Fraud Analytics: Perspectives from the Austrian School of Economics

Imagine alien scientists observing New York’s Grand Central Station. They would notice people entering and exiting metal boxes, and possibly could predict ebbs and flows of activity. However, without access to human meanings, they would not have a complete understanding of the phenomenon, namely that people are using a mass-transit system to commute between home and work.

– Ludwig von Mises

Ludwig von Mises – one of the major figures of the Austrian School of Economics – often used the above parable in his class to demonstrate the difference between two fundamental ways of analyzing human behavior [1, 2]. These are: 1) objective method, where one would observe empirical events, e.g., people rushing back and forth, aimlessly at certain predictable times of day; and 2) teleological method, in which all human behavior is purposive, and one can see the motive is to get from home to the train to work in the morning and the opposite at night.

Mises built the Austrian School of Economics on the tenet that human action is always purposefully directed at achieving chosen ends. This approach clearly hinges on understanding of causal relationships between means and ends.

In recent times, data science armed with black box machine learning models and inexpensive computing power has been hurtling on the tracks of the objective approach. With the advent of big data, automated algorithms routinely scan millions of observations mechanically searching for patterns, correlated to solutions of business problems. Nowhere is the contrast between teleological and objective methods more pronounced, than in the world of fraud analytics.

To illustrate, say data reveals ATM fraud transactions occur in greater frequency around midnight. This information is sufficient to use transaction time as a feature in models. The teleological approach would be to question why this is the case. An explanation could be that fraudsters look to exploit midnight refreshments of daily withdrawal limits. Can we claim that one approach is better than the other? To answer that, let us consider another example.

Say we observe high fraud rates in credit card transactions where ticket size exceeds $10,000. Consequently, it suffices to use high transaction amounts as a red flag. However, the danger lies in the fact that, post deployment, fraudsters might adapt to the solution and attempt transactions of lower amounts. This implies patterns in transaction amount are spuriously correlated to criminal behavior – there is no underlying causality. Noncausal features make it easy for criminals to slip through the cracks of counter-fraud initiatives.

Thus, a lack of investigation into the purposefulness of empirically observed criminal behavior renders solutions vulnerable to shifts in fraud modus operandi. Consequently, model performance rapidly deteriorates – the solution needs to be refreshed frequently to catch up with highly dynamic mutations in fraud trends.

Identification of robust, causal relationships at the feature engineering stage of model building thus helps to enhance shelf life of solutions by making it difficult for criminals to adapt to fraud detection measures. Principles of Austrian economics can also add value to fraud loss forecasts.

Difficult to Predict

Fraud losses are notoriously difficult to predict, as they are dependent on dynamic interplay of internal controls and a plethora of external (often hidden) factors. Extrapolation of historical figures does not result in accurate forecasts. The Austrian School believes that each individual, in buying or not buying and in selling or not selling, contributes his share to the formation of market prices. Prices are thus a composite effect of human interactions.

The fraud ecosystem harbors similar interactions, with criminals chasing undetected fraud activity, while banks strive to make it tougher for them to achieve this. The “price” of successful fraud activity is the amount of cognitive effort criminals need to expend in order to sneak past any bank’s counter-fraud infrastructure. As the level of fraud fighting sophistication varies across banks, so does the ease of committing fraud. To the fraudster, there exists a menu of different banks to choose from, each at a different “price.”

This “price” follows patterns similar to business cycles. If a bank does not relentlessly invest in enhancing its infrastructure, it gets progressively easier for criminals to commit fraud. This implies a gradual drop in “price” and results in an uptick in fraud loss. Increased loss motivates banks to develop better solutions, which increase required cognitive effort, and the cycle continues.

Holistic understanding of this cyclical nature can help in accurate forecasting of fraud losses. Grasping the purposiveness of criminal behavior as chasing gains along the path of least resistance can help to shed light on nonlinearity of fraud activity. The teleological approach can also aid in evaluating models – a solution with counterintuitive features would raise an alarm and mobilize need for further scrutiny prior to implementation.

Perspectives from the Austrian School of Economics can add immense value as data scientists try to design solutions to counter adaptive adversaries in the fraud ecosystem.

References

1. “Money, Method, and the Market Process: Essays by Ludwig von Mises,” 1990, edited by Richard M. Ebeling, Mises Institute, https://mises.org/library/money-method-and-market-process.
2. “The Oxford Handbook of Austrian Economics,” 2015, edited by Peter Boettke and Christopher J. Coyne, Oxford Handbooks Online, https://www.oxfordhandbooks.com/view/10.1093/oxfordhb/9780199811762.001.0001/oxfordhb-9780199811762.