Short communication: Projecting milk yield using best prediction and the MilkBot lactation model

The accuracy and precision of 3 lactation models was estimated by summarizing means and variability in projection error for next-test milk and actual 305-d milk yield (M305) for 50-d intervals in a large Dairy Herd Improvement Association data set. Lactations were grouped by breed (Holstein, Jersey, and crossbred) and parity (first vs. later). A smaller, single-herd data set with both Dairy Herd Improvement Association data and daily milk weights was used to compare M305 calculated from test-day data with M305 computed by summing daily milk weights. The lactation models tested were best prediction (BP), the nonlinear MilkBot (MB) model, and a null model (NM) based on a stepwise function. The accuracy of the models was ranked (best to worst) MB, BP, and NM for later-parity cows and MB, NM, and BP for first-parity cows, with MB achieving accuracy in projecting daily milk of 0.5 kg or better in most groups. The models generally showed better accuracy after 50 d in milk. Best prediction and NM had low accuracy for crossbred cows and first-parity Holstein and Jersey cows. The MB model appears to be more precise than BP, and NM had low precision, especially for M305. Regression of model-generated M305 on summed M305 showed BP and MB to be equally efficient in ranking lactations, but MB was better at quantifying differences.