Over the decades, crises and subsequent regulatory change has transformed the risk-led financial services sector. However, financial organisations also have had reputations as early technology adopters, some quite “wild west” in their quest for differentiation and profitability. On one hand, the industry prioritises risk mitigation to ensure financial services firms are well-capitalised and compliant and hence our savings, pensions and loans are deemed safe, but on the other hand it encourages risk-taking to drive product differentiation, enhance returns and ultimately drive wealth creation. Right at the intersection in 2017, we find machine learning.
Why machine learning now?
Machine learning applies “intelligence” to discover useful patterns in complex data sets, sometimes with predetermined knowledge, and other times not. Models adapt to changing data. Machine learning mathematical methodologies, have been around for several decades, but as data sets have become larger, storage of them easier, and computer power cheaper, data and computer-intensive processes that aid machine learning methods have become more affordable and impactful.
In finance, machine learning helps identify impactful factors that effect a product’s performance or risk, predict prices, detect anomalies for fraud, and in other cases automate and mimic the skills of allegedly clever strategic investors. Machine learning technologies are not unique to finance, helping bring transformation to robotics, medical and vision processing industries among others.
In financial services, machine learning is becoming more popular, but not yet ubiquitous. MathWorks surveyed financial quantitative professionals last year and found machine learning use has quadrupled compared to a similar group we surveyed in 2014; yet 60% of respondents are still not currently using machine learning. The study also revealed that nearly half of those questioned had the intent to apply or test the technology in the next year. It is therefore becoming another quantitative tool in the financial services industry’s broad tool-box, set to address challenging problems and build exciting new use cases.
A risk-sensitive uptake
However, there is caution, uncertainty and delay among many. First, the methodology depends on true data, increasingly imported from novel but perhaps unvalidated data sources which might underpin new features in an investment strategy. For others, more sceptical, why complicate? Simpler models can and do work equally well, often with less expense and with greater transparency and reproducibility, and it’s easier to intervene at points of regime change, i.e. when a market performs one way today but another tomorrow due to a significant unanticipated event. In addition, financial services industry participants and clients are not always comfortable with its use, though of course others are excited by it.
In the risk industry, emerging regtech innovation both excites and worries. “Black box” machine learning prediction or classification may offer accuracy in risk forecasting, for example, of default scenarios and rating methodologies in the credit industries, but it can also be something of an unreproducible, un-transparent “black box”. Model (in)transparency was a problem identified during the global financial crisis – overcomplicated, obtuse, unvalidated models developed in silos causing systemic impacts beyond their remit – and hence why regulators care about it. Accuracy may not actually be accurate, particularly in the high dimensional data universe in which finance operates. Correlation is easy to infer, but it may not mean causality, and when machine learning fails to distinguish one from the other, overfitting occurs and predictions fail. What is true one day can be untrue and costly the next, leading to business losses and in some cases regulatory penalties.
However, machine learning also automates and adapts well, extending credit risk model shelf lives and adding insight. It can also apply objectivity to “model selection”, helping determine the right model features logically and objectively, challenging your preferred model because it might have been your PhD speciality or the vogue in the latest text book. Machine learning can take a fresh, automated look at your data, feature and model selection, and help mitigate risk as part of good model governance and model risk management. In the emerging risk segments too – e.g. anti-money laundering (AML) and know-your-customer (KYC) applications - classification, anomaly detection and predictive capabilities of machine learning technologies are truly driving analytics.
At its very best, machine learning can improve risk model efficacy, efficiency and longevity, and support good model governance, but participants among regulators, regulated and non-regulated firms need to appreciate methodology risks too. Regtech firms can help experiment to find that sweet spot, helping support and challenge risk-sensitive larger institutions, but we must also be sensitive to regulators and supervisor priorities which may override.
Ultimately, a sound understanding of both data and methodology helps enable effective adoption and avoid risky pitfalls. Use with care, know your data and know your model just as regulators encourage you to know your customer. Machine learning will support, not replace, traditional statistics and equation-based modelling, including in risk management.