Adaptive Algorithms for Non-Stationary Data Environments

Abstract

Maintaining model performance becomes difficult when the data distribution fails to remain steady in dynamic contexts where the data generated is constantly changing. Typical machine learning algorithms struggle to cope with concept drift, which leads to lower prediction accuracy and reliability. The goal of this research is to design an adaptive learning system for forecasting financial markets and monitoring industrial equipment in non-stationary contexts. The proposed approach combines Adaptive Random Forest and Wolf Pack Search Optimization (WPS-ARF) model, where WPS optimizes ARF hyperparameters for enhanced adaptation, while ARF performs robust forecasting and fault detection in non-stationary environments. Preprocessing using Min-Max normalization ensures uniform feature ranges and stable learning, while PCA reduces dimensionality, removes redundancy, lowers computational cost, and mitigates overfitting. The model is coded in python and tested with financial and industrial datasets. Results show that WPS-ARF achieves Mean Squared Error (MSE) of 0.010, a Stability (CV) of 3, and a response time of 0.15 s for financial forecasting, while attaining 98% fault detection accuracy, a CV of 4, and a response time of 0.18 s for industrial monitoring, outperforming baseline models in accuracy, stability, and computational efficiency. As a result, the proposed model improves the predictive performance and robustness in the non-stationary environment, and can be used for decision-making in the financial and industrial systems.