Forecasting redenomination risk: unleashing the superior predictor via machine learning

Witzemann, Frederik; Arnorsson, Astthor

Forecasting redenomination risk: unleashing the superior predictor via machine learning

License

Citation

Witzemann, F, Arnorsson, A. Forecasting redenomination risk: unleashing the superior predictor via machine learning. 2023. handle: http://hdl.handle.net/10230/58411

This citation was generated automatically.

Abstract

We are the first to construct a general model for forecasting redenomination risk by using the difference in a country’s EUR-denominated CDS spreads compared to Germany. The goal of this work is to further close this gap by harnessing the power of machine learning algorithms, with a particular focus on techniques such as Random Forests, XGBoosting, and Lasso regressions. The results have important implications for policymakers, financial institutions and investors. Relying mainly on daily financial market data, we find that machine learning models significantly improve forecasting accuracy in both a structural crisis and a period of economic recovery. We conclude that XGB and Lasso are the best performing forecasting models, as for the latter modeling approach we get correlations between predicted and actual values as high as 0.6 for some countries. Finally, when we look at the measures of variable importance in our machine learning models, we indeed find that Lasso selects different types of predictors with much larger macroeconomic reference as relevant, which seems to explain the difference in prediction accuracy.