Understanding how individual lives unfold is one of the central questions of sociology. Despite extensive theorizing about factors shaping various outcomes, our ability to predict what happens in people’s lives remains limited. This lack of predictive success possibly stems from small datasets and inflexible modelling of individual life sequences and their interdependencies.

We overcome these limitations by conceptualizing life events – such as education, employment, and family milestones – as temporal sequences and train machine learning models inspired by sequence models from NLP on them. Using Dutch registry data and a vocabulary where each token encodes a life event attribute, we pre-train transformer models with objectives analogous to Masked Language Modelling.

Our sequences integrate individual life trajectories, but also information from population-scale opportunity networks. Learning from these data, our model captures the broad regularities in life-course dynamics across 18 million individuals in the Dutch population.

We assess our model’s capability to predict outcomes such as fertility, marriage and divorce, and income. Across multiple tasks, we find performance comparable to best-performing task-specific models. Our results show that foundation models trained on life-course sequences enriched with network data can provide a unifying framework for forecasting diverse demographic, economic, and health outcomes.

please note that this is an internal talk. If you are interested in participating in person, please write an email to