Estimation of Spectral Risk Measures

Ajay Kumar Pandey, Prashanth L. A., Sanjay P. Bhat

Introduction

In high-stakes domains such as financial portfolio management and autonomous robotics, quantifying uncertainty is critical for safety and performance. Traditional metrics like Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR) have become industry standards but possess inherent limitations. VaR is not a "coherent" risk measure because it fails to satisfy sub-additivity—meaning it does not always reward diversification. While CVaR is coherent, it treats all losses in the "tail" (the worst-case outcomes) with equal weight, which may not align with human risk perception or specific engineering requirements.

Spectral Risk Measures (SRMs) offer a more powerful and flexible generalization. By allowing users to define a "risk-aversion function," SRMs can assign progressively higher weights to increasingly severe losses. Despite their utility, estimating SRMs from finite data samples is mathematically complex. Prior to this work, the field lacked tight, two-sided concentration bounds—guarantees that specify the number of samples ($n$) required to ensure an estimate is within a desired distance ($\epsilon$) of the true risk value.

The SRM-Trapz Estimator

This paper introduces SRM-Trapz, a novel estimation scheme based on numerical integration. Unlike simple empirical averages, SRM-Trapz utilizes the trapezoidal rule to integrate the product of the risk-aversion function and the quantiles of the data, estimated via the Empirical Distribution Function (EDF).

The authors address two primary scenarios:

  1. Bounded Support: For distributions where losses are naturally capped, the authors prove that estimation error concentrates exponentially.
  2. Unbounded Support: For "light-tailed" distributions (such as Gaussian or Exponential), they propose a truncation-based approach. By ignoring extreme outliers beyond a threshold that scales with the sample size, they derive the first tight concentration bounds for these common distributions, improving previous "power-law" decay guarantees to superior exponential decay.

Experimental Setup

To validate the estimator, the authors conducted tests on synthetic datasets (Gaussian, Exponential, and Uniform) and a complex real-world simulation. The simulation utilized the SUMO (Simulation of Urban MObility) platform to model stochastic traffic delays in Chennai, India.

SUMO simulation map of IIT Madras

The simulation environment, based on OpenStreetMap data of the IIT Madras area, involved 7,000 cars and over 500 buses. This setup provided a realistic testbed for evaluating risk-averse routing, where the goal is to find paths that avoid catastrophic delays rather than just minimizing average travel time.

Results and Discussion

The SRM-Trapz estimator demonstrated high accuracy across all tested distributions. As shown in the convergence analysis, the absolute error between the true SRM and the estimate decreases consistently as the sample size $n$ increases.

Error convergence of SRM-Trapz

In the traffic routing experiment, the estimator was integrated into a "Successive Rejects" (SR) multi-armed bandit framework. A key insight was that the "safest" path is not necessarily the "fastest." While Route 1 had the lowest expected (average) delay, the SRM-based approach correctly identified Route 2 as the optimal choice for a risk-averse traveler. Route 2 avoided the extreme "nightmare" traffic jams that occasionally plagued Route 1. The modified SR algorithm identified this lowest-risk route with 91% probability using only 1,000 samples per route.

Conclusion

This work provides a robust mathematical foundation for using Spectral Risk Measures in practical applications. By providing explicit constants in their concentration bounds, the authors enable engineers to calculate exact sample requirements for high-confidence risk estimation. Future research may look toward establishing matching lower bounds and extending these techniques to heavy-tailed distributions where extreme events are even more frequent.

Summary updated

Abstract

We consider the problem of estimating a spectral risk measure (SRM) from i.i.d. samples, and propose a novel method that is based on numerical integration. We show that our SRM estimate concentrates exponentially, when the underlying distribution has bounded support. Further, we also consider the case when the underlying distribution is either Gaussian or exponential, and derive a concentration bound for our estimation scheme. We validate the theoretical findings on a synthetic setup, and in a vehicular traffic routing application.
BibTeX
@inproceedings{pandey2021estimation,
  author = {Ajay Kumar Pandey and Prashanth L. A. and Sanjay P. Bhat},
  title = {{Estimation of Spectral Risk Measures}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i13.17444},
}
Figure