Some in the United States have proposed raising the retirement age in order to spend less on Medicare and Social Security. But it’s not simple to predict whether such a policy change would save as much as expected, in part because it depends on how the elderly will react. Predicting this requires also knowing why older people save their money instead of spending it in their old age, a behavior that has been observed and noted in research.
Untangling this if-then exercise quickly gets complicated, but Chicago Booth’s Tetsuya Kaji, NYU’s Elena Manresa, and University of Chicago Harris School of Public Policy’s Guillaume Pouliot have devised a machine-learning method to help. And if ML could better explain behavior, such as what drives people to preserve money in their old age, policy makers could make more-informed decisions.
To understand how their ML model works, says Kaji, consider the art world. “What do you think it means to have an authentic painting?” he asks. Say you have a work that looks like it was painted by Leonardo da Vinci, and most critics say it’s authentic. Is it?
In the world of art, authenticity is defined on the basis of what the majority of critics agree, so successful forgers have to trick the critics. In the 20th century, a Dutch painter named Han van Meegeren forged paintings by a number of famous artists, including Vermeer. After World War II, when he was accused of having sold Dutch property to the Nazis, van Meegeren confessed and was found guilty of the lesser charges of forgery and fraud.
The researchers’ ML method, explains Kaji, is similar to the cat-and-mouse game played by critics and forgers. Say a “critic” ML algorithm is trained to identify paintings by Rembrandt, much like an autonomous car is trained to identify pedestrians. But a “forger” ML algorithm tries to generate paintings that can be misidentified as a real Rembrandt by the critic ML. As the forger ML learns, it starts to create paintings that increasingly look authentic and confuse the critic ML.
A mathematical version of this is the method that the researchers apply to complicated questions, says Kaji. In their method, a “generator” creates fake economic data using a model backed by economic theory, and a “discriminator” classifies whether the data are genuine. The competing algorithms, in a method the researchers call an adversarial estimation, establish the parameters needed to make realistic, more-accurate predictions in situations where behavior changes can affect outcomes.
The researchers applied their method to the question of why the elderly spend money at a puzzlingly slow rate. One algorithm looked for features in economic statistics that distinguish the observed behavior of the actual elderly, while the other looked for an economic model that reproduces such statistics to trick its opponent.
The economic model that this cat-and-mouse game generated aligns the complicated incentive structure in a way that mimics the elderly’s actual behavior, which makes it possible to try out different policies in a simulated world and see what happens. The researchers used their model to assess several possible explanations, including that older people don’t know how long they’ll live, want to leave money to heirs, or are concerned with rising medical costs.
The ML-inspired method doesn’t require selecting and using features that might best represent the elderly’s behavior. Instead, the method adaptively learns from data and figures out for itself an optimal set of features. And the researchers find the model’s explanation for why older people spend slowly—that they, even if not rich, want to leave money to their heirs—to be reasonable and plausible. They say their method opens up new possibilities to address deeper and more sophisticated policy questions than was previously feasible.