RouteTS Improves Time Series Forecasting with Frequency-Time Routing.

Traditional time series forecasting models often struggle with data that combines global periodic patterns and localized, non-stationary changes. Models operating solely in the time domain can misalign periodic structures over long periods, while frequency-domain models tend to over-smooth sudden spikes. This research introduces RouteTS, a novel framework designed to address this by dynamically routing different components of the time series to their most suitable computational domains. RouteTS analyzes the frequency spectrum of the data, using amplitude routing to identify and separate dominant periodic components from residual, localized variations. The dominant frequencies are then processed using a complex-valued linear predictor in the frequency domain, which effectively preserves their periodic structure. Conversely, the remaining spectral energy, representing local changes, is converted back to the time domain and modeled by a lightweight Multi-Layer Perceptron (MLP). This dual-domain processing allows RouteTS to achieve superior prediction accuracy across various real-world datasets while also offering significant computational efficiency. The framework provides a principled solution to the challenge of simultaneously handling both global periodicity and local transience in time series data.