Systems Analysis of Decision-Making Under Ambiguity With Comparative Out-of-Sample Experimental Study of Geometric Dispersion Theory and Cumulative Prospect Theory

• Published in: IEEE transactions on systems, man, and cybernetics. Systems Vol. 55; no. 8; pp. 5192 - 5206
• Main Authors: Al-Najjar, Camelia, Malakooti, Behnam
• Format: Journal Article
• Language: English
• Published: IEEE 01.08.2025
• Subjects: Ambiguity; Data models; Decision making; Dispersion; Gain measurement; Loss measurement; Modeling; out-of-sample; Predictive models; Reviews; sources of ambiguity; tail-separability; Uncertainty; Vectors
• ISSN: 2168-2216; 2168-2232
• DOI: 10.1109/TSMC.2025.3564865

In this article, we develop a new descriptive model for ambiguity decision-making called Ambiguity Geometric Dispersion Theory (A-GDT). In out-of-sample predictions, we find that A-GDT is experimentally superior to all models which it generalizes; specifically, cumulative prospect theory (CPT), subjective expected utility (SEU), alpha-maxmin expected utility (<inline-formula> <tex-math notation="LaTeX">\alpha </tex-math></inline-formula>-MEU), vector expected utility (VEU), and expected utility theory (EUT). We ran an experimental study of 150 decisions under ambiguity by 310 subjects by operationalizing payoffs, probability, and ambiguity. We show that subjects exhibit behavior that contradicts many decision models, such as CPT, SEU, <inline-formula> <tex-math notation="LaTeX">\alpha </tex-math></inline-formula>-MEU, and EUT; but not A-GDT. In out-of-sample studies with the same number of parameters, the A-GDT model predicts both representative (aggregate) agent and individual preferences substantially better than other well-known ambiguity models. Specifically, a three-parameter A-GDT is substantially superior to CPT and all other important models. Furthermore, significance testing in a paired comparison of models shows that subjects' behavior matches the A-GDT model markedly better than other models. A new rate-of-degradation analysis demonstrates that A-GDT's predictions are far more stable than other models as the in-sample size is decreased. In addition, the A-GDT model not only can resolve typical Ellsberg-like behavior, but it can also resolve hypothetical and real-life ambiguity problems with multiple sources of ambiguity including the paradoxes by Machina.