Market Shocks - Risk-Averse Portfolio Optimization

Overview

Built a calibration-centered portfolio risk model that replaces a static Gaussian scenario generator with a conditional Neural Spline Flow trained on ten years of S&P 500 daily returns. The model uses VIX as a market-fear context signal and supports Mean-CVaR optimization for a 30-stock portfolio.

Outcome

• Diagnosed why a baseline flow still failed in crisis regimes: collapsed correlations, under-estimated volatility, and too-narrow predictive intervals.
• Added a PCA projection and rolling pairwise-correlation context channel to restore crisis calibration.
• Showed that calibration, not headline Sharpe, was the real improvement for downstream risk-aware optimization.