Combining multiple forecasts for multivariate time series via state-dependent weighting

We present a model-free forecast algorithm that dynamically combines multiple forecasts using multivariate time series data. The underlying principle is based on the fact that forecast performance depends on the position in the state space. This property is exploited to weight multiple forecasts via a local loss function. Specifically, additional weights are assigned to appropriate forecasts depending on their positions in a state space reconstructed via delay coordinates. The function evaluates the forecast error discounted by the distance in the space, whereas most existing methods discount the error in relation to time. In addition, forecasts are selected with the function for each time step based on the existing multiview embedding approach; by contrast, the original multiview embedding selects the reconstructions in advance before starting the forecast. The proposed prediction method has the smallest mean squared error among conventional ensemble methods for the Rössler and the Lorenz'96I models. The results of comparison of the proposed method with conventional machine-learning methods using a flood forecast example indicate that the proposed method yields superior accuracy. We also demonstrate that the proposed method might even correctly forecast the maximum water level of rivers without any prior knowledge.