Leveraging Structural Insights: Enhancing Inflation Forecasting With Dynamic Stochastic General Equilibrium-Informed Long Short-Term Memory Models

Abstract

This study investigates the use of long short-term memory (LSTM) models for forecasting US inflation, integrating structural insights from the Federal Reserve Bank of New York dynamic stochastic general equilibrium (FRBNY DSGE) model to enhance both accuracy and theoretical grounding. Traditional models often struggle with the complexities of macroeconomic dynamics, while machine learning models like LSTM excel at capturing nonlinear relationships but lack interpretability. By using the DSGE model to guide variable selection, this research incorporates key economic sectors into the LSTM framework. An incremental analysis reveals that financial market variables, particularly credit spreads, significantly improve forecasting performance, while government-related variables contribute minimally and may reduce model reliability. The out-of-sample forecasting exercise demonstrates that the LSTM model produces projections more aligned with historical inflation patterns compared with the DSGE model, which often oversimplifies macroeconomic fluctuations. These findings highlight the potential of combining machine learning techniques with economic theory to create robust and reliable forecasting tools. The proposed approach not only enhances the accuracy of inflation forecasts but also provides a practical framework for central banks and financial institutions aiming to refine their models by leveraging the strengths of machine learning within a theoretically consistent structure. This integration represents a valuable advancement in macroeconomic forecasting methodologies.