(Bloomberg Markets) -- Oxford University professor Doyne Farmer traces his research exposing risks in the financial system to the roulette wheels of Las Vegas.
In the 1970s, Farmer and two fellow physics students at the University of California at Santa Cruz built a computer small enough to hide in a shoe that helped them predict roughly where roulette balls would land. At casinos in Vegas, they communicated with toe-controlled switches and transmitters, also in their shoes, about what bets to make. The gadget was legal, but they feared their winnings—about a 20% return on their wagers—would lead to trouble. So they quit after a couple of years.
“We were nervous about getting our kneecaps broken,” he explains.
Today, in a more bucolic setting—the Institute for New Economic Thinking at the Oxford Martin School—Farmer is drawing on decades of complexity research that began with roulette. After winning acclaim as a pioneer of chaos theory, which helps explain the unpredictability of complex systems such as the weather, he jumped into markets, co-founding one of the early quantitative investment firms in the 1990s. Now, Farmer and a band of central bank researchers are focusing on the tangled web of global finance, using a tool of the natural sciences called agent-based models to find dangers lurking in the system and uncover ways to avoid them.
Agent-based models, used in fields from biology to sociology, are bottom-up, simulating the messy interactions of hundreds and even millions of agents—human cells or attitudes or financial firms—to explain the behavior of a complex system. The nonlinear interplay can produce unexpected phenomena, such as economic booms and busts, providing insights into the causes of events and the best responses. Epidemiologists have successfully deployed the models for years to test strategies to control everything from obesity to the spread of infectious diseases, including the flu.
Central banks worldwide began experimenting with the agent-based approach after macroeconomists and their standard models were blindsided by the 2008 financial crisis. The European Central Bank, where Farmer is a consultant, as well as central banks in Canada, Germany, South Africa, and the U.K. have taken the lead in building the models to research financial risk. The U.S. Federal Reserve’s regulatory staff is also exploring their use.
Today, central banks mostly stress-test financial firms individually. But agent-based models are giving regulators a better read by accounting for the systemwide impact of shocks. In the simulations, a shock to a single firm cascades through the network of banks and asset managers, creating feedback loops that significantly amplify the initial losses. It’s the kind of contagion that a decade ago spread from the U.S. subprime mortgage market through lenders, money managers, and insurers, creating a liquidity crisis that doomed Lehman Brothers Holdings Inc. and infected the global economy.
“There are efforts at central banks to integrate systemic stress testing,” says Co-Pierre Georg, a research economist at Germany’s central bank (but he doesn’t speak for it). “The banks are highly interconnected and highly leveraged. We now know from Lehman that if something happens to one big bank it can be devastating to the entire economy.”
Farmer, 67, a gray-bearded scientist whose papers have garnered more than 34,500 citations, can be a provocateur. He sees central banks as a beachhead for a bigger challenge to mainstream economics. In a coming book, he says economists in academia resist new approaches such as agent-based models.
It’s a sentiment shared by Bank of England Chief Economist Andrew Haldane, who’s helping lead the push for alternative research. In 2017, Haldane called his profession “insular” for its subpar track record in citing work from other disciplines in its academic journal papers. Farmer says the prestigious top five economic journals rarely if ever publish worthy agent-based model papers, including his own. The American Economic Review says papers are evaluated on their merits without bias.
Many macroeconomists shrug off Farmer’s criticism. He’s a physicist, they say, not an economist, and agent-based models have yet to provide real-world verifiable results.
“In science you need a little bit of the rebel, and Doyne is definitely that,” says Georg, who’s also the South African Reserve Bank chair in financial stability studies at the University of Cape Town. “He challenges you to think about your assumptions.”
Since missing warning signs of the Great Recession, economists have improved their DSGE (dynamic stochastic general equilibrium) models, which remain the workhorses of central bank forecasting. The approach is top-down, aggregating the behavior of the economy into a few representative agents—a household, a firm, and a government.
