Using agent based models for stress testing

Traditional modelling approaches are ill-suited to the type of scenario-based risk management that stress tests are concerned with. Agent-based models (or ABMs), in contrast, provide a natural framework for exploring challenging scenarios faced by complex adaptive systems — as I shall illustrate later in this post with a simple mortgage model.

Agent Based Models allow flexible and highly heterogeneous modelling of balance sheets

Coupled with powerful computational simulation, ABMs are well-suited to the exploration of balance sheet risks and to determine appropriate interventions to mitigate those risks. Embracing alternative modelling paradigms is a key step towards extracting business value from stress tests, over and above banks’ regulatory obligations.

Performing a stress test of a large global bank is no mean feat. The exercises set by regulators require banks to run macroeconomic scenarios described by hundreds of variables — which banks then expand out to thousands of risk factors which impact their balance sheet. The paths of these risk factors are then modelled to determine the capital impact of a given scenario — but this is of only limited use for day-to-day risk management.

While banks have worked hard to get to grips with the requirement set by regulators, they have been slower to adopt new modelling paradigms. ABMs are complementary analytical tools that integrate with banks’ existing stress testing data and modelling apparatus and act as a force-multiplier by allowing risk managers to extract new insight and explore “what-if?” scenarios.

Let’s look at a simple example of an agent-based model and see how it can be used to explore risks in a stress testing set-up.

Credit Models are key to stress testing and a natural fit for an agent-based approach

Agent Based Mortgage Book Modelling

Stress testing the mortgage book is a challenge faced by just about every bank. Standard models used to stress the mortgage book typically take a macro scenario and run it through a macro-to-micro model — which uses historical statistical relationships to map macro variables (e.g. GDP, Unemployment, HPI etc.)into arrears and other key portfolio-specific metrics. These metrics are then employed in a life-cycle model to map arrears into provisions, impairments and write-offs.

The agent-based approach is very different in a number of ways. Firstly, an agent-based model starts from the bottom-up. I start by modelling each household as an individual autonomous agent. These agents are initialised with a variety of information: age, income, wealth etc. which are drawn from real data.

To capture the mortgage book life cycle, the model also needs to capture real-world processes — mortgages being written, households paying their mortgage each month, income and wealth shocks, defaults and arrears.

Banks also feature in the model; they have rules which govern their underwriting standards (loan to value and loan to income), as well as a pricing strategy.

This basic framework describes an agent-based simulator of a bank’s mortgage book. The simulator is able to generate all of the mortgage-book dynamics that occur in the real-world: banks write mortgages, accrue interest income, and take impairments. Running that simulation millions of times generates sufficient paths to capture the complex dynamics at work in banks’ balance sheets. These complex emergent dynamics tend to be obscured by restrictive statistical models or when dealing with aggregated portfolio-level data.

More interestingly for risk managers, they can also experiment with combinations of their “policy parameters” — for example changing their underwriting standards, or pricing strategy to explore the impact that these have on key metrics.

It is also possible to ensure that the metrics generated by an agent-based model are consistent with regulatory metrics — or the accounting standards of the day. This is a particularly relevant challenge at the moment. Many banks are getting to grips with IFRS9 — a new accounting standard that changes the way banks provision against losses.

Agent-based models are an ideal set-up for producing IFRS9-consistent outputs. The agent-based set-up is rich enough to capture key concepts embedded in IFRS9. It is straightforward to apply a SICR definition to map a mortgage into the correct stage. And simulating forward throughout the lifetime of a financial instrument enables the modeller to quantify the ECL. These are much trickier to do from traditional statistical models — particularly given the absence of historical data!

Traditional modelling paradigms are simply not well-suited to capturing tipping points introduced by stage allocation, or producing distributional outputs like those required to compute ECL. If you’re interested, you can read more about IFRS9 and agent-based models.

In summary, agent-based simulation is an important complementary technology that should sit alongside traditional risk-management models. Banks can use it to extract additional insight from their existing stress testing infrastructure, as well as providing challenge and alternative insight into their existing models and risk-management decisions.

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.