Deep learning on mixed frequency data

In deep learning, it is common to encounter data observed at different frequencies. Mixed data sampling (MIDAS) is an efficient technique for handling mixed frequency data, where a high frequency predictor is converted into a set of low frequency variables using the frequency alignment approach and parametric function constraints. This efficiently prevents the proliferation of parameters, ensuring the consistency of data frequency. We introduce the MIDAS technique into the deep learning architecture and develop a novel deep learning‐MIDAS (DL‐MIDAS) model, which enables to conduct deep learning on raw mixed frequency data directly. Its efficacy is then illustrated through extensive Monte Carlo simulations and a real‐world application. The simulation experiments show that the DL‐MIDAS model is able to explore nonlinear patterns in mixed frequency data and achieves more stable and accurate prediction results than several competing models, such as the artificial neural network for mixed frequency data (ANN‐MIDAS), long short‐term memory (LSTM), and MIDAS regressions. Additionally, the real‐world application of predicting the inflation rate of China also confirms the strength of DL‐MIDAS. The model can exploit high frequency information contained in financial market to produce timely and accurate prediction results on the inflation rate.