The paper describes a machine learning process for identifying errors in banks' supervisory reports on loans to the private sector, which are employed by the Bank of Italy in the production of monetary and financial institutions' balance sheet item statistics. In particular, the paper proposes a 'Revisions Adjusted - Quantile Regression Random Forest' algorithm, in which the predicted acceptance regions of the reported values are calibrated through an individual 'imprecision rate' estimated on the basis of the entire history of each bank's reporting errors as logged by the Bank of Italy.
The Revisions Adjusted - Quantile Regression Random Forest provides very satisfying results in terms of error detection, especially for the loans to the households sector, for which the process identifies three quarters of banks' errors and 93 per cent of correctly reported values. The robustness analysis shows that the proposed algorithm outperforms well-established alternative outlier-detection methods based on probit and logit models.
Published in 2022 in: Quality & Quantity, v.56, pp. 4025-4059.
