Learning from revisions: an algorithm to detect errors in banks’ balance sheet statistical reporting

Ensuring and disseminating high-quality data is crucial for central banks to adequately support monetary analysis and the related decision-making processes. In this paper, we develop a new machine learning procedure for the identification of errors in banks’ supervisory reports on loans to the private sector, which are employed in the Bank of Italy’s production of statistics on Monetary Financial Institutions’ (MFIs) Balance Sheet Items (BSI). We model a “Revisions Adjusted–Quantile Regression Random Forest” (RA–QRRF) algorithm in which the predicted acceptance regions of the reported values are calibrated through an individual “imprecision rate” derived from the entire history of each bank’s reporting errors and revisions collected by the Bank of Italy. The analysis shows that our RA-QRRF approach provides very satisfying results in terms of error detection, especially for loans to the households sector, and outperforms well-established alternative outlier detection procedures based on probit and logit models.