This article investigates the interindividual variability of underlying glucose dynamics and the relative predictive power of exogenous inputs in different frequency bands (FBs) for online subcutaneous glucose prediction for subjects with type 1 diabetes mellitus. Auto-regressive (AR) models and AR models with exogenous inputs (ARX) are developed based on two groups of ambulatory subjects and two groups of in silico subjects using different combinations of FBs. Some important modeling parameters are studied with respect to their influences on glucose prediction. Four issues are of particular interest and discussed based on the illustration results, suggesting that: (i) In different frequency bands, the underlying glucose dynamics act differently across subjects for online glucose prediction; (ii) A global AR model can be developed for one subject in some FB and then used to make online glucose predictions for other subjects in the same FB, revealing little interindividual variability; (iii) The exogenous inputs have different influences in different FBs for prediction of future subcutaneous glucose concentration; (iv) The exogenous inputs may not excite the glucose signals in some FBs, that is, the inclusion of the exogenous inputs may not result in more accurate models in comparison with standard AR model in some FBs. (c) 2013 American Institute of Chemical Engineers AIChE J, 59: 4228-4240, 2013