QRNN-MIDAS: A novel quantile regression neural network for mixed sampling frequency data

Abstract

Abstract Text of abstract In the big data era, it is common to encounter data observed at different frequencies. This raises the problem of how to explore the heterogeneous nonlinear relationship between variables on mixed sampling frequency data. To this end, we develop a novel quantile regression neural network for mixed sampling frequency data called QRNN-MIDAS by introducing the Mixed Data Sampling (MIDAS) technique into the framework of quantile regression neural network (QRNN). The proposed QRNN-MIDAS model enables QRNN to handle raw mixed sampling frequency data directly. Specifically, we conduct frequency alignment on each high frequency input variable according to the given maximum lag order. Then, a convenient parametric weight function is imposed on the frequency alignment vector and a low frequency variable is obtained. This strategy allows the QRNN-MIDAS model to extract valuable information from raw mixed sampling frequency data, which is helpful to explore the heterogeneous nonlinear relationship between variables in real time. To illustrate the efficacy of QRNN-MIDAS, both Monte Carlo simulation studies and real-world data applications are considered. The numerical results show that the QRNN-MIDAS model outperforms several competing models in terms of goodness-of-fit and predictive ability. In addition, US quarterly GDP growth and China’s monthly inflation forecast results also illustrate the superiority of the QRNN-MIDAS model, and provide more timely, accurate and comprehensive forecasts for decision-making.