Forecasting inflation with machine learning

This dissertation explores the integration of machine learning techniques and narrative-based data for enhancing inflation forecasting, with a focus on the Russian economy. Traditional econometric models often fail to capture nonlinear dynamics and shifting macroeconomic conditions, particularly in volatile contexts. To address these limitations, this study employs a hybrid approach combining macroeconomic indicators with narrative data extracted from Russian news articles using Latent Dirichlet Allocation (LDA). Predictive modeling is conducted using machine learning algorithms, including Random Forest, LASSO, and Elastic Net. Macroeconomic data, sourced from platforms such as RosStat and Investing.com, and narrative data, obtained via web scraping from Lenta.ru, were preprocessed to ensure consistency and stationarity. The analysis reveals that integrating text-derived features with economic indicators improves forecasting accuracy across multiple horizons. Random Forest consistently outperforms other models, particularly in short-term forecasts, underscoring its ability to leverage both structured and unstructured data. Key contributions include the development of a scalable framework for inflation forecasting in non-Western economies, validation of narrative-based predictors, and incorporation of geopolitical factors like sanctions into the analysis. The results show that combining macroeconomic data with narrative-based information leads to better predictions than using either type of data alone, highlighting the value of qualitative insights in understanding and managing economic fluctuations.