Year: 2026
Type: article
Abstract: In practice, statistical learning models are faced with multiple uncertainties, including data noise, outliers and distributional shift, which cause large estimation bias in traditional risk measurement models or even cause decision failure. Based on distributionally robust optimization, this paper constructs a statistical learning framework. An uncertainty set based on the empirical distribution is constructed with Wasserstein distance, and then the unknown distribution shift is transferred into a computable mathematical constraint. On this basis, a robust conditional value of risk is constructed as the optimization objective, which reflects the tail loss under the worst-case distribution and improves model robustness against outliers. Finally, duality theory is used to transfer the original minimax problem into an equivalent convex optimization form, and then the stochastic mirror projection algorithm is designed to solve the model parameters effectively. The experimental results based on artificially simulated datasets show that the parameter estimation error of the model output parameters can be reduced to 0.122 when the outlier ratio is 20%, and the cumulative final value of the decision made according to the model output can reach 0.92 in extreme cases. On real financial datasets, the risk prediction calibration error of the model output at a 0.90 confidence level can be reduced to 0.094, which proves that the model is effective in a complex market environment. The robust risk measurement model constructed in this paper can achieve stable output and accurate quantification of risk estimation in the face of uncertain conditions, and provides theoretical support and computational tools for statistical learning in high-risk situations. Not applicable.
Source: Discover Computing
Author: Yunfei Cheng
Institution: University of Glasgow
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $1,520
OpenAlex ID: W7222713629