Natural sciences are inspiring ground-breaking simulation tools. Silvia Pavoni looks at what banks can do with them.

Science of risk management

The behaviour of proteins is unusual. The linear sequence of amino acids that forms proteins has no instructions on how to fold into a stable shape. And shape is important as a protein with the wrong form would not work as it should do or, worse, could be toxic.

Understanding this process has implications that go beyond natural science, for the same approach that can help simulate how a protein’s three-dimensional structure develops can also be applied to study how the economy or a banking system changes. Agent-based models, where the interactions of various agents part of a system are observed and their future actions simulated, are one such approach.

The idea that big data and emerging modelling techniques can help manage risk is becoming more prevalent, as is the recognition that risk managers need new approaches and faster tools to deal with uncertainty. “[Models used so far,] in case of crisis, have nothing to say, frankly,” says Justin Lyon, the CEO of Simudyne, a fintech company that creates agent-based modelling software. Its application to banking is still relatively new.

A grain of sand 

Giulia Iori, professor of economics at City University in London, first became interested in agent-based models during her PhD research in a branch of physics called statistical mechanics, for which she dived into the protein-folding phenomenon. She notes how other natural science models are lending inspiration to economics and macroprudential activities, all interested in the concept of equilibrium of systems – from models that study amorphous solids such as glass, which do not have a specific structure that defines them, to the physics concept of ‘self-organised criticality’ used to explain a pile of sand where grains are progressively stacked on top of each and which suddenly collapses.

“Everything is stable for a while, then an additional grain of sand falls and brings down a lot more sand from the [pile],” says Ms Iori. “So the reaction of the system is not proportional to the shock inferred: you let a grain of sand slide many, many times and nothing happens, then an additional one creates a systemic effect.” This means that the system has spontaneously organised itself to reach a state that is very unstable so all it takes is a very small shock to create an avalanche.

“[This is an] idea that we try to use for the banking system too: the idea that a crisis is not necessarily a reaction to a [large] shock but because of the complexity of the system, where banks interact with each other and are exposed to each other, exchange money. This system naturally evolves towards a state that is unstable,” says Ms Iori. “So you may give it a shock and nothing happens, you give it another one and everything falls.” Intuitively, for anyone that has ever operated in or observed the banking sector, this may sound familiar.

Military precision 

Banking supervisors are acutely aware of the challenges of preserving the balance of complex systems. Much as for the natural sciences, risk management, whether at a macro-prudential level or the level of individual banks, needs the attention of dedicated research, which has so far only happened in isolated cases and confined geographical pockets.

“Banking supervision is an area of research that is relatively under-explored by academics and central banks,” says the Bank of England’s chief economist, Andy Haldane. This is a gap the bank is attempting to fill. “We have made promising progress using techniques such as big data, machine learning, semantic algorithms, experimental methods and agent-based models to better understand behaviour in the financial system and to provide new tools to help banking supervisors in their work,” says Mr Haldane.

The US Treasury’s office of financial research started this process early, shaken by the inability of traditional tools to see the fault lines that led to the financial crisis. In 2013 it asked Mitre, a US organisation with nearly 8500 staff and which runs federally funded research projects that span across crucial sectors from national and cyber security to healthcare, to come up with a new solution.

“What we set out to do was to develop a dynamic platform to do stress testing at the system level rather than the level of individual financial institutions, which was really the prevailing approach leading up to and through the financial crisis,” says Brian Tivnan, Mitre’s models and simulation chief engineer, who led the team that developed such a solution for the US Treasury.

Mitre’s own financial application of agent-based modelling derived from the experience in other fields. The organisation’s first such tool was deployed for the military to build a high-resolution model of coalition forces in the counter-insurgency activities in Afghanistan more than a decade ago.

A bank-based approach 

From biology to the military to macroprudential activities, simulation – through agent-based models or otherwise – can find applications within commercial banks too. The software that Simudyne has built is based on Mitre’s intellectual property. The idea behind the partnership is that the more central banks and commercial banks will adopt those models, the safer global banking systems will be, to the benefit of the US too.

Traditional risk models had focused on the value at risk at individual institutions and then, with the introduction of stress testing by banking supervisors after the financial crisis, on taking a picture of how each bank would perform should a series of shocks be introduced in its environment. Those were static, primarily backward-looking methods. The idea now is to allow the risk function to consider how the interactions (rational or otherwise) within the system, connecting counterparties and taking into consideration human behaviours such as panic, could lead to instability and systemic collapses. By incorporating human behaviour into possible scenarios, agent-based models can help understand the impact of policy decisions on markets and the economy, according to Mitre.

