AI-Driven Inflation Forecasting in the Aftermath of COVID-19

Abstract

This paper develops a hybrid approach to inflation forecasting that combines large language models (LLMs) with traditional econometric methods. We construct a novel indicator of inflation expectations by applying GPT-4o to classify Twitter discussions among economists and financial commentators in the United States, the Eurozone, and Poland. Embedded in an ARIMAX framework, this indicator delivers systematic predictive content and consistently anticipates inflation dynamics up to 1 year ahead. Out-of-sample tests show that AI-augmented forecasts achieve statistically significant accuracy gains over both professional consensus forecasts and benchmark ARIMA models, particularly at medium and long horizons. Robustness checks—including alternative smoothing and lag structures, model variants, prompt formulations, and metadata weighting—confirm the stability of results across regions and specifications. By integrating high-frequency, sentiment-based signals with econometric forecasting, the study demonstrates how generative AI can complement conventional tools, offering central banks, policymakers, and financial market participants an early-warning mechanism during periods of uncertainty and structural change. While the approach is constrained by its reliance on Twitter data and the interpretability limits of LLMs, it highlights the transformative potential of hybrid AI-econometric frameworks for real-time macroeconomic analysis.