Belief at Risk: Quantifying Agentic AI Model Risk with LLM-Inferred Bayesian State Filters

Battle Count: 67

Highly relevant to quantitative trading through its regime-detection framework, belief-state portfolio construction, and risk measures (VaR, CVaR, drawdown). The POMDP formulation directly connects to algorithmic trading as sequential decision-making under uncertainty. The Belief-at-Risk measure provides a novel model-risk metric for AI-driven trading systems. However, the paper focuses more on governance and validation than on alpha generation or execution optimization. The illustrative portfolio (SPY with regime-dependent exposure) is a simplified demonstration rather than a production trading strategy.

Moderate complexity. The Bayesian filtering step is straightforward (matrix multiplication + element-wise update). The main complexity lies in: (1) designing effective prompts and schemas for LLM structured outputs, (2) selecting and justifying latent regimes and transition matrices, (3) implementing calibration diagnostics and entropy monitoring, and (4) integrating the pipeline with data feeds and portfolio construction. The provided Python code significantly reduces implementation burden. No deep learning training is required—only API calls to OpenAI.

Python source code is publicly available on GitHub (github.com/mfrdixon/belief-at-risk). Uses publicly accessible data from Massive.com and OpenAI API. Complete numerical-experiment parameter table is provided. However, results depend on OpenAI API responses which may vary over time, and the LLM component introduces non-determinism. The Bayesian filtering and portfolio construction code is fully reproducible given the same LLM outputs.