AI and machine learning for risk management
AI and machine learning solutions for risk management are rapidly becoming mainstream. These developments offer excellent opportunities for improving risk management. My co-author, Saqib Aziz, and myself expand on this core idea in our recent article, “AI and Machine Learning for Risk Management”. In this post I emphasise the breadth of advances in developing AI and machine learning solutions across all areas of risk management. This highlights some of the fundamental changes coming to the industry of risk management as well as showcasing some of the most popular machine learning and AI techniques for the main areas of risk management: credit risk, market risk, operational risk, and compliance (or ‘RegTech’).
Credit risk was one of the first beneficiaries of machine learning technologies. To some extent because it needed the most help. Traditional credit scoring was historically confined to financial quantitative measures and regression testing techniques, in order to classify borrowers into good and bad risks. The problem is that there is far more information than just basic financial measures that is potentially useful in informing good lending decisions. In our study we note a company called ZestFinance which is one of a fast-growing group of startups showing the huge potential of machine learning to enable better credit risk management. ZestFinance teamed up with Baidu in 2017 to use customer search engine history and past purchasing data to provide improved lending decisions on Chinese customers wishing to buy from Baidu shopping platforms. The new data, analysed using a suite of clustering, classification, and more advanced machine learning techniques, was able to increase Baidu lending by 150 percent in just two months with no increase in credit losses. The clustering and classification machine learning techniques applied are the drivers behind all manner of useful application – from spam email detection to Amazon recommendations and search engine results. A growing body of research and practical evidence indicates they will also become the mainstays for credit risk management.
The growth in machine learning and AI techniques in market risk management shines a spotlight on a different set of techniques within the machine learning arsenal - those of powerful simulations and reinforcement learning. There is also a bit more of a move towards pure AI in market risk management, if we adopt a definition of AI as automated decision making based on machine learning testing in conjunction with other advanced techniques. ‘Model risk’ – the risk that trading models are not valid or have become invalid due to changed market conditions – is a key risk in market risk management. Popular machine learning services such as yields.io have cropped up offering model validation based on machine learning that involves constantly running millions of simulations of model performance and investigating simulated deviations from expected outcomes. There is also a focus on reinforcement learning; a branch of testing at the cusp between machine learning and AI, due to its focus on designing algorithms to self-learn from experience and make improved decisions over time. This is particularly useful for market trading algorithms that are faced with the fundamental problem of other market participants altering their trading behaviour based on their learning of how the algorithms are making decisions. Up to two-thirds of algorithm trading profits can be lost due to these ‘market impact costs’. Reinforcement learning and other advanced AI algorithms can train themselves to understand how other trader reactions affect their decisions and profitability and thus avoid some of these substantial costs and the risks associated with reliance on algorithms.
In managing operational risk, such as customer and employee fraud, clustering and classification techniques can be useful, but increasingly used are techniques based on network analysis and deep learning. This choice is driven by the complexity inherent in how fraud is committed. If we take something like money laundering, which carries a critical reputational and legal risk for the firm, it will always happen in a clandestine manner. So simple regressions, collecting information via customer documentation, or visual inspections of client data are unlikely to spot money laundering. New techniques being applied from the AI and machine learning toolkit include network analysis (identifying connections and shared behaviours between people), or deep learning (to model hidden connections between input data). These techniques have both shown superior results in operational risk management contexts. Many banks and consortiums of banks have, as a result of the convincing evidence, now started investing heavily in machine learning to address this form of operational risk.
Lastly, we mention RegTech, technology solutions to address compliance issues associated with the risk management function of the financial firm. Here we see the diversity of machine learning and AI techniques once again, as a key focus is on textual analysis and natural language processing (NLP). This has both a cost reduction function for the firm, as machine solutions are increasingly more cost effective than are people at reading the large volumes of regulatory requirements. These solutions can also offer better regulatory understanding, as packages are now available that offer reliable shortened summaries of new regulatory requirements due to the steady improvement in reading comprehension offered by NLP techniques.
This post is just a summary of the potential, and we go into a lot more detail in the article itself. But hopefully it will serve to whet the reader’s appetite to learn more about the potential of this area, especially for those readers who were aware of some of the individual advancements, but not familiar with the overall scale of progress that has been made in making machine learning and AI the new mainstream approaches to risk management.
Michael Dowling, Rennes School of Business