<p>Most experimental datasets provide essential information needed for validating predictive models. However, selecting the most appropriate predictive model for a given dataset is often a tedious and ambiguous task, particularly for geo‑materials whose mechanical behavior exhibits substantial variability and weakly defined selection criteria. In response to this drawback, this study evaluates the predictive performance of three models: Multiple Linear Regression (MLR), Support Vector Machine (SVM), and Decision Tree (DT) in estimating both unsoaked and soaked California Bearing Ratio (CBR) values from soil index properties. A dataset comprising liquid limit (LL), plastic limit (PL), optimum moisture content (OMC), maximum dry density (MDD), and percentage passing the No. 200 sieve (PP200) was used to develop and validate the models. The dataset consists of 80 observations for unsoaked CBR conditions and 80 for soaked CBR conditions, giving 160 experimental datasets in total. The novelty of this study lies in the comparative multi-model framework applied to both soaked and unsoaked CBR conditions using experimentally derived lateritic soil data, enabling condition-specific performance assessment and model data compatibility insights for geotechnical prediction. All the models yielded appreciable outcomes, with DT outperforming others in terms of non-linear behavior across both datasets. For the soaked dataset, DT-I achieved the best performance with MSE = 0.0062 and MAE = 0.0601, while for the unsoaked dataset, DT-2 produced the best accuracy with MSE = 0.0099 and MAE = 0.0610. Although SVM provided robust predictions for both conditions, it required careful parameter optimization to reach optimal performance. This indicates that the nature of any dataset and the features of the intended model to be employed are essential and should be matched for proper implementation, which leads to befitting results.</p>