Abstract Context Aging varies between individuals, with profound consequences for chronic diseases and longevity. One hypothesis to explain the diversity is a genetically regulated molecular clock that runs differently between individuals. Large human studies with long enough follow-up to test the hypothesis are rare due to practical challenges, but statistical models of aging are built as proxies for the molecular clock by comparing young and old individuals cross-sectionally. These models remain untested against longitudinal data. Objective We applied novel methodology to test if cross-sectional modeling can distinguish slow vs accelerated aging in a human population. Methods We trained a machine learning model to predict age from 153 clinical and cardiometabolic traits. The model was tested against longitudinal data from another cohort. The training data came from cross-sectional surveys of the Finnish population (n = 9708; ages 25-74 years). The validation data included 3 time points across 10 years in the Young Finns Study (YFS; n = 1009; ages 24-49 years). Predicted metabolic age in 2007 was compared against observed aging rate from the 2001 visit to the 2011 visit in the YFS dataset and correlation between predicted vs observed metabolic aging was determined. Results The cross-sectional proxy failed to predict longitudinal observations (R2 = 0.018%, P = 0.67). Conclusion The finding is unexpected under the clock hypothesis that would produce a positive correlation between predicted and observed aging. Our results are better explained by a stratified model where aging rates per se are similar in adulthood but differences in starting points explain diverging metabolic fates.