Buffered Spatial Cross-Validation for 5G SS-RSRP Error Estimation: A Multi-Campaign Drive-Test Benchmark

Validation geometry changes what a received-power prediction error describes. We benchmark established cross-validation designs on 2,967,199 aggregated 25 m location records from four UK operators and three Ofcom 5G SS-RSRP campaigns. A 10 km blocked, 1 km buffered design, selected using a separate development campaign, is compared with random, training-count-matched random, and unbuffered blocked validation using fixed inverse distance weighting (IDW), LightGBM, and XGBoost models. The buffered design has lower absolute disagreement with a prespecified same-campaign 20 km reference in all 36 original endpoints, reducing mean signed shortfall from 5.31 to 0.71 dB. The mean disagreement benefit is 1.83 dB for the boosted trees and 10.13 dB for IDW; blocking contributes more than the additional buffer. Post-outcome analyses correct exact-distance boundaries and endpoint weights, separate directional coverage errors, and add 12 coordinate-only random forest endpoints. The reference and distance diagnostics remain dependent on the same measurements, and only three campaigns support campaign-level sensitivity. Buffered validation still understates measured error for large contiguous-region holdout and later-campaign transfer by 4.69 and 4.47 dB. These results support reporting realized separation, training counts, and model-specific effects, while choosing validation geometry for the intended transfer setting rather than treating internal reference agreement as a universal deployment guarantee.