In this paper, we propose an inverse validation approach for an agent-based model of a prediction market with substantively heterogeneous traders. Our prior research has used empirically derived trader types to investigate whether they can generate plausible market outcomes, we ask which features of these trader types are actually required to successfully reproduce emerging market-level patterns. We will evaluate our inverse procedure against a traditional forward approach considering price adjustments, mispricing, participation and volume patterns, as well as sensitivity to particularly wealthy traders (e.g., whales), whose high-volume activities have been shown to distort prediction market outcomes. We aim to identify which dimensions of trader heterogeneity are necessary, redundant, or replaceable to simulate prediction markets using agent-based modelling. In doing so, our research also addresses a recurring challenge in empirical agent-based modelling: how to move from plausible microfoundations to more reliable claims about which features, patterns, and distinctions actually matter most.

Agent-based models (ABMs) are increasingly used in public health research to assess interventions that operate through complex, non-linear pathways. Social network structure in ABMs determines how behaviours and interventions spread. Many Agent based public health models rely on stylised network structures that are not empirically grounded. The complete network structure cannot practically be observed, so a feasible alternative is ego-centric network sampling. We present a method for generating synthetic social network structures from ego-centric survey data that adapts stochastic actor-oriented modelling as a generative algorithm, accounting for uncertainty and suitable for embedding within population ABMs. The approach is demonstrated in a case study on smoking cessation in England, using ego-centric data from the Smoking Toolkit Study.

Existing daily activity generation models produce data that are statistically consistent at the community level. However, frequent activity switching is often observed when individual agent behaviors are examined. To address this issue and generate daily activities that reflect regional characteristics, we previously proposed a simulated annealing based model using the Survey on Time Use and Leisure Activities, Japan. However, that approach required excessive computation time. In this study, we apply combinatorial optimization with machine learning to significantly reduce generation time while maintaining statistical consistency and behavioral continuity.

This study addresses the growing need for a rigorous evaluation of Bayesian simulation-based inference methods in agent-based modeling (ABM). We conducted a simulation study to compare state-of-the-art Bayesian inference approaches, including approximate Bayesian computation, neural posterior estimation, and neural density ratio estimation, in realistic economic ABM settings. The study emphasizes the applicability of its findings to models of complex economic systems. The study design uses pseudo-empirical data from multiple ABMs with properties representative for modeling economic complexity. Factors such as model complexity, computational budget, and sample size are varied in a partial factorial design. This approach allows to evaluate estimator properties, such as bias, consistency, and coverage. The study follows the BASIS framework for simulation studies, including the ADEMP planning structure. We assess the statistical significance of the findings using Monte Carlo standard errors.

This work contributes to the establishment of clearer guidelines for the use of Bayesian simulation-based inference methods in ABM. In addition, it advances methodological transparency in the comparison of inference methods for ABM by emphasizing uncertainty reporting and neutral design principles.

In recent years, increasing attention has been dedicated to testing and validating opinion dynamics models against empirical data. At the same time, the psychometrics literature highlights that opinion data are inherently ordinal, and transforming them into numerical scales introduces significant distortions. Such distortions have been shown to substantially affect model dynamics and predictions---so much so that, in some cases, one model can be transformed into another through the application of these distortions. In this article, we introduce a procedure for standard validation of opinion dynamics which takes into account the effect of distortions. As we show, their inclusion allows for a more robust validation process and, overall, better results. Specifically, we test such a methodology on five opinion dynamics models by using both simulated and empirical data from the European Social Survey.