Do machine learning techniques outperform autoregressive distributed lag models in inflation forecasting?

Abstract

Following the COVID-19 pandemic, Romania and other Central and Eastern European (CEE) countries faced some of the highest inflation rates in the European Union, creating a pressing need for accurate short-term forecasts to guide monetary policy. This study compares modern machine learning (ML) methods-Long Short-Term Memory (LSTM) neural networks, Random Forests (RF) and Support Vector Regression (SVR)-with traditional Autoregressive Distributed Lag (ARDL) models in forecasting Harmonised Index of Consumer Prices. Using quarterly data for Romania (2006Q1-2023Q4) and monthly data for nine CEE economies (2006M1-2025M3), we incorporate unemployment and sentiment indicators derived from the Romanian Central Bank reports and the European Commission's Economic Sentiment Indicator (ESI). We further evaluate model performance through simulation experiments that include high persistence, moving-average non-invertibility, nonlinear regimes, and structural breaks. Across both empirical and LSTM and SVR models-they frequently deliver lower forecast errors than ARDL, with LSTM achieving up to 53% reductions in mean squared error relative to na & iuml;ve benchmarks. However, ARDL remains competitive when sentiment indices are the main predictor. These findings highlight that while advanced ML models can capture nonlinear dynamics and regime changes, traditional econometric tools still provide valuable robustness, particularly in sentiment-driven contexts. Overall, integrating ML, econometric approaches, and sentiment analysis offers a more reliable toolkit for short-horizon inflation forecasting under economic uncertainty.