Importance sampling for portfolio credit risk
Reduces Monte Carlo variance for rare credit losses using a likelihood-ratio correction.
How it works
Importance sampling draws from a proposal distribution that favours rare losses and reweights each path. In a factor copula one can shift the factor and tilt conditionally independent default probabilities. The likelihood ratio corrects both changes when used together and the proposal must cover all relevant target states.
Worked example
Suppose a loss event has target probability 1% and proposal probability 20%. Each proposal event contributes 0.01/0.20 = 0.05 to an estimate of probability. A fifth of draws contribute 0.05 in expectation, recovering 1%. In a heterogeneous portfolio weights vary by state and can be extremely uneven.
Validation and limitations
Use exact or quadrature benchmarks for simpler books, and compare the weighted variance and effective precision for the intended payoff. A high event rate can coexist with a very noisy estimate. The credit tranche report quantifies this failure under weak factor dependence.