Unemployment forecasting: Integrating economic narratives with machine learning

This dissertation explores the potential of integrating textual data with macroeconomic indicators to enhance the predictive accuracy of unemployment forecasting models. Nowadays, traditional unemployment forecasting models rely heavily on historical quantitative data, but recent studies suggest that unstructured textual data may improve prediction accuracy since it is capable of capturing a different dimension of unemployment that may be overlooked when only using quantitative data. Building on methodologies from prior research, this analysis employs a hybrid approach to assess the predictive power of machine learning models trained with three different types of data, namely: macroeconomic, textual, and a combination of both. The findings reveal that while the topics generated from economic narratives underperform in short- to medium-term predictions compared to traditional macroeconomic models, they show significant improvements in long-term forecasts, achieving enhancements of 3.83% to 13.81% over macroeconomic methods. The combined approach also shows limited effectiveness for short-term predictions, however, it begins to show positive results in medium-term forecasts. In long-term forecasts, this approach shows variations ranging from -0.21% to 5.12% relative to macroeconomic methods, although its performance frequently falls below the Autoregressive benchmark. In conclusion, textual data derived from economic narratives can enhance unemployment forecasting, particularly for longer horizons, however, their inconsistent performance, along with the complexity of the news retrieval process presents important trade-offs to consider when adopting this approach.