Promoting Fairness and Exploring Algorithmic Discrimination in Financial Decision Making Through Explainable Artificial Intelligence

In this white paper a comprehensive toolbox is developed, grounded in an ethical “rights to
explanation” framework, deploying state-of-the-art machine learning/artificial intelligence models,
through the lens of explainability.

Harnessing these explainable artificial intelligence algorithms within the toolbox, we propose implementing an ensemble of model-agnostic techniques, to improve fairness in financial decision making, with a particular focus on US home mortgage loan applications with a granular public dataset.

We also highlight variability in these techniques, imposing various pragmatic scenarios that explore real-world decision making, alongside equality of opportunity and equality of outcome conditions. We highlight potential pitfalls, nuances, and possible innovations in applying these techniques, while providing the ability to simultaneously assess the impact of any specific variable in decision making, and a model’s performance in such decision making, with established machine learning criteria.

Furthermore, we showcase the trade-off between fairness and model performance optimization with a protected characteristic (age) that might form the basis of plausibly discriminatory practices in such a context. Our study aims to be in the spirit of Agarwal, Muckley, & Neelakantan (2023), Kelley, Ovchinnikov, Hardoon, & Heinrich, (2022), Kozodoi, Jacob, & Lessmann (2022), and Kim & Routledge (2022), among others. We lastly identify areas for future research.

Fairness and Discrimination in Lending Decisions: Multiple Protected Characteristics Analysis

We build upon the comprehensive toolbox developed in Jain, Bowden and Cummins
(2024), extending its applicability to multiple protected characteristics.

We explore a way in which several characteristics can be simultaneously considered for multi-dimensional fairness promotion and potential mitigation of plausibly discriminatory practices. In the spirit of Jain, Bowden and Cummins (2024), once again we do this with a particular focus on US home
mortgage loan applications with a granular public dataset.

Finally, we address a prior deficiency, namely a worse overall model accuracy/performance as measured by Area Under the Curve (AUC). The improved AUC can be attributed to a better True Positive Rate of correctly classified loan acceptances, which is achieved with the aid of hyperparameter tuning.

Specifically, we use Stratified K-Fold Cross-Validation combined with overfitting- robust hyperparameter tuning facilitated with the aid of a Grid Search. These were discussed but not explicitly implemented in the use case of Jain, Bowden and Cummins (2024). We document that even a narrow set and range of hyperparameters (mitigating the computational cost of employing the Grid Search) is sufficient to elicit these improvements.

Lastly, we provide recommendations on the implications of our results including where a
human-in-the-loop

Enhancing Financial Crime Detection By Implementing End-to-end AI Frameworks

Economic crime, encompassing money laundering, fraud, scams, and various other
illegal financial activities, continues to evolve with the emergence of sophisticated Artificial
Intelligence (AI) technologies.

This white paper explores the dual-edged nature of AI in the financial sector. While AI tools are increasingly being exploited by criminals to commit financial crimes, they also hold the key to more robust and effective detection and prevention strategies.

This paper delves into the array of AI techniques currently leveraged by malicious criminals, including deepfake technologies, phishing and spear phishing, automated social engineering, credential stuffing, synthetic identity fraud and others.

Furthermore, it provides a comprehensive analysis of AI techniques capable of countering
these threats. Key focus areas include Neural Networks for unusual patterns and behaviours,
gradient boosting algorithms for risk assessment, reinforcement learning for optimisation of
decision making, Markov chains for temporal patterns and anomalies over time, Naïve Bayes
for real-time classification and decision trees for interpretable detection.

The culmination of this paper is the presentation of a state-of-the-art end-to-end AI-driven solution that integrates AI technology to offer a holistic and dynamically adaptable approach to financial crime detection and prevention. By implementing this framework, financial institutions can significantly enhance their capabilities to identify, mitigate, and prevent financial crimes, ensuring a more secure financial ecosystem.

Using Automation and AI toCombat Money Laundering

Money laundering, which is the criminal activity of processing criminal proceeds to disguise their origin is one of the gravest problems faced by the global economy, and its size is growing rapidly. It is estimated that 2- 5% of the global GDP or US$800 billion to US$2 trillion is being laundered every year across the globe.

Banks have begun to understand that their legacy rules-based systems cannot effectively mitigate risks related to money laundering. There is a need to embrace advanced technology that can effectively solve their problems of getting involved in money laundering cases. This white paper outlines a case study focusing on the effectiveness and limitations of Artificial Intelligence (AI) in detecting and preventing money laundering activities. It will provide an overview of the design, architecture, implementation, and testing of such a strategy.

ESG Greenwashing And Applications of AI For Measurement

“ESG greenwashing” refers to the strategic communication tactics firms use to
selectively disclose their ESG conduct to stakeholders.

ESG greenwashing strategy, while it may attract and satisfy stakeholders at the beginning, may cause different issues for firms later, such as adverse publicity, lobbying, or boycott campaigns by consumer or pressure groups or divestment by socially responsible investors. The complex impacts of ESG
greenwashing underscore the imperative of discerning and quantifying instances of such practices. We aim to consolidate recent literature reviews of ESG greenwashing, methodologies to measure ESG greenwashing and developing applications of AI, text analysis and machine learning models to advance such measurement.

This white paper makes significant contributions to policy developments, such as the greenwashing regulations of the UK FCA and the European Parliament.

Simplifying Compliance through Explainable Intelligent Automation

We discuss how explainability in AI-systems can deliver transparency and build trust
towards greater adoption of automation to support financial regulation compliance among
banks and financial services firms.

We uniquely propose the concept of Explainable Intelligent Automation as the next generation of Intelligent Automation. Explainable Intelligent Automation seeks to leverage emerging innovations in the area of Explainable Artificial Intelligence. AI systems underlying Intelligent Automation bring considerable advantages to the task of automating compliance processes. A barrier to AI adoption though is the black-box nature of the machine learning techniques delivering the outcomes, which is exacerbated by the pursuit of increasingly complex frameworks, such as deep learning, in the delivery of performance accuracy.

Through articulating the business value of Robotic Process Automation
and Intelligent Automation, we consider the potential for Explainable Intelligent Automation
to add value. The solution framework sets out the Explainable Intelligent Automation
framework, as the interface of Robotic Process Automation, Business Process Management
and Explainable Artificial Intelligence. We discuss key considerations of an organisation in
terms of setting strategic priorities around the explainability of AI systems, the technical
considerations in Explainable Artificial Intelligence analytics, and the imperative to evaluate
explanations.

Explainable AI For Financial Risk Management

We overview the opportunities that Explainable AI (XAI) offer to enhance financial risk
management practice, which feeds into the objective of simplifying compliance for banking and
financial services organisations. We provide a clear problem statement, which makes the case for
explainability around AI systems from the business and the regulatory perspective.

A comprehensive literature review positions the study and informs the solution framework proposed. The solution framework sets out the key considerations of an organisation in terms of setting strategic priorities around the explainability of AI systems, the institution of appropriate model governance structures, the technical considerations in XAI analytics, and the imperative to evaluate explanations.

The use case demonstration brings the XAI discussion to life through an application to AI based credit risk management, with focus on credit default prediction.