We address the question of how to identify fast road races in running by automatically decomposing race results into athlete performance and race condition components. Our approach does not require explicit modeling of influencing factors such as course terrain profiles. Favorable conditions have a substantial impact on race results in road running, and can be critical for meeting championship qualifying standards or for achieving personal bests. We frame this problem as an instance of weighted nonnegative matrix factorization and validate our approach using 6,000 real-world 10k race results from recent local to regional level races. Extensive experiments on both this real-world data and simulated data demonstrate the robustness of this method to high missing value rates and its ability to reduce bias in estimating race conditions compared to mean- or median-based approaches. Our approach also successfully recovered seasonal patterns in race conditions. The number of races and the rate of missing values were found to be the most important properties affecting accuracy, while the number of athletes had less impact.