Coronavirus: Making policy outside the database

In the middle of the coronavirus pandemic, fiscal policymakers, health professionals, and others are focused on critically important short-term decisions – whether to cut payroll taxes, send checks, act as the payer-of-last-resort, and so on – and rightly so. When policymakers were making similarly difficult decisions in 2009, Doyne Farmer and Duncan Foley wrote that one would assume that leaders in the US and abroad “are using sophisticated quantitative computer models to guide us out of the current economic crisis. They are not.” The same is true today.

Farmer and Foley pointed out that policymakers rely on two types of models to determine their response: empirical statistical models – which are fit to past data – and general equilibrium models – which assume a perfect world, thereby ruling out crises. These models have less-than-perfect explanatory ability due to their strong assumptions. Their main strength is high predictive power in stable periods. If GDP grew by 2 percent last year and we all maintain our routines, it’s generally a pretty good guess that GDP will grow by 2 percent this year.

However, as we see in times such as 2007-09 and today, the predictive power of these models becomes relatively nonexistent once the relationship between the models’ dependent and supposedly independent variables stop reflecting the behavior of the “complex networks of agents and institutions, stocks and flows, goods and services, money and credit”. Beyond the human tragedy, it is hard to know the full impact of empty streets, shuttered local retailers, and halted international supply lines. But clearly the statistical relationships of the empirical models and the assumptions of the general equilibrium models do not provide useable information to leaders during a pandemic. In the words of Bill Janeway in Doing Capitalism in the Innovation Economy, we are “living outside the database”.

Why are we still here?

Much like the financial crisis made it clear that agent-based models are a better tool for understanding our our economy, the current health crisis and lack of a useful economic framework for evaluating fiscal policy responses. As I summarize below in Legacy economists and penguins, network effects have locked in the dominant economic framework of modeling the economy as a system in general equilibrium, despite its many known limitations and promising alternatives.

In normal times, that might seem fine – let academics be academics. But when it becomes clear that standard economic research and its effect on policies during crisis can have direct implications on millions of lives, the failure to explore better approaches begins to appear negligent.

One of the most promising approaches for large-scale deployment in economics is agent-based models – computational models that simulate the actions and interactions of agents such as firms and households to assess individual-level and system-wide impacts. These models are not new. Although economics has been slow to adopt them, these models have gained traction in areas such as epidemiology, physics, and biology, with many successful applications.

An example of an agent-based model

In 1996, the Department of Energy commissioned Sandia National Laboratories to build the Aspen agent-based model to study the economic effects of traumatic events such as energy grid failures and telecommunications shutdowns. In contrast to most general equilibrium models, in Aspen time is divided into days instead of years (useful with a rapidly spreading virus), agents make decisions based on artificial intelligence learning models instead of with perfect foresight (potentially useful for identifying behaviors during and after forced lockdowns), and a banking sector conducts open market operations instead of simply not existing (useful when the Federal Reserve has dropped the fed funds rate to zero and is purchasing $700 billion of assets while supplying liquidity to markets). Perhaps most important for understanding economic policy during and after a pandemic, agents in the model interact. Aspen agents, much like humans, “adapt their behavior dynamically, according to changing economic conditions and past experience.” Unfortunately, Sandia’s work on Aspen stopped around 2000.

Today, a similar model could have been used to guide policymakers in determining fiscal policy responses, which should be designed to provide maximum support to the all-important health policy decisions. We have an opportunity now, in preparation for the next crisis, to use the exponential increases in data, computing power, and behavioral and neuroscience research over the past two decades to develop an updated Aspen.

Next time we face a crisis, hopefully policymakers will have a more useful guide.

The National Science Foundation took preliminary steps to fund development of a similar model in the wake of the financial crisis, and a few academic and private sector groups are working on agent-based models. These steps are positive, but the necessary large-scale commitment has been missing. Hopefully one small positive will come amidst the pain and suffering caused by the pandemic: better recognition of the harmful effects of unrealistic economic models and stronger support – from the federal government, philanthropies and philanthropists, universities, and others – for development of a better way.