The Bias-Variance Frontier of Risk Estimation: When Spectral Methods Dominate Monte Carlo

We study the bias-variance trade-off in Expected Shortfall estimation, comparing spectral (order-statistic) methods with parametric and bootstrap Monte Carlo. We prove that spectral ES is the minimum-variance unbiased estimator (MVUE) in the class of model-free risk estimators and characterize the Pareto frontier between bias and variance. Parametric Monte Carlo achieves lower variance through regularization but introduces irreducible model bias that does not diminish with sample size. We derive the crossover condition: spectral methods dominate when the variance gap between methods falls below the squared model bias, which occurs at sample sizes of approximately 500--1000 for heavy-tailed distributions. The characterization extends to all coherent, law-invariant risk measures via Kusuoka's representation theorem. All results are formally verified (889 verified declarations, 168 theorems, 0 axioms) and supported by Monte Carlo experiments on \(t(3)\) portfolios. The findings imply that regulatory backtesting at the Basel III standard window (\(n = 250\)) operates in the Monte Carlo regime, while stress testing and long-horizon risk management benefit from spectral methods. A companion paper (Nagy, 2026b) develops the structural estimation-theoretic foundations: Cramér-Rao efficiency, minimax sample complexity, a model risk diagnostic, and phase transition characterization.

Keywords: Expected Shortfall, spectral risk measures, bias-variance trade-off, MVUE, Monte Carlo, model risk, Kusuoka representation, formal verification