Systemic Stress-Testing Framework

Conventional stress testing asks what a portfolio loses in a scenario someone thought to write down, which means it is blind to precisely the scenarios nobody anticipated. This project builds a multi-asset stress-testing engine that works in both directions: forward, by propagating shocks through a dependence structure that tightens in the tail, and backward, by solving for the smallest, most plausible set of moves that would breach a given loss threshold. CoVaR sits at the centre, quantifying how much of the portfolio's tail risk is conditional on distress elsewhere rather than intrinsic to it.

• Risk factor mapping. Decomposition of a multi-asset portfolio into equity, rates, credit, FX and commodity factors, with instrument-level revaluation rather than a linear delta approximation, so that convexity survives large shocks.
• Tail dependence modelling. Calibration of a copula structure that allows correlations to rise in the tail, replacing the Gaussian assumption under which diversification appears to hold exactly when it does not.
• CoVaR estimation. Quantile regression of portfolio returns on the distress state of each systemic driver, yielding Delta-CoVaR as the marginal contribution of that driver to the portfolio's conditional tail loss.
• Forward scenario engine. Historical replay of named crisis episodes alongside hypothetical and regulatory-style scenarios, all expressed as shocks to the same factor set for comparability.
• Reverse stress testing. Constrained optimisation over the factor space for the most probable shock vector that produces a target loss, penalised by Mahalanobis distance so that the answer is a plausible scenario rather than an arbitrary corner of the space.
• Performance layer. Full revaluation across scenarios is expensive, so the inner pricing loop is implemented in C++ and exposed to the Python orchestration and reporting layer.

• Reverse stress testing surfaces vulnerabilities that no scenario library contains, because it is not constrained by the imagination of whoever wrote the scenarios.
• A large share of measured tail risk is conditional rather than standalone: Delta-CoVaR and standalone VaR rank the same positions very differently, which changes where a hedge should be placed.
• The choice of dependence structure dominates the choice of shock magnitude. Getting the copula wrong matters more than getting the scenario severity wrong.