Most precrisis versions at central banks didn’t include a financial sector because it wasn’t considered relevant in making aggregate forecasts for measures such as gross domestic product and inflation, says Frank Schorfheide, chair of the economics department at the University of Pennsylvania and a visiting scholar with the Federal Reserve Banks in Philadelphia and New York.
Since then, some economists have added a financial sector to capture its impact on the economy. “There were certainly many aspects of the financial crisis that were not on the radar screen of things to monitor by central banks that turned out to be important,” Schorfheide says. “People quickly learned what kind of data they had to collect to understand what just happened and how to modify the models to provide a better narrative.”
Central banks are turning to agent-based models to exploit the wave of new business and social data sets. One of the most data-rich models was built by Farmer and a team including Robert Axtell of George Mason University in Fairfax, Va. They used data from the U.S. Census Bureau, Internal Revenue Service, housing sales, and mortgages to set the behavioral rules for buyers and sellers in the Washington, D.C., housing market from 1997 to 2009. The research, which showed that mortgage lending policy was the key driver of the housing bubble, helped establish agent-based models as an asset for central banks.
Bank of England researchers adapted the Washington model to the U.K. housing market and found that an increase in the size of the “buy-to-let” rental sector could boost the volatility of house prices. Central banks in Hungary and Denmark recently published papers based on the U.K. model.
In 2018, Grzegorz Halaj was a financial stability specialist at the ECB when he used an agent-based model to study liquidity shocks. He tapped balance sheet data for 130 of the largest banking groups in the European Union and aggregated public figures for asset managers. In the simulation, after lenders suffer a drain on deposits, those without an adequate capital cushion cut their interbank lending. In some cases they also dump assets at fire-sale prices. The price drop spills over to fund managers holding the same assets, who then suffer client redemptions and are forced to sell more. That further depresses prices and amplifies the banks’ losses—in some cases possibly leading to defaults, according to the ECB working paper by Halaj, now a principal researcher in the financial stability department at the Bank of Canada.
“It’s part of being a scientist to do things that are novel and important”
For Farmer, the upshot is that policymakers need to develop tests that capture the panoply of losses that financial firms face. In an April paper, Farmer and Alissa Kleinnijenhuis, a graduate student at Oxford, describe a preliminary model they created of a systemwide stress test for the European financial system. It reveals that financial losses from shocks could be three times greater compared with a traditional stress test and that current capital buffers may be too small.
“From the research, we can say unambiguously that there is some amplification of losses from shocks, and I would be very surprised if the amplification isn’t very significant, particularly in times of distress,” Farmer says. “So there is a serious problem to worry about.”
While regulators are far from creating a comprehensive testing model, the BOE is already moving down this path. It’s incorporating feedback loops in areas such as counterparty risk and asset fire sales into its stress tests, said a group of three BOE researchers—Marco Bardoscia, Marc Hinterschweiger, and Arzu Uluc—in an email to Bloomberg Markets. Germany’s central bank has built models to assess the magnification of losses from contagion and is now discussing whether to include them in stress tests.
Farmer, who grew up in the wide-open desert terrain of New Mexico, is fond of chasing big, original ideas. “It’s part of being a scientist to do things that are novel and important,” says Farmer, who likes to explore the mountains of his home state with a backpack and a tent.
He leaped into uncharted territory in the 1970s, when he was a physics graduate student specializing in cosmology at UC Santa Cruz. He switched his focus to chaos and complex systems—a subject so new that there were no professors who could teach it. That didn’t deter Farmer and a small group of fellow students, who formed the “Dynamical Systems Collective” to advise one another on their dissertations.
The Santa Cruz students joined researchers across the country to explain why there’s turbulence in the natural world. Consider the weather. At the time, it was assumed that weather changes came from external disturbances hitting the atmosphere. The scientists showed that the volatility is generated from within, caused by chaos, in which a small disturbance in the initial conditions of a complex system is amplified exponentially; that’s why the weather is so hard to predict. The discoveries, which influenced fields from math to the social sciences, were chronicled in James Gleick’s best-selling 1987 book, Chaos: Making a New Science.
