LLM-Generated Counterfactual Stress Scenarios for Portfolio Risk Simulation via Hybrid Prompt-RAG Pipeline

Battle Count: 121

Highly relevant for quantitative risk management and portfolio stress testing. The paper provides a scalable framework for generating macroeconomic counterfactual scenarios that translate into portfolio VaR/CVaR multiples, directly applicable to risk committee scenario libraries, capital adequacy assessment, and tail-risk monitoring. The three-channel factor model (volatility, linear PCA, nonlinear) offers interpretable decomposition of stress transmission. However, it is positioned as a complement to existing stress-testing frameworks rather than a trading signal generator. The moderate stress levels (1.1-1.5x VaR multiples) are more relevant for regulatory/supervisory risk assessment than for short-horizon trading decisions. The reproducibility and auditability features make it suitable for institutional deployment.

Moderate-to-high complexity. The pipeline involves: (1) IMF WEO data ingestion and country profile construction, (2) MiniLM embedding and FAISS index construction, (3) structured prompt engineering with 30 variants per country, (4) LLM API calls with JSON parsing and tolerant extraction, (5) two-layer plausibility filtering (hard gates + soft scoring), (6) DeBERTa-based NLI regime classification, (7) PCA factor estimation with sign alignment, (8) linear and nonlinear beta estimation, (9) regime-specific covariance estimation, (10) Monte Carlo simulation (20,000 paths × 63 days), (11) VaR/CVaR/MDD computation, (12) ANOVA variance decomposition, (13) fairness diagnostics, and (14) comprehensive artifact hashing and manifest management. However, all components use standard libraries (FAISS, MiniLM, scikit-learn PCA, standard econometric models) and run on commodity hardware. The full factorial grid (7 countries × 4 configs × 30 prompts × 2 models) requires significant API calls but is parallelizable.

Exceptionally high reproducibility. The pipeline incorporates: (1) snapshot-frozen artifacts (IMF WEO baselines, headline CSVs, cached ETF prices, MiniLM embeddings, FAISS index, PCA factors, covariance matrices), (2) SHA-256 hash-verified run manifests linking every figure/table to immutable artifacts, (3) deterministic decoding modes with near-zero temperature, (4) fixed random seeds for PCA and Monte Carlo simulation, (5) version-locked models and retrievers, (6) explicit run metadata schema with all hashes. The pipeline is described as 'snapshot-replayable' - given frozen artifacts and recorded seeds, all scenarios and risk metrics can be regenerated up to floating-point and Monte Carlo noise. Strict bit-level determinism across hardware is not claimed but all other controls are in place.