Deep Time-series Forecasting Needs Kernelized Moment Balancing
Existing deep time-series forecasting methods fail to achieve true distribution balance because they only match moments for a few predefined functions. The KMB-DF framework solves this by adaptively selecting informative balancing functions from a Reproducing Kernel Hilbert Space (RKHS). This approach enforces sufficient distribution alignment, leading to state-of-the-art accuracy across various models.