Chaos theory eventually penetrated pop culture. In Steven Spielberg’s 1993 film Jurassic Park, Jeff Goldblum plays a mathematician specializing in chaos. Farmer says Goldblum called him to help prepare for the role. “He wanted to understand how a chaos scientist speaks,” he says.
In 1981, Farmer went to the Center for Nonlinear Studies at Los Alamos National Laboratory—where freewheeling research was the norm. Seven years later he started the Complex Systems Group at the lab in New Mexico, bringing together theoretical scholars who helped develop the nascent field of complexity studies.
After a decade at Los Alamos, Farmer was lured into investing. Scientists helped plant the seed: At conferences, they’d ask him if he’d considered applying his insights about the nature of chaos to the stock market. He didn’t know how to trade, but he knew how to make short-term predictions in complex systems such as fluid flows. So he set himself a goal of making $5 million in five years—a sum he’d substantially surpass.
In 1991, Farmer started Prediction Co. in Santa Fe, N.M., with two physicist friends, Norman Packard and James McGill. After about five years, Prediction’s automated statistical arbitrage strategy was earning risk-adjusted returns about five times the S&P 500’s. The company traded proprietary capital for UBS AG, and the partners sold Prediction to the bank for $100 million in stages, ending in 2005.
Farmer had left Prediction in 1999 with the financial freedom to follow his own research interests. He landed at the Santa Fe Institute, where he revised a paper showing how changes in market ecology, composed of trend followers and value investors, can cause crashes. He says he had several conversations with legendary hedge fund investor George Soros about the paper. When Soros’s Institute for New Economic Thinking formed a partnership with Oxford in 2012, the university hired Farmer to run its complexity economics program.
Complexity economists, while few in number, have a shared ambition to knock holes in a cornerstone of DSGE models: rational expectations, the idea that everyone in the economy understands each other’s decisions, and they’re mutually compatible in their pursuit of their self-interest. Since the crisis, mainstream economists have added “frictions” to the models—such as lenders denying loans to creditworthy businesses—to make them more realistic.
“It’s part of being a scientist to do things that are novel and important”
Behavioral economists have shown that people aren’t perfect calculators and often make rule-of-thumb judgments, such as concluding that rising markets will keep going up despite other evidence to the contrary. This is the kind of real-world behavior that agent-based models capture. The agents are also programmed to behave differently to express an economy’s diversity. Heterogeneity is the hallmark of the models. In Farmer’s Washington housing model, 100,000 households made varying buying and selling decisions based on their income and savings.
One challenge unmet by complexity economists is understanding the behavior of complicated humans. Economics is harder than physics, Farmer says. People don’t obey rules as reliably as atoms do. In his stress-test model of the European financial system, for instance, he had to make assumptions about what assets bankers might dump after their firms suffered shocks. “They will likely sell the most liquid assets because they will lose less money,” he says.
Agent-based models also suffer from the black-box problem. The inner workings are so complex, with thousands of agents running in different directions, that it can be difficult to pinpoint the main drivers of a model’s findings. That doesn’t sit well with central bankers who need to know the reasons behind their decisions, says Georg of the University of Cape Town.
“With DSGE models, we know exactly how A follows from B, and I can explain that to my governor,” he says. “But in an agent-based model, I can’t do that. So how do you communicate these results with the hierarchy? That’s the big missing piece.”
Farmer says it could take years of research to make the models a mainstay of central banking. To achieve this, researchers will need lenders to provide them with more detailed balance sheet data showing the linkages among them—something European banks have resisted because of privacy and cost concerns. He says the payoff—addressing risks before they cascade into meltdowns—seems worth the effort.
“Agent-based models can be a game changer,” he says. “I’m convinced we can solve these problems.”
Bielski is a senior editor on Bloomberg News’s investing team in New York. With Lucy Meakin in London
To contact the author of this story: Vincent Bielski in New York at firstname.lastname@example.org
To contact the editor responsible for this story: Christine Harper at email@example.com
For more articles like this, please visit us at bloomberg.com
©2019 Bloomberg L.P.