Get started
Work

Data Science for Internal Audit in Banking: Refinement of an Internal Audit Alarmistic System with Machine Learning

Year: 2022
Type: dissertation
Abstract: This report presents the work developed during the academic internship required for obtaining the Master’s Degree in Data Science and Advanced Analytics. The internship took place in the area of Data & Analytics of the Department for Internal Audit of Caixa Geral de Depósitos (Portugal), from the 14th of September 2020 to the 13th of June 2021. The internship’s goal was the introduction of machine learning to the Department of Internal Audit. In particular, the implementation of three machine learning pipelines to aid in audit activities of the institution, which systematically analyze operations that stand out from the implemented alarm system. The alarm system triggers alerts when an event disobeys a predefined methodology. Each triggering event is reviewed and processed individually by the auditors, either by being classified as a confirmed error or as a false positive. Confirmed errors frequently lead to recommendations to rectify the operations, while false positives are closed without a recommendation. The alerts’ triggers are defined by sets of arguably general and manually implemented rules, resulting in high trigger frequencies and low precisions. Trigger frequency, precision, and cost of miss rate differ for each alert. Based on the alerts’ trigger history data, three types of alerts were selected for improvements. The deployment of machine learning pipelines with classification models optimized the triggers' specificity while maintaining high sensitivity, which reduced the number of daily events that have to be reviewed by the auditors. This optimization maximizes the efficiency and productivity of the general alarm system and decreases the auditors’ workload.
Author: Crisóstomo, Laura Inês Bleeker Casquinha
Language: English
Open Access status: green
OpenAlex ID: W7027603684
Cites: 13
Cited by:
Related to: 10