Explaining OECD Fertility Divergence: Clustering and Machine Learning Insights

This study investigates fertility divergence among 33 OECD countries from 2014 to 2023 using a two-step, data-driven framework. First, dynamic-time-warped K-Means and tsfresh-HDBSCAN clustering identify six distinct fertility trajectory types, from “high-welfare stability” to “ultra-low decline.” Second, Gradient Boosting Machines, Mixed-Effects Random Forests, and sequence-to-one LSTMs predict annual fertility using seven variables, including childcare spending, parental leave, urbanization, and ART access. Explainable AI tools—TreeSHAP and partial dependence plots—reveal critical thresholds: fertility rises only when childcare spending exceeds 0.8% of GDP and ART access surpasses an index of 0.55. However, these effects diminish above 68% urbanization due to housing-cost pressure. Notably, identical policies yield contrasting impacts across clusters, challenging one-size-fits-all approaches. Korea’s ultra-low cluster, for instance, shows limited returns without addressing housing affordability and ART coverage. The findings underscore the need for integrated, cluster-specific policy packages combining childcare, housing, and reproductive support to reverse fertility decline. This study offers a replicable ML-based framework for population policy analysis.