Development and application of machine learning models in US consumer price index forecasting: Analysis of a hybrid approach

Abstract

This study aims to apply advanced machine-learning models and hybrid approaches to improve the forecasting accuracy of the US Consumer Price Index (CPI). The study examined the performance of LSTM, MARS, XGBoost, LSTM-MARS, and LSTM-XGBoost models using a large time-series data from January 1974 to October 2023. The data were combined with key economic indicators of the US, and the hyperparameters of the forecasting models were optimized using genetic algorithm and Bayesian optimization methods. According to the VAR model results, variables such as past values of CPI, oil prices (OP), and gross domestic product (GDP) have strong and significant effects on CPI. In particular, the LSTM-XGBoost model provided superior accuracy in CPI forecasts compared with other models and was found to perform the best by establishing strong relationships with variables such as the federal funds rate (FFER) and GDP. These results suggest that hybrid approaches can significantly improve economic forecasts and provide valuable insights for policymakers, investors, and market analysts.