The field of time series forecasting is witnessing significant developments, with a focus on improving the accuracy and robustness of forecasting models. Researchers are exploring new architectures and techniques to capture complex patterns and relationships in time series data, including the use of partially asymmetric convolutional neural networks, sparsity-robust foundational forecasters, and physics-informed neural networks. These innovations aim to address the challenges of modeling time series with strong heterogeneity in magnitude and/or sparsity patterns, as well as the need for economically consistent predictions. Noteworthy papers in this area include:

• A novel convolutional architecture that achieves state-of-the-art results on popular time series datasets by adaptively fuzzifying temporal data and using a bilateral Atrous algorithm to reduce calculation demand.
• A robust forecasting architecture that improves overall prediction accuracy by reducing magnitude- and sparsity-based systematic biases.
• A deep learning framework that integrates a Gated Recurrent Unit architecture with physics-informed neural network principles to provide a robust, interpretable, and scalable solution for high-dimensional nonlinear time series forecasting.
• A unified ensemble framework that enhances the performance of foundation models for sales forecasting in real-world supply chains by combining hierarchical and architectural ensemble strategies.
• A novel evaluation metric that transforms time series into images to leverage their inherent two-dimensional geometric representations, and a multi-component loss function that enhances structure modeling during training.