And there are applications beyond risk management too, as agent-based modelling simulations can also be used in investment, such as in benchmarking exercises to discover best execution standards to help meet the more stringent requirements imposed by the second reiteration of Europe’s Markets in Financial Instruments Directive (better known as MiFID II).

Agents are simply representations of people, investors, banks or even central banks and these relationships can be calibrated to behaviours that can be observed, according to Mr Lyon, who says: “This means that the likelihood of individual funding counterparties to pull funds can be modelled directly – and the parameters governing this behaviour can be varied to explore a range of other outcomes.” These models can be used both to evaluate the sensitivity of, say, a liquidity risk scenario to a number of alternative behavioural assumptions as well as to perform reverse stress testing, where risk managers can understand the way in which behaviour would have to change for a bank to get into trouble.

Commercial interests 

How does a protein fold into a functional shape is not too dissimilar a question to how does any other system of interconnected agents move in time.

Commercial banks’ interest in simulation is growing too. Barclays, for example, was one of Simudyne’s first clients. Barclays’ chief risk officer, CS Venkatakrishnan, says that although the agent-based modelling project at the bank is at the early research stage, it could have interesting uses in capturing new risks and emergent phenomena “including, for example, modelling Brexit scenarios”. He adds that agent-based models complement other modelling techniques because they do no rely explicitly on historical data, in the way that typical risk models do, and this makes them suited to analysing ‘what-if’ scenarios. They are able “to model complex systems such as the financial network and to capture feedback loops and contagion risk, which is difficult to achieve in more established methods”, he says.

Big data and the use of new sets of data can help risk managers in other areas. Looking at individual banks’ vulnerabilities, for example, unstructured data such as social media interactions or personal habits captured by mobile phone data have proved useful indicators of customers’ reliability and character, according to Evgueni Ivantsov, chairman of the European Risk Management Council. “Some people say this [provides] better predictions than traditional methods,” he says, which rely on structured data such as financial accounts, typically used in credit-worthiness assessments.

For example, the time taken to respond to a mobile message and text content analysis might indicate certain personal characteristics – not responding promptly or answering in a manner that could be interpreted as clumsy might suggest that that individual could behave similarly in its relations with the bank. Or a customer who gets around to recharging their mobile phone only when the battery goes completely flat might not be organised enough to repay a loan in time should a problem occur. The link between mobile phone habits and character might raise eyebrows; and the invasion of privacy that such analytics are based on would likely cause offence to some customers. Mr Ivantsov points out that these tools have found easier applications in Asia, where privacy issues are generally less of a concern.

With added complexity 

Every new solution requires a whole new set of considerations. Na Zhou, a principal at consultancy Oliver Wyman, who has looked into the use of data science in banking, says: “Thanks to alternative data you can build a more fine-grain picture of your customers and your customers’ customers and the interaction between them. This is enabled by both data availability and computation power. But with added opportunities comes added complexity: questions over whether data has been verified at source; whether it has been cleaned; privacy considerations; the legal foundation on which this data can be used; and European legislation such as GDPR.”

With all the challenges that banks face, UBS’s head of risk methodology and UK chief risk officer, Rahul Dhumale, agrees on the importance of simulation techniques. In the Swiss bank’s case, artificial intelligence has helped identify new types of risks by quickly analysing large numbers of reports written about a specific subject or the bank itself and flushing out which specific concerns are raised, for example. Or to quickly pick up on the growth of products that proved problematic in the past, such as banks re-entering the securitisation market after its seeming disappearance post-financial crisis.

Agent-based modelling can also help address so-called ‘fat-tail’ risks – events that have low frequency but high impact, where historical data is scarce or non-existent and traditional methodologies ineffective, according to Simudyne’s Mr Lyon. These include cyber attacks, where hackers’ success rates are low, despite the many thousands of attacks a bank might suffer daily, but consequences of a breach can be dramatic. Climate change-related risks are also part of this group, and even less well understood and accounted for than a more immediate concern such as cyber security.

“The scope of risk and the scrutiny that risk management teams are under have both greatly increased, particularly for non-financial risks,” says Thomas Garside, a partner at Oliver Wyman who specialises in risk management. “We see a focus on conduct, compliance and culture, on operational resilience, of which cyber risk is a part, and then emerging risks such as climate.”

Speed and clarity of response is important too. At UBS, Mr Dhumale’s focus has been on being able to report to the board of directors and senior management, in “plain English” and quickly, the way an external event would impact the bank. So the focus is on both being able to improve the accuracy of predictions but also to keeping an eye on the bigger picture and on the purpose those predictions serve.

