Business cycle and realized losses in the consumer credit industry

We investigate the determinants of losses given default (LGD) in consumer credit. Utilizing a unique dataset encompassing over 6 million observations of Italian consumer credit over a long time span, we find that macroeconomic and social (MS) variables significantly enhance the forecasting performance at both individual and portfolio levels, improving R 2 by up to 10 percentage points. Our findings are robust across various model specifications. Non-linear forecast combination schemes employing neural networks consistently rank among the top performers in terms of mean absolute error, RMSE, R 2 , and model confidence sets in every tested scenario. Notably, every model that belongs to the superior set systematically includes MS variables. The relationship between expected LGD and macro predictors, as revealed by accumulated local effects plots and Shapley values, supports the intuition that lower real activity, a rising cost-of-debt to GDP ratio, and heightened economic uncertainty are associated with higher LGD for consumer credit. Our results on the influence of MS variables complement and slightly differ from those of related papers. These discrepancies can be attributed to the comprehensive nature of our database – spanning broader dimensions in space, time, sectors, and types of consumer credit – the variety of models utilized, and the analyses conducted. • We study drivers of losses in defaulted consumer credits. • Adding macroeconomic and social variables improves recovery rate forecasting. • Non-linear forecast combinations using ANNs offer the best forecasts. • ALE and SHAP values reveal relationships between recovery rates and key predictors. • Lower output, rising public debt costs, greater uncertainty are linked to higher LGD.