA is an institutional-grade variance mispricing engine that shifts the focus from fragile directional price prediction to the mathematical extraction of volatility alpha.

By utilizing Amazon’s Chronos-Bolt zero-shot foundation model on an NVIDIA H100 (during the hackathon), the system executes batched tensor inference across the S&P 100 to identify spreads between forecasted variance and live market Implied Volatility (IV).

• Batched GPU Inference: Stacks 100+ tickers into a single PyTorch tensor for simultaneous parallel forecasting.
• Vega-Weighted Alpha: Ranks opportunities by actual dollar-payout potential ($Edge = Spread_{adj} \times \nu$) rather than raw percentage spreads.
• Microstructure Realism: Includes dynamic risk-free rate scaling, 21-day expiry alignment, liquidity gating (OI > 500), and bid-ask spread penalties.
• Agentic Routing: Uses Gemini 3.1 Pro in a deterministic, zero-temperature mode to generate machine-readable JSON trade tickets.

We extract annualized volatility ($\sigma$) from the foundation model's 10th and 90th percentile quantiles using a log-return standard deviation proxy:

$$\sigma_{predicted} = \sqrt{\frac{252}{15}} \times \frac{\ln(q_{90} / q_{10})}{2.56}$$

Greeks are calculated via a closed-form Black-Scholes implementation:

$$d_1 = \frac{\ln(S/K) + (r + \frac{\sigma^2}{2})T}{\sigma\sqrt{T}}$$ $$\nu = S \cdot \phi(d_1) \cdot \sqrt{T}$$

• Model: amazon/chronos-bolt-base
• Agent: google/gemini-3.1-pro-preview
• Compute: NVIDIA H100
• Frontend: Gradio (Institutional Monochrome)

1. Clone the repository:

   git clone [https://github.com/yourusername/AURA.git](https://github.com/yourusername/AURA.git)

2. Install dependencies:

   pip install -r requirements.txt

3. Run the engine:

   python app.py

This project is an educational proxy for variance trading. Options Greeks are approximated via Black-Scholes (European-style). Real-world U.S. equities use American-style options. Use for live trading at your own risk.