Uncertainty in Individual-Level Longitudinal Forecasting: A Methodological Synthesis

Uncertainty is a defining challenge in forecasting individual longitudinal data, yet existing methods address only fragments of the amount of uncertainty that accumulates across the forecasting pipeline. This paper offers a methodological synthesis that delineates the full set of components relevant to uncertainty in longitudinal psychological data and examines how major forecasting uncertainty frameworks, such as Bayesian and probabilistic models, resampling methods, conformal prediction, or quantile-based approaches, capture the underlying forecasting uncertainty. Using an illustrative forecasting pipeline, we show that so far, no existing framework provides comprehensive coverage, and that their strengths are fundamentally complementary. Our findings suggest that the most promising paradigm for quantifying forecast uncertainty in such pipelines is a Bayesian approach, in which multiple sources of uncertainty can be jointly propagated and coherently integrated into a unified posterior predictive distribution. Building on this analysis, we derive principled recommendations for constructing uncertainty-aware forecasting pipelines and identify methodological gaps that require targeted innovation, including unified frameworks that propagate all major uncertainty sources, improved handling of dependence and non-stationarity, and hybrid approaches. This synthesis establishes a foundation for advancing uncertainty quantification in longitudinal forecasting and for developing next-generation methods tailored to psychological time-series.