From Theory to Practice with Agent-Based Modeling

In a March press conference, Federal Reserve Board chair Jerome Powell said that U.S. economic fundamentals were strong and financial stability vulnerabilities were not at high levels. There was no indication of any financial crisis.

For those who care to be vigilant, however, signs of a future downturn may be perceptible in market players and their response to a variety of market forces such as the rise of corporate debt, cybersecurity events, or geopolitical and trade tensions. Simudyne has a model suited to the task of identifying the most likely crisis scenarios.

A London‐based fintech venture, Simudyne says that it is offering a commercial version of a financial crisis model developed by the U.S. Treasury's Office of Financial Research (OFR) that would be of particular interest to central banks and private-sector banks, as well as sovereign and pension funds, and specifically to the risk managers at all such entities.

“It represents a multimillion‐dollar investment by the U.S. government to build a robust model of the U.S. economy that can be used for analyzing indications of financial vulnerability,” says Simudyne co-founder and CEO Justin Lyon.

The firm announced its licensing rights to the model late last year and has started work with Barclays in the training and employment of the Simudyne simulation platform as a crisis and general risk management and decision-making tool.

On April 8 this year, Simudyne announced the closing of a $6 million Series A funding round, led by Barclays, bringing its total financing to $10 million.

OFR Partner

According to Lyon, Simudyne obtained non‐exclusive rights to the model from the federally funded R&D non‐profit Mitre Corp., which had partnered with the OFR on the programming.

Brian Tivnan, chief engineer, Modeling & Simulation Center at Mitre, says the model was developed after the financial crisis “in response to the regulatory agencies looking for next-generation tools for assessing systemic risk and for assessing the stability or lack thereof of the financial system.”

Lyon adds, “Any global bank that has operations in the Americas and wants to understand what scenarios might tip markets over again, what might cause asset prices to decline, and what an emerging crisis might look like, can save themselves five years' worth of work” by making use of the simulation model.

Lyon explains that Simudyne gained this capability through adoption and integration of an agent-based model (ABM) developed under the guidance of the U.S. Treasury.

ABM simulations - sometimes referred to as multi-agent systems or systems for modeling behavior - can be powered by artificial intelligence. They are useful for analyzing human and institutional behavior in complex worlds, environments, or markets, testing a multitude of scenarios, and enhancing decision-making.

Bottom‐Up Development

ABMs differ from traditional risk models that might rely on historical information and employ top‐down assumptions such as the belief that all markets are efficient. Instead, the model‐building approach is bottom‐up, utilizing software agents and, in the instance of analyzing financial markets, agents that represent investors, banks, hedge funds, providers of funding and market makers whose actions would come into play during a financial crisis.

The agents are assigned rules of behavior that reflect how different financial market participants might operate, change and respond during times of crisis and in turn, alter the environment around them.

As Richard Bookstaber, an author and expert in agent‐based modeling, describes it, “The result is a dynamic that can lead to complex outcomes that often do not seem related to the actions of the individuals and from which a crisis emerges.”

By running multiple versions of what-if crisis scenarios employing representative agents, risk managers can learn from complex group behavior that can in turn produce feedback loops, fire sales, funding runs, liquidity spirals, contagion risks and panics, making note of all emergent trends and system shocks.

Over time, a kind of heat map of a potential financial crisis is created that can guide risk managers or regulators in their decision-making.

“Economic theory can't do the job when it is asked to deal with crisis,” says Bookstaber, who is convinced that agent‐based models can.

And ABMs are not just a tool for financial‐crisis simulations. They have been applied in the military for war game analysis, in health care for analyzing how infectious diseases spread, as well as in such fields as transportation, the energy sector and environmental sciences.

Only now, Lyon says, is there easier access to massive computing power, cloud computing and artificial intelligence technology, spurring wider acceptance and adoption of ABMs. (See Agent-based Computational Economics)

Influential Writings

Lyon credits Bookstaber with the growing awareness of agent‐based modeling in the financial sector. A former risk manager for Wall Street and hedge fund firms, Bookstaber wrote A Demon of Our Own Design (2007) and The End of Theory: Financial Crises, the Failure of Economics and the Sweep of Human Interaction (2017).

Bookstaber touched on ABM in several OFR working papers, starting in 2012 with Using Agent‐Based Models for Analyzing Threats to Financial Stability. He also authored Agent‐Based Models for Financial Crises for the Annual Review of Financial Economics (2017).

An Agent‐based Model for Financial Vulnerability, a 2014 OFR paper co‐written by Bookstaber, Mark Paddrick of the OFR and Brian Tivnan of Mitre Corp., guided the Mitre technology team on the programming of a crisis model specific to the needs of the financial sector and ultimately integrated into the Simudyne platform.

