Agent-based models and the evolution of risk management

Risk management today is little different from that employed a quarter of a century ago for managing broker-dealer trading desks. It relies on price history, so at its foundation assumes that the future will be drawn from the same distribution as the past. It includes standard stress-testing 'what-if' exercises that are also executed through a historical lens – either explicitly by looking at times in the past where similar events occurred, or by applying the historical relationships between the markets as they appear in past correlations.

One problem with our current approach, which I call Risk Management 1.0, is that the past will not repeat, even if an identical shock occurs. The market today is not the market of the past. Leverage and liquidity are ever changing. There are new strategies and new financial instruments. Nor are today’s investors and borrowers those of the past. They engage differently having learned from – and perhaps having been burned by – past experiences. And, most importantly, they do not operate in a vacuum. The actions of one can affect the environment and the actions of others.

Dynamic tools

Where Risk Management 1.0 is rooted in the past, Risk Management 2.0 must point toward the future. Doing this requires a tool that looks at the market as it stands today, accommodating the complex and ever-changing dynamics of the market. One such tool is the agent-based model.

Agent-based models are used in areas ranging from locating exit ramps on highways to minimise traffic jams to designing nightclubs to anticipate potential stampedes during a fire. I discovered, not surprisingly, that they could find an application with market uncertainties as well.  

I developed the agent-based model approach to risk management while serving in the US Treasury after the 2008 crisis. My objective there was to assess the vulnerabilities of the financial system in the face of the complexities that only seemed to become recognised after that crisis. I have since put this approach to use in an investor-oriented setting with the University of California’s $120bn pension and endowment portfolios.

Non-linear system

To get a sense of agent-based modelling, look at how it is used in assessing the risk of traffic congestion. For traffic flow, the agents are the drivers. Some drivers drive slowly in the left lane, some zigzag around slower cars. Each driver reacts to changes in its environment, which in turn changes the environment of others. The result is a complex, ever-changing and non-linear dynamical system.

To understand the likelihood of congestion in some strip of highway, unleash the agent-based model on the problem. Pepper a simulated roadway with various drivers, and second by second have each driver observe its respective environment. Each second, each driver alters its actions, changing the environment for themselves and for nearby drivers, and each driver adjusts accordingly. Do this many times, and then look at the distribution of congestion.

In the financial markets the agents are investors, market makers, borrowers and lenders. As with the agents in the traffic application, some investors trade quickly, others slowly. Some agents take on leverage, and some of them liquidate rapidly in the face of market shocks. The model looks at the markets as they are today; for example, their current leverage, market and funding liquidity, market concentration across agents, for the sort of events that, like a stalled car on the highway, can lead to unexpected cascades of 'congestion' within the markets.

The human factor

For a bank, risk arises not only from the market, but also from its clients. This means the bank’s decisions can have broad repercussions for its own risk. The attempt to reduce risk by, say, calling in a loan or executing a margin call, might reverberate to actually increase risk – both for the bank and for the system overall. When we combine high-risk events or periods of crisis with the human dimension, the agent-based approach takes on some of the characteristics of the fire-in-the-nightclub stampedes.

For the next stage of risk management, Risk Management 3.0, banks must recognise that the markets, the bank and the clients are all agents to be incorporated in the agent-based model, and that each of these agents’ actions changes the environment for the bank and for others, and thus affects the future decisions of the other agents. The path from Risk Management 1.0 to Risk Management 3.0 demands throwing away any black box that spews out numbers without context or understanding, and focusing on the market along with the risks and decisions of the actual agents that inhabit it.

Richard Bookstaber is chief risk officer for the investment office at the University of California and a former senior adviser to the secretary at the US Treasury.