Macroeconomic Forecasting with LSTM and Mixed Frequency Time Series Data

This paper investigates the potentials of the long short-term memory (LSTM) when applying with macroeconomic time series data sampled at different frequencies. We first present how the conventional LSTM model can be adapted to the time series observed at mixed frequencies when the same mismatch ratio is applied for all pairs of low-frequency output and higher-frequency variable. To generalize the LSTM to the case of multiple mismatch ratios, we adopt the unrestricted Mixed Data Sampling (U-MIDAS) scheme (Foroni et al., 2015) into the LSTM architecture. We assess via both Monte Carlo simulations and empirical application the out-of-sample predictive performance. Our proposed models outperform the restricted MIDAS model even in a set up favorable to the MIDAS estimator. For real world application, we study forecasting a quarterly growth rate of Thai real GDP using a vast array of macroeconomic indicators both quarterly and monthly. Our LSTM with U-MIDAS scheme easily beats the simple benchmark AR(1) model at all horizons, but outperforms the strong benchmark univariate LSTM only at one and six months ahead. Nonetheless, we find that our proposed model could be very helpful in the period of large economic downturns for short-term forecast. Simulation and empirical results seem to support the use of our proposed LSTM with U-MIDAS scheme to nowcasting application.