Bookstaber is currently chief risk officer in the University of California's Office of the Chief Investment Officer, overseeing a $120 billion investment portfolio and continuing to develop ABMs - specifically, one designed to meet the needs of pension funds and asset management firms.

“I think we may commercialize it down the road,” Bookstaber says. He does not have a commercial interest in Simudyne, though Lyon says that it was Bookstaber who guided him to connect with Mitre regarding access to the ABM crisis model.

Beyond Traditional Modeling

In his writings, Bookstaber has described agent‐based modeling as a tool to help overcome the limitations of more traditional risk modeling. He views ABMs as a form of “Risk Management 3.0,” which in concert with other models can provide “a better understanding of how the crisis will unfold, where the dominoes will fall, and how bad it might become.”

Lyon of Simudyne is equally enthusiastic, while highlighting the multiple uses of ABMs and, in turn, his commercial offering.

Additional applications include design and execution improvement of trading algorithms. “Under MIFID II, it is incumbent that trading algorithms not contribute to a disorderly market, and so, ABM simulations can help,” Lyon says.

The models can also be used to test future regulatory scenarios and weed out instances of fraud and money laundering. “Simulations can help create data sets that can then train algos to be much more robust, allowing them to detect a wider range of criminal behavior,” Lyon adds.

Bookstaber has written about how ABMs can be used in policy analysis; to assess data needs, demonstrating the value of new data sources and motivating data acquisitions; and as a tool to encourage liquidity providers in volatile financial markets, helping investors assess investment opportunities that may emerge in the face of fire sales.

Available Platforms

Of course, challenges come into play when turning to ABM technology for crisis or risk management. One is finding the most appropriate modeling platform.

There are currently a host of open‐source ABM platforms and libraries, including BehaviorComposer, NetLogo, Repast, StarLogo, and Swarm Simulator. Commercial offerings, aside from Simudyne, come from firms such as AnyLogic, EntendSIM, Improbable and Vensim.

According to Lyon, a commercial offering has the advantage of enterprise-wide capabilities and security features, critical for financial firms. “Many of the open-source projects that are out there have not been updated for many years, are Java-based, and don't scale to meet the needs of a production environment in a bank,” the Simudyne CEO says.

Other challenges relate to talent: “It is still very hard to find the right people, who can work technically with agent-based modeling. The numbers are still very low,” says Anand Rao, a PwC Advisory partner and innovation lead for U.S. analytics.

Rao says that the use of ABMs by risk managers often requires a change of mindset, as it does not result in a simple mathematical equation. Reaping benefits from agent-based modeling can take a lot of patience on the part of users and can require processing a great many scenarios before acquiring useful data.

The PwC consultant finds that when an ABM indicates emergent behavior, some clients are reluctant to accept the unexpected. “Some of them will say, 'I don't believe it,' when what they should say is, 'We will test it in the market.' What we are finding is that the level of experimentation that is required to reap all the benefits is not there yet.”

Questioning and Challenging

Lyon notes that unlike other methodologies, the freedom of choice that comes with the design of ABMs and the calibration required can be stressful. The modeler must carefully consider, “What do you keep in the model?” and “What do you keep out?”

For some, it is much easier to take a more statistically-based risk management approach, even though those statistics may fail to identify emergent risks. Lyon points out that Simudyne has developed calibration tools to help users refine the accuracy of their models.

He adds that users need to understand that ABMs “are not a panacea or crystal ball, but rather an addition to the risk manager's tool box.” He notes that Andrew Haldane, chief economist of the Bank of England, has referred to ABMs as “challenger models,” a new and different way to look at risk and a complement to traditional approaches. (See An Interdisciplinary Model for Macroeconomics [2017] by Haldane and Arthur Turrell)

Bookstaber says it is important for the ABM user to have expertise in the relevant subject matter and understand how all the components of a given financial market or environment - be they hedge funds, banks or money market funds - are related and would tend to interact.

“If you don't structure an ABM correctly, you might do a thousand runs, but it won't do you any good because you don't have good or accurate representation of all the participants,” he explains. “You have to constantly update for the changing realities of the markets.”

Whatever the challenges, Simudyne's Lyon is optimistic about ABMs' future. He said in a March Traders Magazine interview: “Ultimately, simulation is going to become the gold standard of risk modeling, and those leveraging its potential will benefit from the competitive edge provided by being able to test drive their decisions, fail fast without consequences and create solutions that drive growth.”

Katherine Heires is a freelance business journalist and founder of MediaKat llc.