We present, to our knowledge, the first methodological study aimed at enhancing the prognostic power of Cox regression models, widely used in survival analysis, through optimized data selection. Our approach employs a novel two-stage mechanism: by framing the prognostic stratum matching problem intuitively, we select prognostically representative patient observations to create a more balanced training set. This enables the model to assign equal attention to distinct prognostic subgroups. We demonstrate the methodology using an observational dataset of 1,799 patients with resected colorectal cancer liver metastases, 1,197 of whom received adjuvant chemotherapy and 602 who did not. In our study, as is current standard practice, the comparator was training prognostic models on the entire cohort (referred to as "model 1"). Models trained on the untreated and treated subgroups, matched through our approach (referred to as "model 3"), showed an improvement of up to 20% in bootstrapped C-indices compared to model 1. Notably, model 3 exhibited superior calibration, with a 6- to 10-fold improvement over model 1. Additional performance metrics aligned with these findings, and robustness was confirmed through bias-corrected bootstrapping. Given the ongoing development of numerous linear prognostic models and the general applicability of our approach to any observational data, this method holds significant potential to impact biomedical research and clinical practice where prognostic models are utilized.