The use of rolling averages to analyse training data has been debated recently. We evaluated two training load quantification methods (five-zone, seven-zone) fitted to performances over two race distances (50 m and 100 m) using four separate longitudinal models (Banister, Busso. rolling averages and exponentially weighted rolling averages) for three swimmers ranked in the top 8 in the world. A total of 1610 daily load measures and 108 performances were collected. Banister (standard error of the estimate (SEE) 0.64 s & 0.62 s; five-zone & seven-zone quantification methods), Busso (SEE 0.73 s & 0.70) and exponentially weighted rolling averages (SEE 0.57 s & 0.63 s) models fitted more accurately (p < 0.001) than the rolling averages approach (SEE 1.32 s and 1.36 s). The seven-zone quantification method did not produce more accurate performance predictions than the five-zone method, despite being a more detailed form of training load quantification. Four neural network models were fitted and had lower error (SEE 0.38, 0.41, 0.35 and 0.60 s) than all longitudinal models but did not track as predictably over time. Exponentially weighted impulse-response models and exponentially weighted rolling averages appear more effective at predicting performance using training load data in elite swimmers than a rolling averages approach.