Phenology exercises a critical control on annual carbon uptake by terrestrial ecosystems. Autumn phenology, while less studied relative to spring phenology, may also constrain annual net ecosystem productivity (NEP). Using 17-year (1992–2008) records of C flux phenology (CFP) derived from continuous eddy covariance (EC) measurements at the Harvard Forest (HF), here we show that the autumn phenology played a more significant role than the spring phenology in controlling annual NEP. We found that the onset of carbon uptake (CU) in spring only explained 39% of annual NEP, compared to 66% of end of CU in autumn. Though neither onset nor end of gross primary productivity (GPP) was correlated with annual NEP, the autumn lag, i.e., the time lag between ends of GPP and CU, was found to have a particularly high potential in explaining annual NEP (R2 = 0.82, p < 0.001). We further showed that the autumn lag can be modeled as a function of entirely autumn (September–November) meteorological variables, including the water vapor pressure deficit, global shortwave radiation and the surface soil temperature, indicating the autumn lag and consequently the annual NEP can be modeled in areas lacking EC measurements. The usefulness of the modeled autumn lag was evidenced in its capability to explain 70% of annual NEP at HF site. A validation of the empirical function derived from HF site using 13-year (1999–2011) independent data at the University of Michigan Biological Station (UMB) forest was promising. Estimates of autumn lag using the exactly same meteorological variables proposed at HF site but different regression coefficients were highly correlated with the observed autumn lag (R2 = 0.87, p < 0.001) at UMB site. The correlation decreased slightly (R2 = 0.83, p < 0.001) if the regression coefficients found at HF site was also used, which subsequently explained 46% of annual NEP (p = 0.011) for UMB site. These results advocate for the inclusion of autumn phenology in terrestrial ecosystem models in order to predict the interannual variability of C sequestration more accurately, but also indicate challenges in deriving appropriate models even for the same plant functional types.