The BPNN‐MIDAS Model and Its Unrestricted Variant: A Nonparametric Approach to Mixed‐Frequency Forecasting With Applications to Stock Markets

Currently, analyzing mixed frequency variables without losing information is a challenging task. To solve this problem, we propose a BP neural network based nonparametric mixed frequency data sampling regression (BPNN‐MIDAS) model, and further develop its unrestricted variant (i.e., BPNN‐U‐MIDAS) to markedly enhance the flexibility in capturing the patterns of high‐frequency data. First, we give their specific structures and develop the Adam‐based estimation procedure to optimize the model parameters. Then, numerical experiments show that the proposed models can integrate data of different frequencies and achieve nonlinear modeling with minimal information loss, which in turn provides more accurate analysis and prediction results. Finally, we illustrate the advantages of the proposed models by predicting the daily stock returns of the Chinese A‐share and US Nasdaq markets, which provide more accurate and comprehensive predictions for decision‐making.