Can alert models for fraud protect the elderly clients of a financial institution?

ABSTRACT

Using account-level transaction data at a major financial institution, we predict the incidence of suspicious activity that can be related to the external financial fraud of its elderly clients. The data consists of over 5 million accounts of clients aged 70 years and older, and over 250 million transactions extending from January 2015 to August 2016. Our main focus is to improve the detection of alerts within a proprietorial transaction monitoring system. Using logistic regression, random forest and support vector machine learning techniques, together with corrections for imbalanced alert samples, we provide a new alert model for the protection of elderly clients at a financial institution, with out-of-sample predictive accuracy. Our findings show the relative influence of client traits and account activity in our select external fraud alert models.

KEYWORDS:

• Alert models
• fraud
• elderly clients
• financial institutions
• machine learning

JEL Classifications:

• G21
• G17
• D12
• C55

Acknowledgements

The authors would like to thank for suggestions: Tom Butler, John Byrne, Peter Cowap, Andreas Hoepner, Ahmed Mansoor and Andrew Vivian. We are grateful to the authors’ institution for financial support. Cal Muckley is also grateful for support provided by an Operational Risk industry consortium comprising: Bank of Ireland, Citibank Europe Plc, Deloitte Ireland and Institute of Banking. We are responsible for all remaining errors.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Gaurav Kumar http://orcid.org/0000-0001-9067-4161

Notes

1 In 2015, FINRA (Financial Industry Regulatory Authority) established a helpline for senior citizens concerned that they are being taken advantage of by their broker or investment bank. In the same year NASAA (North American Securities Administrators Association) launched a senior protection website to help combat elderly financial exploitation. This serves as a barometer for the growing level of mistrust between clients and their banks, especially regarding senior citizens, and the efforts being made to combat this trend. In addition, both at the state and federal levels, FINRA and NASAA have been instrumental in empowering and obligating financial planners to combat the financial exploitation of elderly clients.

2 Section 3310 of the Bank Secrecy Act requires financial institutions to develop and implement an anti-money laundering programme, which covers elderly financial exploitation as one of the activities to be monitored. These regulations are implemented by the US Department of the Treasury.

3 The Conduct Costs Project is the first Associated Research Project of the CCP Research Foundation. It builds on the work done at the London School of Economics Conduct Costs Project, with substantially the same team.

4 Section 3310 of the Bank Secrecy Act requires financial institutions to, at a minimum:

1. Establish and implement policies and procedures that can be reasonably expected to detect and report suspicious transactions.

2. Establish and implement policies, procedures, and internal controls reasonably designed to achieve compliance with the Bank Secrecy Act and its implementing regulations.

3. Provide for annual independent testing for compliance to be conducted by member personnel or by a qualified outside party.

4. Designate and identify to FINRA personnel responsible for implementing and monitoring the day-to-day operations and internal controls of the programme and provide prompt notification to FINRA regarding any change in such personnel.

5. Provide ongoing training for appropriate personnel.

5 It should be noted that the alert data had to first be cleaned prior to analysis. This mainly consisted of removing erroneous and duplicated alerts.

6 With regard to our models, we include Flagged Ratio (outflows) and Flagged Ratio (transactions) in our set of predictors at the expense of the more fundamental risk indicators that comprise these two variables, i.e. Aggregated Flagged Outflows, Aggregated Outflows, Number of Flagged Outbound Transactions, Number of Outbound Transactions.

7 C4.5 is an algorithm used to generate decision trees. Typically, C4.5 assigns the proportion of correct counts at the leaf node as the estimated probability. The C4.5 decision tree learner is one of the industry standards for evaluating cost sensitive learning algorithms.

8 Other approaches have been put forward to help address the problem of unbalanced datasets which we include here for completeness but do not make use of in our work. Kang and Cho (2006) apply an ensemble method of under-sampled support vector machines to two synthetic and six real datasets and find that it outperforms other methods, with regard to the G-measure, especially in the case of extreme imbalance. Kubat and Matwin (1997) put forward the idea of selectively under-sampling a representative subset of the majority class, in what they dub One-sided Selection. We leave the use of these alternative approaches for future work.

9 Using solely over sampling of the minority class gives very weak out-of-sample predictive performance in our data.

10 Other metrics commonly used for performance evaluation e.g. Precision, F-measure, Kappa Statistic etc., are not suitable for our purposes.

11 For an imbalanced dataset, the probability cut-off should be chosen so as to incorporate the prevalence of the minority class and the relative cost of misclassification.

12 In comparison to the relatively thorough review received by Issue cases, Second Review cases can be particularly prone to human error in their classification. They can receive a relatively preliminary draft review by fraud analysts. Hence, it is not surprising that our optimal alert model for Issue cases outperforms its counterpart model for Second Review cases. Issue cases are more systematically related to client and account traits than are Second review cases.

13 The Gini index is based on the Gini coefficient which lies between 0 and 1 and measures the class purity of a random forest leaf node. A value close to zero represents a leaf node with an overwhelming majority of one class, whereas a value close to one represents an almost 50:50 split. As a result, a decrease in the sum total of each leaf node's Gini coefficient is a desirable outcome. The Gini index, therefore, measures the relative contributions towards a mean decrease for each predictor.

Additional information

Funding

This project is co-funded by Enterprise Ireland and the European Regional Development Fund, under Ireland’s European Structural and Investment Funds Programs 2014–2020. Cal Muckley would like to acknowledge the financial support of Science Foundation Ireland under Grant Numbers 16/SPP/3347 and 17/SP/5447.