Learning evolving populations from multi-year data: A self-contained model for short-term synthetic population forecasting

Abstract

Whereas synthetic populations constitute a fundamental component of agent-based simulation models, synthetic population forecasting seeks to generate such populations at future time horizons, enabling the simulation of hypothetical scenarios with representative agents. Existing approaches typically rely on static, single-year snapshots augmented with external data sources. In this work, we propose a novel framework for short-term forecasting that captures temporal trends directly from multi-year data, eliminating the need for external inputs. Our approach combines a projected Bayesian Network with a projective Iterative Proportional Fitting procedure. For each component, we perform attribute-level hyperparameter selection to differentiate stable relationships from those exhibiting meaningful temporal dynamics. Experimental results across diverse scenarios including varying training sample sizes, forecast horizons (one to five years), and attribute configurations, demonstrate that the proposed method yields improved distributional accuracy compared to static models, while maintaining comparable levels of realism. Additionally, we introduce a guardrailbased validation mechanism to detect and mitigate erroneous extrapolations.