Previous studies have shown that offline classification accuracy does not always correlate with real-time performance. However, the relationships between real-time performance and other offline measures, such as feature space metrics, are not clearly understood. We evaluated controller performance in intact limb (ITL) and amputee (AMP) subjects using online and offline tests in four limb positions and with three limb loads. We quantified the Pearson correlation coefficients between three offline metrics (offline accuracy, repeatability index, separability index) and four real-time metrics (completion rate, remaining time, movement efficacy, stopping efficacy). Our results showed that repeatability index had the strongest correlations with all real-time metrics (ITL: r ≥ 0.91, AMP: r ≥ 0.68). This suggests that pattern recognition control algorithms and training protocols should aim to optimize repeatability as opposed to accuracy or separability.