Unemployment is a key economic factor for local and global economies. Having a working population ensures economic well-being, financial opportunities as well as allowing individuals to participate and contribute to society. Unemployment is, however, influenced by many economic as well as non-economic factors and it can be difficult to account for all driving factors when making predictions. Nevertheless, having an accurate and robust system for unemployment forecasting could provide policy and decision-makers with vital information ahead of time, allowing preemptive measures to be taken and unemployment rates kept low. This manuscript investigates the prospect of long-short term memory (LSTM) based neural networks for forecasting unemployment rates based on openly accessible real-world data. However, as forecasting model performance is tied to hyperparameter assortment, a refined metaheuristics algorithm based on the elk herd optimizer (EHO) is introduced to ensure favorable outcomes. Simulations on real-world data suggest favorable performance with the best-performing models attaining a mean absolute error (MAE) as low as 6.31%.