18 Predicting monthly total sperm production of dairy bulls using additive and fixed age-based training windows

Selection of elite young bulls using genomic data has shortened the generation interval and increased pressure to collect large quantities of semen at an early age. Younger bulls pose a challenge in forecasting total sperm production, due to non-linear relationships between management factors and semen production. Our aims were to compare prediction methods for forecasting monthly total sperm (TS), evaluate additive versus fixed-window training sets, and identify management factors that improve TS forecasting. Data were from a commercial AI company and included 8,060 monthly collection records from 1,118 Holstein and Jersey bulls between 10 and 28 mo of age. Potential explanatory variables included: year and season of collection; barn location; number of collections per month; breed; scrotal circumference at 10–11 mo of age; TS in three previous collection months; and age at arrival, first collection, and current collection. Training and testing sets were split by age at collection. In the additive approach, all prior data were used as training, while in fixed-window, records from 3 previous months were used. Five-fold cross validation was used to train models (R v3.5.1). Prediction models included linear regression (LM), random forest (RF), and Bayesian regularization neural network (BRNN), with performance measured by root mean squared error (RMSE) and correlation (r) between actual and predicted TS of testing sets. Models with fixed training sets (RMSE = 19.07, r = 0.847) performed better than additive training sets (RMSE = 19.82, r = 0.833). RF (RMSE = 19.25, r = 0.855) performed slightly better than BRNN (RMSE = 19.25, r = 0.828) and LM (RMSE = 19.87, r = 0.836). The most important management variables affecting TS were: collection frequency, TS in previous collection months, and age at collection. Preliminary results suggest fixed-window training of RF models yield best TS prediction.