AI meets economics: Can deep learning surpass machine learning and traditional statistical models in inflation time series forecasting?

This study examined the forecasting ability of deep learning (DL) and machine learning (ML) models against benchmark traditional statistical models for the monthly inflation rates in the USA. The study compared various DL and ML models like transformers, linear regression, gradient boosting (GB), extreme gradient boosting (XGBoost), and adaptive boosting (AdaBoost) with traditional baseline time-series models like autoregressive integrated moving averages (ARIMA) and exponential smoothing (ETS) with Holt-Winters seasonal method utilizing data sourced from the Federal Reserve Bank of St. Louis. The study consistently showed that all DL and ML models outperformed the traditional approaches. In particular, the Transformer (RMSE = 0.0291, MAE = 0.0221) exhibited the lowest error rates, suggesting high forecasting accuracy for the monthly inflation data. In contrast, ARIMA (RMSE = 0.2038, MAE = 0.1895) and ETS (RMSE = 0.1619, MAE = 0.1455) models displayed higher error rates. Due to the ability of DL and ML models to process large volumes of data and identify underlying patterns, they are particularly effective for dynamic and evolving datasets. This study explored how novel DL modeling techniques could augment the accuracy and reliability of economic forecasting. The results indicate that policymakers and economists could leverage DL and ML models to gain deeper insights into inflation dynamics and implement more informed strategies to combat inflationary pressure. Hybrid methods should be further investigated in future studies to enhance economic forecasting accuracy.