2026-05-05T00:38:44.207Z http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai

ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai:2:11699 2025-11-07T22:10:32Z 2

A multivariate methodology combining recurrent neural networks with the modified variable neighborhood search algorithm for unemployment forecasting 2025 http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai/record/2/11699 https://www.sciencedirect.com/science/article/abs/pii/S1568494625014565 L. Babic L. Jovanovic J. Cadjenovic M. Mohamed Elsayed M. Hammad M. Zivkovic V. Simic N. Bacanin Tracking unemployment trends is a vital task as it serves as a critical indicator of economic health, influencing decisions on job creation, monetary policies, and workforce development. Beyond economic implications, unemployment fluctuations affect social well-being, public health, and social stability. Predicting unemployment rates allows governments, economists, and businesses to forecast labor market dynamics and make informed strategic decisions. While traditional forecasting methods like time series analysis have been widely used, they often struggle with capturing complex, nonlinear patterns in data, leading to reduced predictive accuracy. Machine learning models, particularly recurrent neural networks (RNNs) and their advanced variants, such as gated recurrent units (GRUs), offer promising alternatives by handling large datasets, uncovering hidden patterns, and adapting to dynamic economic conditions. However, challenges such as data quality, interpretability, and the need for customized hyperparameter optimization persist. This study proposes a hybrid RNN and GRU-based approach for unemployment forecasting, employing a modified variable neighborhood search algorithm for hyperparameter tuning. Furthermore, sHapley additive exPlanations (SHAP) analysis is conducted to amplify interpretability by elucidating model behavior and feature importance. The experimental findings reveal the potential of this hybrid methodology for improving forecast accuracy and decision-making reliability with simulations carried out on real-world data that yield a mean square error of just 0.005567 for the best-performing model. article 10.1016/j.asoc.2025.114143 114143 1568-4946 Applied Soft Computing