This paper proposes a methodology to identify anomalies in the reporting of quarterly insurance asset data, detecting changes in the identification codes of the assets, which negatively affect the quality of the statistics that the Bank of Italy compiles. The proposed model is based on a record linkage approach using supervised machine learning classification models.
The model, tested on 2019-2022 data, shows robust performance for different levels of data quality and for different asset types: in most cases, the model is able to detect almost all cases of anomalies while ensuring a low rate of data erroneously classified as anomalous (false positives). Since June 2022, the model has been included in the ordinary procedures for managing the quality of insurance data and the feedback received so far confirms its effectiveness.
