58: Mixed-Frequency Time Series Forecasting via Depth-Separable Neural Networks

Thomas M. Menino Convention & Exhibition Center 

This paper introduces the Depth-Separable Neural Network (DSNN), a novel neural-network-based framework for forecasting mixed-frequency data. Through a multi-depth architecture built from deep ReLU networks, the DSNN simultaneously performs adaptive frequency alignment and models complex nonlinear dependencies. Moreover, a parameter-sharing mechanism is adopted across the alignment networks, making the architecture separable by depth and computationally scalable even with a large set of higher-frequency predictors. We establish an approximation result for the DSNN over a broad class of hierarchical composition functions, and derive a non-asymptotic prediction error bound for its least squares estimator. Simulation studies demonstrate the finite-sample performance of the proposed method, and an empirical application to forecasting U.S. quarterly macroeconomic variables using monthly and daily indicators, highlights its superior predictive accuracy over existing mixed-frequency methods. The DSNN thus provides a theoretically grounded, scalable, and effective tool for complex mixed-frequency data analysis.

frequency alignment

least squares estimation

mixed-frequency data

non-asymptotic properties

ReLU neural network 

Business and Economic Statistics Section