Comparative computer simulation and empirical analysis of MIDAS and artificial neural network-UMIDAS models for short- and long-term US GDP forecasting

Purpose

This study aims to evaluate linear mixed data sampling (MIDAS), nonlinear artificial neural networks (ANNs) and a hybrid approach for exploiting high-frequency information to improve low-frequency gross domestic product (GDP) forecasting. Their capabilities are assessed through direct forecasting comparisons.

Design/methodology/approach

This study compares quarterly GDP forecasts from unrestricted MIDAS (UMIDAS), standalone ANN and ANN-enhanced MIDAS models using five monthly predictors. Rigorous empirical analysis of recent US data is supplemented by Monte Carlo simulations to validate findings.

Findings

The empirical results and simulations demonstrate that the hybrid ANN-MIDAS performs best for short-term predictions, whereas UMIDAS is more robust for long-term forecasts. The integration of ANNs into MIDAS provides modeling flexibility and accuracy gains for near-term forecasts.

Research limitations/implications

The model comparisons are limited to five selected monthly indicators. Expanding the variables and alternative data processing techniques may reveal further insights. Longer analysis horizons could identify structural breaks in relationships.

Practical implications

The findings guide researchers and policymakers in leveraging mixed frequencies amidst data complexity. Appropriate modeling choices based on context and forecast horizon can maximize accuracy.

Social implications

Enhanced GDP forecasting supports improved policy and business decisions, benefiting economic performance and societal welfare. More accurate predictions build stakeholder confidence and trust in statistics underlying critical choices.

Originality/value

This direct forecasting comparison offers unique large-scale simulation evidence on harnessing mixed frequencies with leading statistical and machine learning techniques. The results elucidate their complementarity for short-term versus long-term modeling.