Evaluating the Statistical Significance of Grid Search Optimization in Support Vector Regression for Dengue Forecasting Using Nested Time-Series Cross-Validation

Background Support Vector Regression (SVR) is widely employed for epidemic forecasting because kernel methods effectively model nonlinear, multivariate, and lagged climate-disease relationships. While practitioners routinely apply exhaustive grid-search hyperparameter optimization, whether such computationally expensive tuning produces statistically significant predictive enhancements under strict, leakage-free time-series validation remains an open empirical question. Objective To rigorously evaluate the statistical significance and practical utility of grid-search hyperparameter optimization for SVR-based weekly dengue case forecasting. Methods Utilizing the publicly available DengAI benchmark for San Juan, Puerto Rico, and Iquitos, Peru, from 1990 to 2013 (N = 1,872 city-weeks), this study constructs a strictly leakage-free pipeline. This pipeline incorporates median imputation, training-exclusive Z-score outlier mitigation, standardization, and Principal Component Analysis. We compared the default SVR configuration against a grid-tuned SVR optimized over the regularization constant, epsilon-insensitive tube, and kernel coefficient. Three validation protocols were applied: a chronological 80:20 hold-out, rolling-origin time-series cross-validation, and nested time-series cross-validation utilizing five outer folds. Statistical inference relied on paired Wilcoxon signed-rank tests applied to outer-fold errors, complemented by Cohen’s dz effect sizes, percentile-bootstrap 95% confidence intervals, and Holm-Bonferroni multiplicity correction. Methodological robustness was examined by systematically varying the number of retained principal components. Results On the chronological hold-out evaluation, the grid search reduced the Mean Absolute Error (MAE) by a marginal 0.30% in Iquitos and 0.59% in San Juan. Under nested cross-validation, the mean difference in MAE between the tuned and default models demonstrated a 0.024-case reduction in Iquitos (p = 1.0000) and a 0.494-case increase in San Juan (p = 0.1875). Neither result achieved statistical significance following multiplicity correction. Furthermore, the selected hyperparameters exhibited severe instability across the outer validation folds, particularly for San Juan. The optimization process also incurred an additional 11.48 seconds and 14.80 seconds of computational wall-clock overhead per city, respectively, without delivering a corresponding accuracy benefit. Conclusion Within the context of the DengAI benchmark and a robust leakage-free preprocessing pipeline, the default SVR configuration is statistically indistinguishable from its grid-tuned counterpart. These findings suggest that hyperparameter optimization in epidemic forecasting should not be executed automatically. Instead, it must be strategically guided by a pre-registered Smallest Effect Size of Interest and an explicit computational stopping rule. The comprehensive methodological framework presented in this study offers a potentially transferable framework for evaluating optimization utility in other low-fold time-series forecasting domains.