Machine learning improves warnings of Italian municipal distress
Machine learning outperforms Italy's regulatory indicators in predicting municipal financial distress, a Bank of Italy paper finds. The best models achieve F1-scores of 0.272 for 2009–2015 and 0.297 for 2016–2023, against a regulatory benchmark of 0.133.
Better detection, uneven reliability
Matteo Bergamaschi, Michele Cascarano, Anna Laura Mancini, Francesco Rinaldi and Giacomo Ziglio analyse 7,778 municipalities over 2009–2023, covering 116,670 municipality-year observations and 897 distress events.
They compare six models across 99 configurations, separating periods before and after accounting harmonisation.
Random Forest leads median ROC AUC at 0.873 in 2009–2015 and 0.895 in 2016–2023, but logistic regression is more stable across specifications.
At F1-optimised thresholds, logistic regression achieves precision of 0.192 and recall of 0.469 in the first period; a convolutional neural network achieves 0.423 and 0.229 in the second.
Regulatory indicators record precision of 0.175 and recall of 0.107 over 2010–2023.
Faster data sacrifice predictive accuracy
Finalised OpenBDAP accounts provide the strongest predictions but arrive with delays.
Daily treasury advances from the Treasury Operations Information System (SIOPE) offer earlier liquidity monitoring, yet perform weakly alone and do not improve average accuracy when added to full accounts.
Budget forecasts combined with advances retain useful predictive information.
Two- and three-year target windows generally outperform one-year forecasts; histories beyond five years add little.
Interpretability tests identify debt-service pressure, advance reliance and capital-payment capacity as predictors.
Risk scores, not automatic sanctions
The authors position model alerts as inputs to supervisory review, not autonomous decisions.
Deployment requires calibration of false-alarm costs against the costs of missed crises.
Longer target windows also broaden the positive class, so better scores do not prove that more distant defaults are easier to predict.
IN: