Predicting Demographic Contraction as an Upstream Signal for Health-System Planning: A Persistence-Corrected...

Demographic contraction changes the scale and geography of healthcare demand while weakening the fiscal and workforce base that sustains services. The general objective was to evaluate a persistence-corrected one-year-ahead demographic forecasting framework and to determine whether it can provide a defensible upstream signal for health-system planning; the primary model forecasts demographic change rather than health-system capacity itself.

The analytic sample retained 672 of 4616 documented country-year observations (14.6%) from 47 countries (2000-2020), representing a coverage-selected panel with sufficiently complete cross-domain data. Requiring consecutive observations produced a 522-row dynamic panel from 45 countries and 167 one-year-ahead forecasts for 2016-2020. DPC-GB learned deviations from the current population-growth rate under expanding-window validation. Analyses compared momentum and cross-domain feature sets, matched gradient-boosting hyperparameters for direct and correction targets, lagged structural covariates, held country groups out of training, and used country-cluster uncertainty estimates. Incident-onset probabilities were examined only as a supplementary exploratory analysis.

Among the feature specifications, momentum plus non-health predictors had the lowest observed RMSE (0.20491), followed by momentum only (0.20503), the full DPC-GB model (0.20531), and momentum plus health predictors (0.20680); persistence had RMSE 0.21920. The full model's observed RMSE was 6.34% lower than persistence, but the country-cluster bootstrap interval for the absolute difference [-0.0038, 0.0366] crossed zero (one-sided p = 0.0968), so improved forecasting across countries was not established. With identical gradient-boosting hyperparameters, direct level prediction had RMSE 0.2259 versus 0.2053 for persistence correction, although their country-cluster difference interval included zero. After Holm correction across four exploratory transfer endpoints, evidence remained for fertility forecasting; hospital-bed density was not significant after adjustment (adjusted p = 0.072). Only one incident contraction occurred among 136 eligible forecasts, precluding validation of probability calibration, discrimination, or operating thresholds.

DPC-GB is a retrospective proof-of-concept for persistence correction in demographic forecasting, with possible downstream relevance to health-system planning. These data do not establish improved point forecasting across countries. Generalizability is constrained by complete-case selection. Momentum plus non-health predictors yielded the lowest observed RMSE, but its small numerical advantage over momentum only does not establish superiority; the three-feature momentum-only specification provides a parsimonious alternative to the full 42-feature model. Incident-risk probabilities remain exploratory and cannot support validated early-warning or threshold claims.