Accurate yield prediction in perennial crops is challenging when hierarchical datasets contain repeated response values because lower-level evaluation can inflate apparent model performance. We re-audited a four-year Coffea arabica var. Esperanza L4A5 dataset from a lowland agroforestry trial in Costa Rica. The 2020–2022 development dataset comprised 45 response-unit-year records representing 15 recurring management × fertilization combinations, while the corresponding 15 combinations in 2023 were retained as a temporal holdout. Decision Tree, Random Forest (RF), Support Vector Regression (SVR), and ridge regression were evaluated using forward temporal validation and compared with training-mean and previous-year persistence benchmarks. On the 2023 holdout, SVR (R² = 0.53; RMSE = 0.308 kg plant⁻¹) and RF (R² = 0.51; RMSE = 0.312 kg plant⁻¹) improved on the training-mean baseline, but previous-year yield persistence performed substantially better (R2 = 0.96; RMSE = 0.091 kg plant−1). Leakage-aware aggregation therefore materially changed conclusions about model performance, and added algorithmic complexity did not improve prediction beyond recent yield history in this short time series. Longer multi-site time series and genuinely pre-harvest predictors are required before operational forecasting can be justified.