Conventional economic models do not capture the complexity of human behaviour
HAVING GONE deep into the red during the covid-19 pandemic, governments will grapple in 2021 with getting their finances back in order. After the global financial crisis of 2007-09, those in rich countries tightened their belts too much, choking the economic recovery. This time they will want to be cleverer about it. Some will be more ambitious, seeking to redesign their welfare systems: the pandemic will strengthen public support for stronger social safety-nets. And policymakers in poor countries will want to alleviate poverty and sustain economic development.
How to balance all these aims? An experiment might tell you if a particular tool works, and findings from projects on basic income, such as that run in Kenya by Give Directly, a charity, will influence governments’ thinking. But experiments can be neither broad nor timely enough to help governments set a plethora of tax and subsidy rates every year. Conventional economic models do not capture the complexity of human behaviour: that people change what they do as tax rates rise, or that corrupt officials might pocket some public funds. So in 2021 governments will be tempted to throw computational power at their policymaking, using artificial intelligence (AI) to simulate the economy, and the effects of new policies.
“Agent-based” models simulate the behaviour of different types of participants in the economy by allowing them to respond to each other over time: if a public servant can get away with pocketing more money, or a taxpayer with paying less tax, then they will do so. Some simulate surprisingly realistic behaviour by using machine learning to “train” the model using vast sets of data. One such approach is Policy Priority Inference, developed by researchers in Britain and Mexico and sponsored by the UN’s development programme. Already used in Mexico, it takes governments’ spending plans across a range of categories and works out, based on its simulation of corruption, inefficiencies and spillovers, whether a government is likely to hit its development goals, and where more (or less) money should be spent. More poor countries could see the appeal of such an approach.
“基于主体”模型（ABM）会模拟经济体中不同类型的参与者之间的长期相互影响来预测他们的行为：如果公务员贪污或纳税人逃税可以不受责罚，那么他们就会这么做。一些模型使用机器学习，通过海量数据来“训练”模型，模拟的行为真实得令人吃惊。由英国和墨西哥的研究人员开发、联合国开发计划署赞助的“政策优先顺位推断”项目（Policy Priority Inference）就使用了这种方法。它已经在墨西哥使用，将各种类别的政府开支计划输入模型，根据它对腐败、低效和溢出效应的模拟，计算出政府是否有可能实现其发展目标，以及应该在什么方面多花钱或少花钱。更多贫穷国家可能会看到这种方法的吸引力。
Interest in rich countries could be piqued, too. Researchers at Salesforce, a software company, and Harvard University have used simulations to show that, much as computers can learn to play Go and develop strategies that might not occur to humans, they can also suggest combinations of tax and spending that maximise economic performance, and which bureaucrats might not have dreamed up. So why not turn to AI for fresh ideas?
None of this means that economists or bureaucrats will find themselves out of work in 2021. Interpreting the models’ results requires expertise. Politicians will not cede their power to raise and lower tax rates. But policymakers and researchers keen to experiment in the aftermath of the pandemic will have an opportunity to expand their toolkits.