Accurate channel modeling plays a pivotal role in optimizing communication systems, and fitting field measurements to stochastic models is crucial for capturing the key propagation features and to map these to achievable system performances. In this work, we shed light onto what's the most appropriate alternative for channel fitting, when the ultimate goal is performance analysis. Results show that likelihood-based and average-error metrics should be used with caution, since they can largely fail to predict outage probability measures. We show that supremum-error fitting metrics with tail awareness are more robust to estimate both ergodic and outage performance measures, even when they yield a larger average-error fitting.