**Background:** Body weight is the most economically important trait in sheep production, directly affecting meat yield and breeding income. Biometric measurements can serve as indirect selection criteria for body weight estimation, particularly in settings where weighing equipment is unavailable. The Polish Merino is Poland's most common commercial breed, but the national sheep population lacks sufficient meat-type animals. Crossbreeding Polish Merino ewes with Suffolk rams has been proposed as a strategy to improve meat characteristics. While various statistical methods have been used for body weight estimation in different breeds, no prior study had applied CART, Support Vector Regression (SVR), and Random Forest Regression (RFR) to crossbred populations with varying proportions of Polish Merino and Suffolk genotypes.
**Methods:** The study used data from 344 animals collected between 1990–1995, comprising 133 pure Suffolk sheep, 114 R2 crossbreds (75% Suffolk, 25% Polish Merino), and 97 R3 crossbreds (87.5% Suffolk, 12.5% Polish Merino). The sample included 88 rams and 256 ewes. Predictor variables included genotype, birth weight (BiW), sex, birth type, and seven body measurements: withers height (WH), sacrum height (SH), chest depth (CD), chest width (CW), chest circumference (CC), shoulder width (SW), and rump width (RW). The outcome variable was body weight at 12 months (LBW). Three algorithms were compared: CART (with 10-fold cross-validation, minimum node size of 5), SVR (using Gaussian radial basis kernel function with optimized epsilon and cost parameters), and RFR (500 trees, mtry=5). Model performance was evaluated using Pearson correlation coefficient (r), root mean square error (RMSE), coefficient of determination (R²), Akaike's information criterion (AIC), mean absolute percentage error (MAPE), and standard deviation ratio (SD ratio).
**Key Results:** The strongest correlation with LBW was chest circumference (CC, r=0.72, p<0.05). All other body measurements except sacrum height showed significant correlations with LBW. The CART algorithm identified CC as the primary splitting variable: sheep with CC < 94 cm had a mean LBW of 49 kg, while those with CC ≥ 94 cm had a mean LBW of 63 kg. The highest LBW node (88 kg) was observed in animals with CC ≥ 94 cm, CD ≥ 32 cm, genotype not R2 (i.e., R3 or Suffolk), and CD ≥ 34 cm. Sensitivity analysis for SVR confirmed CC as the most important predictor, while genotype and birth type (twin) had the smallest relative importance. For RFR, sensitivity analysis showed CC, CD, and SW had virtual significance >10%, while BiW had the lowest. For the test set, RFR achieved the best performance across all metrics: highest R² and r, and lowest RMSE, SD ratio, MAPE, and AIC. SVR performed best on the training set but showed evidence of memorization (overfitting), yielding substantially weaker test set results. CART was the weakest performer among the three algorithms.
**Clinical Implications:** The RFR algorithm provides a reliable, non-invasive method for estimating body weight from biometric measurements in Suffolk × Polish Merino crossbred sheep, which can assist breeders in herd management decisions including feed calculation, drug dosing, and determination of optimal slaughter weight. The finding that CC ≥ 94 cm, CD ≥ 32 cm, and genotype at the R3 level or higher (≥87.5% Suffolk) are associated with the highest body weights (up to 88 kg) offers practical selection criteria for breeding programs aimed at increasing meat production. The study demonstrates that algorithm selection is genotype-dependent, as prior work in Thalli sheep found SVR superior to RFR—the opposite of the current findings. Breeders should therefore validate model performance on their specific populations rather than assuming generalizability across breeds.