Stacking machine-learning models for anomaly detection: comparing AnaCredit to other banking datasets

This paper addresses the issue of assessing the quality of granular datasets reported by banks via machine learning models. In particular, it investigates how supervised and unsupervised learning algorithms can exploit patterns that can be recognized in other data sources dealing with similar phenomena (although these phenomena are available at a different level of aggregation), in order to detect potential outliers to be submitted to banks for their own checks. The above machine learning algorithms are finally stacked in a semi-supervised fashion in order to enhance their individual outlier detection ability. The described methodology is applied to compare the granular AnaCredit dataset, firstly with the Balance Sheet Items statistics (BSI), and secondly with the harmonised supervisory statistics of the Financial Reporting (FinRep), which are compiled for the Eurosystem and the Single Supervisory Mechanism, respectively. In both cases, we show that the performance of the stacking technique, in terms of F1-score, is higher than in each algorithm alone.