Interpretable Machine Learning and Economic Data: Volatility Spillover along the Supply Chains
We introduce a financial network approach to quantify the impact of counterparty risk on firms’ daily market risk, measured via conditional volatility. We asses competing econometric and machine learning approaches and assess the economic interpretability of the applied machine learning algorithms. We find that XGBoost in combination with SHAP values describe a sensible choice for large economic data sets which are described by a panel structure. Also, suppliers are exposed to additional fundamental risk that is not captured by the suppliers’ market beta, which gets transferred along the supply chains. The identified risk spillover impact both dimension and quality of the suppliers’ market risk assessment: If customers experience large losses, suppliers’ variance forecasts increase by (up to) 1% and the probability of suppliers’ extreme losses doubles the next day.