Horizon-Aware Forecasting of Hypertension Prevalence and Mean Total Cholesterol in Chinese Men: Retrospective Leakage-Controlled Benchmark Study Comparing Statistical, Machine-Learning, and Neural Models

Abstract Background Forecasts of population-level cardiovascular risk factors can support surveillance and planning, but annual country-level health series are often short. With only 30 or 39 annual values per target, model flexibility can outpace the information available for training, making leakage control, repeated temporal evaluation, and comparison with simple baselines essential. Objective The aim of this study is to compare statistical, classical machine-learning, and neural forecasting models for age-standardized prevalence of hypertension (ASPH) and mean total cholesterol (MTC) in Chinese men and to test whether attention or substantially greater neural capacity provides consistent predictive benefit. Methods The authors analyzed 2 officially sourced annual series: ASPH for Chinese men aged 30 to 79 years (1990‐2019; 30 observations) and MTC for Chinese men aged 18 years or older (1980‐2018; 39 observations). No series was interpolated, extrapolated, imputed, smoothed, spliced, or synthetically extended. Leakage-controlled expanding-window evaluation began after 21 observations and assessed 1-, 2-, and 5-year horizons. Classical machine-learning and neural models used a fixed primary lookback of 3 (L=3), with nested lookback selection disabled; neural models used 10 fixed seeds. Performance was assessed with root mean square error (RMSE), mean absolute error, mean absolute scaled error, root mean squared scaled error, symmetric mean absolute percentage error, directional accuracy, mean error, median error, diagnostic R ², prespecified neural-ablation inference, empirical forecast-interval coverage, and computational-cost reporting. Results Statistical models had the lowest RMSE in all 6 target-horizon cells. Autoregressive integrated moving average won for ASPH at h=1 (RMSE 0.000178), h=2 (RMSE 0.000578), and h=5 (RMSE 0.004324), and for MTC at h=1 (RMSE 0.000400) and h=2 (RMSE 0.001479); error-trend-seasonal state space won for MTC at h=5 (0.006991). Classical machine-learning and primary neural models won no cells. Removing attention reduced RMSE in 5 of 6 ablation cells, while the compact architecture had lower RMSE than Original23K in 5 of 6 cells; formal differences were detected for both contrasts at ASPH h=1 and h=2, whereas ASPH h=5 (n=5) was descriptive only. Predictive intervals often undercovered. The benchmark completed all 7242 planned output-producing fits. Conclusions Simple statistical forecasting remained highly competitive in these short, retrospectively estimated annual series. Greater complexity did not ensure better point prediction, and neither attention nor larger capacity conferred a consistent advantage. Point accuracy did not ensure well-calibrated predictive uncertainty; findings were conditional on the 2 targets, target-specific horizons, and small evaluation samples.