A Comparative Study of Loss Functions for Deep Neural Networks in Time Series Analysis

Currently, deep neural networks are widely used for analyzing temporal data. These networks can adapt their architecture to specific needs and deliver good performance. Researchers and developers frequently update their architecture to meet the requirements, but this process can be quite time-consuming. A crucial aspect of DNN architecture is the loss function, which plays a crucial role in calculating gradients. Most research and applications in time series analysis use the mean squared error (MSE) loss function. In this paper, we aim to explore existing loss functions to address the challenge of selecting the appropriate loss function for DNNs. We conduct experiments on time series datasets to evaluate the impact of different loss functions on DNN model performance. Our findings indicate that the Huber loss function outperforms other loss functions in time series analysis. Additionally, we discuss the potential for custom loss functions as future work, beyond the limitations of existing methods.