Multivariate Time Series Forecasting

The field predicts several changing quantities together, such as electricity demand across regions, weather at multiple stations, or patients’ vital signs, where one variable may provide clues about another. Common approaches include linear and MLP models, recurrent networks, Transformers and state-space models, as well as graph, frequency-based, and decomposition methods; newer work also explores probabilistic generation, foundation models, and irregularly sampled data. Models are commonly compared on benchmarks such as ETT, Electricity, Traffic, Weather, and GIFT-Eval, with growing attention to long horizons, uncertainty, and transfer across datasets. A central tension is that cross-variable links and extra context can improve forecasts when informative, but noisy, shifting, or irrelevant relationships can hurt accuracy, so robustness and efficiency matter alongside average error.