HERA-XAI: Explaining One-Year-Ahead Corporate Rating Forecasts Through Hierarchical Artificial Intelligence (Zenodo (CERN European Organization for Nuclear Research))

Summary

• The HERA-XAI framework extends hierarchical corporate rating forecasting using 9 interconnected Random Forest classification nodes across Production, Trade, and Services sectors.p.9
• Node-level calibration showed substantial threshold variations, such as in Trade RF 3 where optimal thresholds were 0.35 under F1 and 0.70 under Youden criteria.p.12
• Global SHAP importance metrics were evaluated using subsamples of 40 observations per node and sector across terminal classifiers.p.29
• Illustrative policy simulations on Node 6 tested financial ratio shocks from -30% to +30% in increments of 5 percentage points to evaluate transitions from CCC to B rating classes.p.35

Key insights

1. Downscaling Random Forest classifiers to 50 trees substantially lowered computational burden while preserving predictive quality across hierarchical nodes.p.11
2. Optimal classification thresholds diverge significantly depending on whether F1-score, Youden Index, or minimum total error criteria are applied.p.12
3. In Node 1 for the Production sector, DEBTA emerged as the top predictor with an Out-of-Bag permutation importance score of 15.887.p.18
4. Terminal classifiers show sectoral differences in default discrimination, with DEBTA leading in Production (8.831), DEBPROD in Trade (5.558), and FAEQ in Services (9.128).p.19
5. Local SHAP decomposition on an RF8 firm revealed that DEBCOV (-0.0320) and FAPROD (-0.0313) provided the strongest negative shifts away from an AAA classification.p.27
6. Global SHAP rankings identified FAEQ as the primary contributor across all three sectors in Node 9 despite indebtedness measures dominating permutation accuracy.p.31
7. Policy simulations on RF6 demonstrated non-linear classification shifts where a 5% increase in ROA, DEBPROD, or ROS changed predicted ratings from CCC to B.p.38