The Role of Artificial Intelligence and Machine Learning in Risk Management

Lending in China is a risky proposition. When a U.S. bank needs to decide whether to approve a loan to a U.S. customer, it simply accesses the customer’s credit report, which is often the deciding factor. The bank can thus reasonably manage its credit risks based on the historic default rates for the lending categories it specializes in.

In China, however, about 80 percent of potential borrowers have no credit record. That has left lenders with two approaches to credit risk: concentrate on lending to the highly sought-after 20 percent of borrowers with credit records and live with less profitable loans and an over-concentrated lending portfolio, or take a chance on the remaining 80 percent.

There may, however, be a third possibility.  ZestFinance[1] is one of a large number of start-ups using artificial intelligence, or AI, to control the risk of lending to anyone in China. The company ran an experiment in 2017 with Baidu, the Chinese version of Google and Amazon combined, that over two months allowed Baidu to increase lending 150 percent without increasing credit risk.

Behind this promising result was a barrage of machine learning tests run, with the permission of customers, on Baidu’s extensive cache of customer data. The process included checking whether customers’ buying history squared with their reported incomes, but also whether their search-engine histories revealed other relevant information. Decisions on whether to lend were made in seconds.

That’s a snapshot of the potential of AI-driven risk management that we address in a recent article, “AI and Machine Learning for Risk Management. We examine how AI and machine learning are managing risk through data and other novel methods and replacing traditional reliance on static numerical data and limited statistical tests.

Most of these new techniques are grouped under the rubric of “machine learning:” drawing inferences from data rather than a statistical model. To some extent, machine learning allows the model to emerge from the data rather than the other way around. Machine learning can use any type of data, be they numbers, text, or images. The main techniques are classification and clustering: assigning people, events, or companies to groups on the assumption that similar groups act in similar ways. Decision trees and support vector models are common, but regression, a mainstay of traditional statistics, also plays a role, after being adjusted to handle large numbers of explanatory variables. The most useful tool, however, may be neural networks – or deep learning – which is designed to allow hidden connections between explanatory variables to be formulated and reformulated to better predict outcomes.

AI is often confused with machine learning but is actually far more advanced. It builds on machine learning along with other techniques and tries to automate the full process, from data selection to a final lending decision. AI attempts to mimic and then to surpass human intelligence in decision making. AI is relatively rare in risk management, mostly because of a lack of technological expertise, but also because true AI carries its own risks that would have to be managed and justified to often skeptical regulators.

AI and machine learning are having a major impact on managing risk, especially credit risk, market risk, operational risk, and compliance.

Their potential contributions to reducing credit risk are evident from the example of ZestFinance. AI and machine learning may also be useful in managing market risk, and especially trading-model risk. Trading models tend to work initially but then to go awry. Identifying the point at which the market turns against a model is critical, yet it is not always clear whether a model breaks down because of temporary or permanent market changes, and traditional testing techniques don’t provide a reliable answer. That has created opportunities for services such as, which use machine learning to monitor, validate, and adjust trading models, constantly checking millions of variables to identify errant testing results that might warrant further investigation.

Operational risk is often harder to manage than financial risk, given that it involves human decision-making. AI and machine learning can help by handling atypical data – textual descriptions of transactions, network relationships, phone and messaging conversations – and have proven effective in detecting money laundering and fraud. For example, the Nordic KYC Utility, a creation of five major Scandinavian banks, uses a range of machine learning techniques, together with traditional detection tools, to ensure that banks fulfil their “know your customer” anti-money laundering obligations.

Another promising area for AI is regulatory compliance, where the term RegTech has come to include AI developments. One of the biggest players in this field is IBM, which makes use of its Watson expertise and has shown how important major tech companies are becoming to the effort. AI and machine learning can use natural language processing to detect regulatory non-compliance and to read and interpret new regulations. They can also help detect fraud by interpreting conversations between employees.

There are, however, challenges. First, risk management departments often don’t have enough personnel trained in applying these techniques. Second, older firms are generally not set up for the data sharing needs of AI and machine learning. These techniques need effortless sharing and storage of data in a uniform manner across the firm, and many companies keep data in silos and on separate systems.

Third, and perhaps most serious, AI and machine learning might themselves create risk for companies and even economies. Some machine learning techniques, like deep learning, are, at this early stage, black boxes in terms of how they arrive at conclusions. There is also the issue of fairness. All machine learning systems used in the U.S. and Europe for credit and lending have hard coded rules designed to prevent racial and other types of discrimination. However, no system is fool proof, especially considering the amount of data that these systems use.

There has, however, been substantial progress in overcoming these challenges, and it will undoubtedly continue as an enormous amount of investment pours into the field. The future of AI and machine learning for risk management is, in a word, bright.


[1] The authors have no connection with ZestFinance or any other companies mentioned in this post.

This post comes to us from professors Saqib Aziz and Michael Dowling at Rennes School of Business, France. It is based on their recent article, “AI and Machine Learning for Risk Management,” available here.