Type 1 diabetes (T1D) management remains particularly challenging in the presence of external disturbances such as stress or physical activity (PA). While the role of meals has been extensively studied, the glycemic impact of PA is still not fully understood and lacks standardized modeling approaches. This study leverages the Type 1 Diabetes and Exercise Initiative (T1DEXI) clinical trial, the largest publicly available dataset of individuals with T1D undergoing both free-living and structured exercise, to explore inter-individual differences in glycemic control and prediction performance. A filtering and characterization pipeline is defined to extract summary statistics from glucose, heart rate, and PA signals. These features are used for unsupervised clustering of patients. Among the algorithms evaluated, the Spectral Clustering algorithm successfully separates individuals into three groups based on glycemic control metrics. The relevance of the cluster-defining variables is then assessed, and their relationship to the performance of a personalized long short-term memory (LSTM) neural network trained to forecast glucose one hour into the future is analyzed. Results show that patients with better glycemic control, often those achieving higher levels of weekly PA, exhibit lower prediction errors and more stable dynamics. These findings highlight a positive synergy: individuals who engage in regular PA not only experience improved glycemic regulation but also benefit from more reliable predictive models. Given that fear of hypoglycemia remains one of the main barriers to exercise in people with T1D, these results suggest that cluster-related predictors with clinically acceptable performance could support safe PA engagement and empower data-driven decisions.