A Comparison of Data-Driven Synthetic Performance Indicators for Default Prediction

This paper compares two approaches to constructing synthetic performance indicators for corporate default prediction: the expert-informed Synthetic Performance Indicator (ISP) and a Neural Network Synthetic Performance Indicator (NNISP). Both are built upon a common set of economic and financial indices from a large panel of firms operating in the Triveneto macro-region of Italy, including not only large companies but also small and medium enterprises (SMEs). The ISP relies on expert elicitation to assign weights to eight selected indicators, while the NNISP is obtained by calibrating the set of composition weights by training a Neural Network to predict a firm’s liquidity. We assess the predictive power of each score using multiple classification models and benchmark them against the full set of indicators. We show that the ISP offers stable and interpretable performance across model types, while the NNISP exhibits more flexibility in non-linear settings. These findings highlight a trade-off between expert-based interpretability and machine-driven adaptability, pointing to the potential of combined methodologies for future applications in credit risk modeling.