StateFlow Improves Long-Horizon Time Series Forecasting with Dual-State Modeling.

Long-horizon multivariate time series forecasting (LTSF) presents significant challenges due to issues like non-stationarity, regime shifts, and the accumulation of errors over time. Existing models often struggle to accurately capture these complex dynamics. A new recurrent forecasting framework, named StateFlow, has been developed to address these limitations by extending the Variability-Aware Recursive Neural Network (VARNN) for multi-step predictions. StateFlow employs a dual-state recurrent backbone that captures two complementary signals from the historical data. One state tracks the primary temporal dynamics, including trends, seasonality, and recurring patterns, while the other, a residual-memory trajectory, represents structured local prediction deviations driven by one-step prediction errors. This dual-state approach allows the model to maintain a rich understanding of both global patterns and local anomalies. The framework utilizes a chunk-based decoder to summarize these trajectories and map them directly to the future horizon for multi-step forecasting. A two-stage optimization strategy further refines the model: first, the VARNN encoder is trained for one-step base prediction, and then a horizon-specific decoder is optimized for direct multi-step forecasting. Experiments on standard LTSF benchmarks demonstrate that StateFlow achieves competitive performance against various strong baselines, including linear, recurrent, convolutional, and Transformer-based models, all while maintaining a compact and efficient design.