Measuring fragmentation risk in European bond markets using machine learning
Аннотация
This paper asks whether machine learning can forecast euro area sovereign bond spreads and whether the resulting forecasts can track financial fragmentation. Using a new high-dimensional monthly dataset of 4948 macro-financial series for ten euro area countries from December 2008 to February 2025, we run a horse race among thirteen machine-learning models and two simple benchmarks, an AR(1) process and a random walk. XGBoost is the strongest machine-learning model and is never significantly outperformed by the other learners. It is not, however, the most accurate one-month-ahead forecaster: under strict out-of-sample re-estimation the AR(1) and the random walk attain lower point-forecast errors in every country. The value of the machine-learning approach lies elsewhere. SHAP decompositions recover the macro-financial drivers of the predicted spreads. The forecasts form the basis of a fragmentation indicator built by clustering predicted spread paths. The indicator reproduces the core-periphery divide in the windows running to 2022. French spreads then decouple from the core after 2023. By 2024–2025 France and Belgium form a distinct cluster, a new source of fragmentation risk with direct implications for the transmission of a single monetary policy.