Partner Group Uruguay

Simulation of Socio-Demographic Systems

Detailed Description

The Modeling and Simulation Partner Group is an international collaboration between the Statistics Institute (IESTA) at the University of the Republic in Uruguay and the Max Planck Society. Our work lies at the intersection of computational statistics, machine learning, and demography.

These computational tools are being developed in response to the pressing need to anticipate global challenges such as sustained fertility decline and population aging. These phenomena place critical structural pressures on the sustainability of pension schemes, health care systems, and the broader economy. To face these challenges, it is essential to understand how societies change, and to project possible future scenarios by relying on rigorous empirical evidence and transparent uncertainty quantification.

Methodologically, our agenda is organized into two main research lines.

Microsimulation and Simulation-Based Inference (SBI)

A recurrent disconnect in quantitative population science is that while the most abundant data exist at the macroscopic level in the form of aggregate rates, the causal mechanisms driving change operate at the micro level. Recovering unobservable individual behavioral parameters from purely aggregate series is a powerful way to uncover the latent micro-level drivers of population change. While this task represents a complex challenge, we address it by developing demographically interpretable individual-level simulation models of the reproductive process. By coupling these mechanistic models with advanced simulation-based inference (SBI) techniques, such as approximate Bayesian computation and sequential neural posterior estimation, this framework allows us to robustly recover the latent structure of human decision-making using only aggregate age-specific fertility rates. Consequently, we establish a statistically rigorous bridge between macro-level observations and their underlying micro-behavioral drivers, estimating core parameters governing contemporary fertility such as family size preferences, reproductive timing, and contraceptive failure. By extracting these behavioral insights from aggregate series, this approach significantly reduces the data requirements for constructing detailed microsimulation models. Furthermore, it provides the foundation for generating behaviorally explicit demographic forecasts.

Deep Learning for Time Series (Neural Forecasting)

Our group has recently established a research line that leverages rapid advances in deep learning methods to explore their potential for demographic forecasting. Projecting global population dynamics is a central predictive challenge. Historically, standard demographic methodologies have relied on models that impose structural constraints such as mean reversion or long-term asymptotes, which may fail to capture contemporary patterns like sustained fertility decline. To overcome these restrictive assumptions, we explore the use of data-driven global models based on deep learning architectures. We formulate demographic projection as a sequence-to-sequence learning problem, and employ a global cross-learning strategy to mitigate the overfitting risk common in short macroscopic time series. By simultaneously training neural networks on harmonized historical data from dozens of countries, we advance distribution-free probabilistic forecasting via multi-quantile regression. Initially focused on global fertility, this line of research is now moving into the multidimensional forecasting of mortality data.

Back to At a Glance

Research