An advanced optimization framework for hyperparameter tuning in predictive analytics

Exhaustive hyperparameter search techniques such as grid search and random search are widely used in predictive analytics but become computationally expensive and often suboptimal in high-dimensional or nonlinear settings. This challenge is particularly clear in tuning Cycle Reservoir with Jumps (CRJ) models for time series prediction, where multiple interdependent parameters critically affect performance. To address this, we propose enhanced variants of two population-based meta-heuristic algorithms: Cheetah Optimizer (CO) and Moth-Flame Optimization (MFO). By embedding crossover, mutation, and Lévy flight operators, we improve the search dynamics and convergence behavior of these optimizers. The resulting hybrid variants are evaluated on six benchmark datasets, including Henon Map, 10th-order NARMA, Sunspot, Santa Fe Laser, Lorenz Attractor, and Mackey-Glass. Our findings show that the enhanced optimizers consistently outperform their original versions in terms of NMSE, RMSE, and R2. For example, on the NARMA dataset, the NMSE was reduced from 0.0367 (MFO) to 0.0167 (COlevy), and on the Laser dataset, the NMSE dropped from 0.0168 (MFO) to 0.0093 (MFOlevy). These results confirm the effectiveness of our approach for scalable and accurate hyperparameter tuning in CRJ-based predictive analytics.