Stress Testing

One-line definition

Asking "what would happen to the portfolio under this specific, named shock?" — as distinct from VaR/ES, which ask "what's the loss at a given statistical confidence level?"

Why PMs care

VaR/ES are calibrated to historical volatility and assume a return distribution; they systematically understate tail events that haven't happened recently (or ever, in this exact form). Stress tests let a PM reason about specific, plausible-but-rare scenarios directly, independent of any distributional assumption.

Two kinds

• Historical: replay a real past event (e.g. 2008, the 2013 taper tantrum) against today's portfolio.
• Hypothetical: construct a scenario by hand (e.g. "2s10s steepens 50bp and IG spreads widen 75bp") — this is what Scenario in src/pm/scenarios.py represents.

Implementation

This repo keeps shocks explicit and separate rather than one aggregator function — curve, IG spread, HY spread, and equity contributions are computed individually, then summed, so each is auditable on its own. See notebooks/integration/11_scenarios.ipynb and reuse key_rate_return_approximation (curve), spread_pnl (credit), and simple multiplication (equity) per leg.

Common mistakes

• picking only scenarios the portfolio is known to survive
• forgetting shocks can partially offset (e.g. falling rates cushioning widening credit spreads in a risk-off scenario)

Related

• Value at risk
• Key-rate duration
• Backtesting biases