Adaptive Conditional Forest Sampling for Spectral Risk Optimisation under Decision-Dependent Uncertainty

Researcher Marcell T. Kurbucz has introduced Adaptive Conditional Forest Sampling (ACFS), a novel four-phase simulation-optimization framework designed to tackle the challenging problem of spectral risk minimization when uncertainty distributions are decision-dependent. This occurs in real-world scenarios where choices (like investment allocations or operational parameters) actually change the probability distributions of outcomes, making traditional optimization methods unreliable. ACFS combines Generalised Random Forests for distribution approximation, Cross-Entropy Method (CEM) guidance for global exploration, rank-weighted augmentation, and a sophisticated two-stage reranking process before final gradient-based refinement.

In rigorous testing across two distinct data-generating processes—a decision-dependent Student-t copula and a Gaussian copula with log-normal marginals—ACFS demonstrated superior performance. It achieved the lowest median oracle spectral risk on the second benchmark in every configuration, with performance gaps over Gaussian Process Bayesian Optimization (GP-BO) ranging from 6.0% to 20.0%. While statistically tied with GP-BO on median objective for the first benchmark, ACFS's major breakthrough was reducing cross-replication dispersion by approximately 1.8 to 1.9 times on the first benchmark and 1.7 to 2.0 times on the second. This 2x improvement in run-to-run reliability addresses a critical pain point in practical deployment, where consistent performance matters as much as peak performance.

The framework also consistently outperformed other contemporary methods including CEM-SO, SGD-CVaR, and KDE-SO across nearly all experimental settings. The paper's ablation and sensitivity analyses further validate that each component of the four-phase design contributes meaningfully to the overall robustness. By providing more reliable optimization under decision-dependent uncertainty, ACFS moves theoretical risk measures like Conditional Value-at-Risk (CVaR) closer to practical implementation in dynamic environments.