Written by:

Shayne Kavanagh, Senior Manager of Research, GFOA

Sam Savage, Executive Director of ProbabilityManagement.org and GFOA Research Fellow

Matthew Raphaelson, Chair of Financial Applications at ProbabilityManagement.org

Though the full story of what, exactly, caused the collapse of Silicon Valley Bank (SVB) is still coming into focus, it does seem clear that insufficient stress testing played at least some role. A bank stress test determines whether a bank has enough capital to withstand a financial stress or shock. A stress test can address different types of risks, like credit risk, market risk, and liquidity risk.(1) So, a stress test is essentially a risk analysis. The gold standard for conducting risk analysis is computer simulation. Computer simulation is standard practice in fields as diverse as insurance, oil exploration, and climate science – if you need to manage large uncertainty and risk, then you use computer simulation.

The simulation is a lot like the multi-verse convention that has entered popular consciousness via science fiction movies like the Marvel franchise and Everything Everywhere All at Once. Computer simulation works by essentially creating hundreds or thousands of parallel universes and changing key variables across the universes and seeing how it plays out. It is analogous to shaking a ladder before you climb on it to check its stability. Even though the forces when you shake a ladder differ from those when you actually climb on it, it provides important directional information. Similarly in simulation, a precise understanding of the underlying uncertainties is not necessary to yield valid insights.

For example, in SVB’s case one key variable was the rate of inflation and the Federal Reserve’s response in the form of an increase in the Federal Funds Rate. The interest rate increase reduced the value of SVB’s investments.

Perhaps unique to SVB, its customer base of venture capitalists and tech was slammed by the same rising interest rates, which dried up funding for the entire industry.  This resulted in the customer base taking down deposits all at the same time (as opposed to bank models which assume that customers deposits in and withdrawals out more or less offset each other.) SVB was insufficiently capitalized to withstand the simultaneous impact of reduced investment value and high deposit withdrawals.

Another less obvious variable was the tight knit community of venture capitalists. As they became aware of the capital problem above, they warned each other over social media to get their money out, and it triggered a flash run on the bank, resulting in unrecoverable insolvency almost overnight.

A computer simulation that addresses multiple risk factors and interactions between those risk factors would highlight the threshold point at which a combination of interest rate increase and coordinated withdrawals would threaten the Bank’s position. In addition, it could point the way toward alternative strategies that would be less vulnerable to these risks. SVB apparently failed to shake their ladder in enough directions to adequately expose the risks it faced.(2) As a result, they were not adequately aware of their exposures and therefore did not develop adequate countermeasures.

In the past, computer simulation required specialized software and a degree in statistics. Now, thanks to advances in technology, computer simulations that address multiple risks at once can be easily built in Microsoft Excel and user interfaces can be added that make the simulation accessible to non-statisticians. Nonprofit ProbabilityManagement.org, a GFOA partner has created tools and standards to greatly facilitate this sort of modeling. As a result, computer simulations can be more widely applied to risks faced by critical institutions in our society.

While we should learn from our mistakes, it is far preferable to learn from the mistakes of others, in this case SVB.(3) Local governments are critical institutions in their communities and can apply computer simulation to risks they face. We have provided a simple example here created with the tools from ProbabilityManagement.org. This mini-stress test demonstration addresses the risk of running a budget deficit.

You enter your revenue forecast, including estimates of the potential error in your forecast. The error is the difference between the forecast and what actually happens. No forecast will ever be exactly right and can sometimes be quite wrong – the potential error in the forecast presents a risk that revenues will be insufficient to cover planned spending. You can estimate the potential error by looking at the difference between past forecasts and what actually happened. You could also apply your professional judgment to estimate the range of possible outcomes.(4)

Given the potential error in your forecast, the model tells you:

• The chance you will run a deficit or a surplus
• The average size of the deficit or surplus you can expect
• If you are unhappy with your chance of a deficit, given your spending plan, the model tells you the size of the budget you can have in order to incur a level of risk you are comfortable with.

A risk model like this also can stimulate conversation on how to manage risk. Decision-makers are uncomfortable with the level of risk in their spending plan, they might consider risk mitigating strategies. For example, some spending could be deferred until midyear when there will be greater insight into whether actual revenues could support that spending. Or, some portion of spending could be devoted to one-time projects that can be safely deferred if revenues don’t prove sufficient. You can read more about how such a model has been used in real life by a local government in “Speaking Uncertainty to Power: Risk-Aware Forecasting and Budgeting.” In this case, the finance officer worked with her City Council to set spending at a level that gave the City a good chance of generating a surplus, which could be used to replenish the City’s depleted reserves. In successive years, as the City succeeded in replenishing its reserves, the Council opted for a spending plan that provided a smaller chance of a surplus and higher levels of current service.

Simulations can also be built for more sophisticated public finance problems too. GFOA, for example, has built simulations for stress testing: reserves in light of risk of economic recessions and natural catastrophes; debt affordability in light of changing bond ratings and interest rates; and long-term revenue and expenditure forecasts in light of recessions, pension investment performance, health care cost increases, and more.

So, what can you do to learn from the mistakes of SVB? Start by giving the demonstration model a spin. This will begin to acquaint you with how computer simulation works. Then, stay tuned to GFOA by visiting our page on risk analysis and signing up for updates. We have additional resources on stress testing that we’ve developed over the past number of years and will have more announcements, like our initiative with global insurance giant Aon, that can help local governments better manage risks in an increasingly complex,  volatile, and uncertain world.

(1) Description of bank stress test quoted from: https://www.investopedia.com/terms/b/bank-stress-test.asp

(2) “Before the Collapse of Silicon Valley Bank, the Fed Spotted Big Problems”. New York Times. March 19, 2023

(3) Paraphrased from Otto Von Bismarck

(4) As of this writing, GFOA is pilot testing a training course to greatly improve the student’s ability subjectively to develop estimates of probabilistic ranges, such as 90% confidence interval. Early results are quite promising. A summary of results will be published soon.