A MULTI-SCENARIO SYNTHETIC TREE-RING TESTBED FOR OUT-OF-SAMPLE EVALUATION OF DIAMETER-INCREMENT MODELS FOR SPECIES OCCURRING IN KAZAKHSTAN

The study uses synthetic data only: hypothetical trees, annual rings, geometric masks, and the high-frequency measurement channel are not field observations from Kazakhstan. The aim was to develop a multi-scenario testbed for model selection and out-of-sample claim evaluation. One hundred independent runs were conducted under four data-generating scenarios; each run contained 120 trees from four species, 3,120 annual rows, and 2,520 modeling rows. AR (5), OLS, Ridge, ElasticNet, kNN, Random Forest, and Gradient Boosting were compared. A prespecified simplicity rule was used when validation errors were practically equivalent within 1%. OLS was selected in 64% of linear runs, whereas kNN was selected in 96% of nonlinear runs. Median gain over Persistence on unseen trees ranged from 25.27% to 37.23%. Increasing sample size from 40 to 120 trees reduced median bootstrap interval width for ΔMAE from 0.278 to 0.171 mm. Under the 2023–2025 regime-shift scenario, median R² was –0.089 and gate G passed in 24% of runs. The results demonstrate that model ranking depends on the generating mechanism and support claims only within the declared computational testbed.