This paper investigates the evolution of deep learning architectures for time series forecasting, focusing on the transition from language model-based encoders to retrievalaugmented domain-aware frameworks for time series. We analyze the strengths and shortcomings of transformer-based models trained in time series data, as exemplified by NV-Tesseract [1], in handling multivariate and cross-domain datasets for forecasting. Through a diagnosis of architectural failure modes, we motivate a new Domain Aware Representation and Retrieval forecasting approach with added neural contextualization. Experiments on public benchmarks and industrial datasets demonstrate that retrieval mechanisms improve resilience to domain shifts in complex targets, with implications for zero-shot generalization and practical deployment.
Read the full paper here: IEEE