Editor's Note: this week’s faiV highlights more research on financial inclusion and machine learning from the American Economic Association annual meetings, guest-edited by Sean Higgins, a Post-Doctoral Fellow at the Center for Effective Global Action at UC Berkeley, whose research focuses on financial inclusion.
Next week, I'm hoping Jonathan Morduch will fill in for me before I resume normal service the week of February 5th--Tim Ogden
1. Financial Inclusion: I [Sean] organized a session on savings and financial inclusion that looked at the impact of various savings interventions such as commitment devices, opt-out savings plans, and mobile money. Continuing last week’s theme on similarities between developed and developing countries, a savings intervention that has greatly increased savings in the US is opt-out savings plans or “default assignment,” such as being automatically enrolled in a 401(k) plan. In an experiment in Afghanistan, Joshua Blumenstock, Michael Callen, and Tarek Ghani explore why defaults affect behavior: some employees are defaulted into a savings program where 5% of their salaries are automatically deposited in a mobile money savings account, but they can opt out at any time. Those who were defaulted in were 40 percentage points more likely to contribute to the savings account, which is comparable to the effect of the employer matching 50% of employees’ savings contributions
Commitment savings accounts have also been tested in the US and in many other countries. In a study by Emily Breza, Martin Kanz, and Leora Klapper, employees in Bangladesh were offered a commitment savings account, with a twist: depending on the treatment arm, employers sometimes endorsed the product, and employees were sometimes told that their decision would be disclosed to the employer. Only the treatment arm that had both employer endorsement and disclosure of the employee’s choice led to higher take-up, suggesting that workplace signaling motivated employees to save. Another study by Simone Schaner et al. (covered in last week’s faiV) offered employees in Ghana a commitment savings product with the goal of building up enough savings to stop incurring overdraft fees, which are common. Take-up was high, but baseline overdrafters were more likely to draw down their savings before the commitment period ended -- meaning they benefited less from the intervention.
Two important barriers to financial inclusion in the US and around the world are transaction costs and low trust in banks. In a paper I coauthored with Pierre Bachas, Paul Gertler, and Enrique Seira, we study the impact of providing debit cards to government cash transfer recipients who were already receiving their benefits directly deposited into a bank account. Debit cards lower the indirect transaction costs -- such as time and travel costs -- of both accessing money in a bank account and monitoring the bank to build trust. Once they receive debit cards, beneficiaries check their balances frequently, and the number of checks decreases over time as their reported trust in the bank and savings increase"
2. Household Finance: Digital credit is a financial service that is rapidly spreading around the world; it uses non-traditional data (such as mobile phone data) to evaluate creditworthiness and provide instant and remote small loans, often through mobile money accounts. One of the concerns about digital credit is that customers’ credit scores can be negatively impacted, even for the failure to repay a few dollars. In turn, this can leave them financially excluded in the future. Andres Liberman, Daniel Paravisini, and Vikram Pathania find a similar result for “high-cost loans” in the UK (which we would call payday loans in the US). They use a natural experiment and compare applicants who receive loans with similar applicants who do not receive loans to study the impact of the loans on financial outcomes. For the average applicant, taking up a high-cost loan causes an immediate decrease in the credit score, and as a result the applicant has less access to credit in the future.
3. Our Algorithmic Overlords: There were a number of sessions at the AEA meetings on big data and machine learning. My favorite of these showcased a variety of economic applications of machine learning, three of which use big data from mobile phones. Susan Athey et al. use high-frequency location data from mobile phones to estimate a consumer choice model over restaurants and travel time. There are a large number of variables going into each individual’s decision of where to go for lunch, and each individual is different; the benefit of using machine learning is that they can incorporate a large number of variables on both restaurants and consumer preferences into the model. Susan also has an excellent overview of applications of machine learning in economics here.
Mobile phone data can also be used to predict creditworthiness: in a middle-income Latin American country, Daniel Björkegren and Darrell Grissen find that mobile phone call detail records perform just as well at predicting creditworthiness as traditional credit bureau scores (although neither perform particularly well in this sample). The mobile phone data appears to be picking up useful information to predict creditworthiness, and could be especially useful for consumers with no formal credit history or traditional credit score. These data sources and models could also help low-income women, who face a bias in the amount lenders are willing to provide, higher interest rates, and legal frameworks which can make it more difficult for them to access credit.
4. More Machine Learning: After the meetings each year, the AEA offers two-day continuing education courses on a changing variety of topics. This year, one of the courses was Machine Learning and Econometrics taught by Susan Athey and Guido Imbens. The webcasts and slides from the course can be accessed here. As economics increasingly adopts methods from machine learning in the coming years, this class’s combination of practical tools, R code, intuition, and theory make it more than worth your time to watch the webcasts and peruse the course materials.
One of the gems was the intuitive descriptions of various machine learning techniques. I feel like I finally have an intuitive understanding of what stochastic gradient descent and neural nets do (and I had to explain it to a friend yesterday which is always a good test). For example, here’s Susan’s description of the “incredibly powerful” method of stochastic gradient descent (in minute 58 of this video). What we usually do: “Estimating a model is climbing a mountain. In economics the way we approached that problem historically, is if you were climbing up that mountain trying to find the parameters that maximize an objective function, at a particular point in climbing that hill there’s a gradient that tells you in which direction should I change my parameters to get up to the top of the hill and find the parameters that best fit my data. We might spend fifteen minutes of our computation computing the gradient at one point, and then climb up the hill a little bit and work really hard at computing the gradient at the next point.”
The magic of stochastic gradient descent: “At each point in climbing the hill, you evaluate the gradient using just one data point from your data set…you just pick one data point and compute where you should go as if that data point was your only data point. It’s an unbiased estimate of the gradient but it’s incredibly noisy. But instead of doing 10,000 computations to figure out how to make one tiny step, instead 10,000 times you go up and down your hill, up-down-up-down, over here over there, but you’re always kind of going in the right direction. And 10,000 points later you’re almost at the top, while with our old methods you would have gone much more in the right direction but you would have just made one tiny step and you’re nowhere near the top of the mountain.”
5. Inequality: The World Wealth & Income Database group led by Thomas Piketty, Facundo Alvaredo, and Lucas Chancel at the Paris School of Economics and Emmanual Saez and Gabriel Zucman at UC Berkeley presented on global inequality and policy. Recently, the group has been combining data from household surveys, national accounts, and tax records to create more comprehensive measures of income and wealth inequality. One interesting finding they presented was that Brazil’s large reduction in inequality since 2001 -- which is based on income measured in household surveys -- goes away if we instead use a measure that combines data from household surveys, national accounts, and tax records. With the more comprehensive measure, income inequality in Brazil has been flat. They also reported that inequality is increasing in almost every region of the world, and the global top 1% have about 20% of global income. A webcast of this session is available here.
Default assignment into an opt-out automatic savings plan leads to a large increase in take-up of the savings account, comparable to the effect of a 50% savings match (from Blumenstock, Callen, and Ghani).
