A1491
Title: Nonparametric kernel mixed data sampling models
Authors: Deliang Dai - Linnaeus university (Sweden) [presenting]
Abstract: A novel extension of the Fully Non-Parametric MIDAS model is introduced. Our approach leverages kernel methods to create a purely nonparametric framework, where estimation complexity depends solely on sample size. This allows the model to capture complex nonlinearities in both the temporal lag structure and the functional relationship between mixed-frequency series. A key feature is the application of a single kernel function to all lags of a given regressor. For estimation, we employ a kernel-based trend filtering algorithm that provides local adaptivity at a lower computational cost than standard spline regression. Empirical results on simulated and real-world data, including a case study forecasting urban air pollution from meteorological conditions, demonstrate the superior performance of our method over linear MIDAS.