“We’ve been looking at [how to deliver this] since 2008, since the financial crisis and all other crises since: the eurozone sovereign crisis, what is happening currently with the US/China trade issues or the US not agreeing in Congress on the border wall,” says Mr Dhumale. “For all these things, stakeholders need answers quickly, and answers that are directionally correct.” This, he adds, is in contrast to the traditional, lengthy scenario analysis that often ended up being a ‘tick-box’ exercise to satisfy regulators.

Tick-box exercises 

The tick-box exercise has in some cases proven ineffective if not counterproductive, according to Giampaolo Gabbi, affiliate professor of risk management and director of financial institution custom programmes at Bocconi University’s management school in Milan. He believes that the regulatory requirements aimed at guaranteeing enough capital is available to absorb potential losses introduced after the financial crisis, along with the freedom to present internal evaluations – rather than standardised models – for this purpose, created a situation by which chief risk officers had to work on two distinct fronts: models for the regulators and models for day-to-day management. The two groups of models did not always converge.

“This objective [freeing up capital] has very much influenced risk managers but this is not a risk management objective, it’s a capital management objective,” says Mr Gabbi. “This contradiction between risk and capital management has made internal models not always reliable.” 

The discussed introduction of a standardised floor on capital requirements as part of the review of internal models of Basel III (and its subsequent changes, often referred to as ‘Basel IV’) might redirect focus away from the creation of internal regulatory models, reducing flexibility but closing the gap with  day-to-day risk management models, according to Mr Gabbi; this could “avoid that, internally, a bank had the ‘right’ model, more oriented on risks, [while] the one that it asks supervisors to validate is more capital management oriented”.

Luigi De Sanctis, a partner at Oliver Wyman agrees that there is a push for internal convergence of risk models and simulation efforts and this goes beyond capital requirements needs. “There is a strong push from supervisors, investors, senior management and boards of directors to build capabilities to develop scenario-based simulation tools that are consistent across the organisation. Traditionally banks have been running several simulation exercises but these were not necessarily consistent with each other,” he says. Mr De Sanctis refers in particular to scenario-based projection tools used by the treasury and risk functions, which use different methodologies, definitions and data sources and which therefore would produce different reactions to the same scenario. Looking at a bank as a whole, this is “nonsense”, he says.

This research into a better understanding of what threatens the stability of a complex banking system is taking place alongside the higher regard for the risk management function within individual banks. Chief risk officers have moved up in management charts since the financial crisis and they have larger teams and stronger powers to escalate issues to the board. 

A unifying effect? 

There is a sense that simpler and faster simulation tools might bring risk and other departments closer together. UBS’s Mr Dhumale believes that working with the front-office staff to provide better-tuned risk tools that help them in their job is important and that, ultimately, that should also lead to a better understanding of the risks a bank is exposed to, to the benefit of that bank as well as its supervisor. But it is a matter of listening to other functions as much as being listened to.

“Chief risk officers are listened to more, absolutely,” says Mr De Sanctis. “Are they listened to as much as they should? Probably not in my view, and it varies a lot between institutions. In general, I observe a positive correlation between the strength and quality of the risk management function and the long-term performance of a bank.”

As a 2017 Oliver Wyman report on the future of risk management pointed out, banks continue to operate in very uncertain political and social environments across the world and there is “almost no tolerance” for misconduct. On the other hand, investors will likely increase pressure to gain higher returns, which have often disappointed in past years. How risk managers can and are allowed to do their job makes a difference not only in the way the banking system keeps its equilibrium, but also on individual banks’ reputation and bottom line.

“In risk management there are hard elements – models and data – and soft elements – culture, governance, how responsibly people use models, [whether] they understand [models’] limitations,” says Mr Ivantsov at the European Risk Management Council. “You need to be very precise about how to use both elements. [If these were the only two options] rather than having good models run by bad people, it’d be better to have bad models run by good people.”

The challenge for banks now is to make sure they are abreast of all the new techniques in risk management and how science can be used effectively to model risks that the old systems may not have picked up. 

PLEASE ENTER YOUR DETAILS TO WATCH THIS VIDEO

All fields are mandatory

The Banker is a service from the Financial Times. The Financial Times Ltd takes your privacy seriously.

Choose how you want us to contact you.

Invites and Offers from The Banker

Receive exclusive personalised event invitations, carefully curated offers and promotions from The Banker



For more information about how we use your data, please refer to our privacy and cookie policies.

Terms and conditions

Join our community

The Banker on Twitter