The choice of loss function is a critical factor in machine learning, particularly when training neural networks, as it directly influences model convergence, generalization, and forecasting performance. In this paper, we conduct a comprehensive comparison of 20 existing regression-based loss functions in the context of training Long Short-Term Memory (LSTM) neural networks for time series forecasting. In addition, this paper introduces a novel loss function, Angle Loss, designed in the context of Network Security Situational Awareness (NSSA) forecasting. The experimental evaluation is based on real-world time series data representing cybersecurity alerts collected by the Warden system. The performance of each loss function is assessed using standard forecasting accuracy metrics, including Mean Absolute Error (MAE*) and Mean Absolute Scaled Error (MASE*), and visualized through box plots to highlight robustness and variability. The results provide insights into the impact of loss function choice on predictive performance and inform future applications of neural networks in NSSA forecasting.