The objective of the study was to compare five lactation curve models, Ali and Schaeffer, Guo, Wilmink, and third- and fourth-order normalized Legendre polynomials, for their ability to reliably predict 305-d lactation yields from test-day records. The analysis covered 27 589 422 test-day yields from 3 350 638 first three lactations of 1 621 796 Polish Holstein-Friesian cows. All of the functions were fitted to the test-day yields of milk, fat and protein using a multiple-trait prediction (MTP) method. Based on the fitted curves, 305-d yields were calculated and the results were compared with the official lactation yields of cows assumed as true yields. Lactation curve models differed significantly (P < 0.01) when 305-d yields of milk traits were estimated and compared. Legendre polynomials showed the smallest bias and the highest accuracy among five compared models for all milk traits, whereas the Guo model for milk and the Wilmink model for fat and protein had the highest bias and the lowest accuracy. All of the functions produced similar estimates of milk, fat and protein yields at peak day; the estimated day of peak depended on the model.