Innovative Data in Household Finance: Opportunities and Challenges Fall 2020

Can an household-level financial transaction data yield new insights about firm-specific risk? Baker, Baugh, and Sammon develop two new measures characterizing firms' customer bases - the rate of churn in a firm's customer base and the pairwise similarity between firms' customer bases - using an increasingly accessible class of household financial transaction data. They validate our approach by using the data to construct accurate pictures of firm revenue, growth, geographic dispersion, and customer base characteristics. The researchers show that these measures of customer bases are impossible to construct utilizing traditional sources of firm data, but provide important insights into the behavior of both real firm decisions and firm asset prices. Rates of customer churn affect the level and volatility of firm-level investment, markups, and profits. Churn also affects how quickly firms respond to shocks in the value of their growth options (i.e. Tobin's Q). Similarity between firms' customer bases highlights one under-explored type of predictability among stock returns - Baker, Baugh, and Sammon demonstrate that significant alpha can be generated using a trading strategy that exploits our index of customer base similarity across firms.

Coombs, Dube, Kluender, Naidu, and Stepner present new results on the consumption, savings, and income effects of the introduction of the unusually generous unemployment insurance benefits during the COVID-19 pandemic in April, their abrupt expiration at the end of July, and their short-term partial reintroduction through August and September. They use a new dataset of administrative bank account balances and transactions 1.2 million workers and 258,065 recipients of UI. The researchers link these administrative data with a large-scale survey (N = 24,671) of expectations and economic preferences. Coombs, Dube, Kluender, Naidu, and Stepner find that account outflows fell by 20% among July UI recipients in the 12 weeks since expiration relative to non-recipients. They find that consumption drops around expiration were muted owing to accumulated savings out of the expanded UI over the March-July period; end of July savings were roughly three times as large as savings in January. The magnitude of the drop in savings following the expiration was larger in households with low expectations of continuing benefits, no children, low risk aversion, and high discount rates. The researchers also find that the temporary Lost Wages Assistance program provided a small but temporary boost to savings and consumption, and the timing of this boost varied based on the staggered adoption by states.

Vellekoop and Pettinicchi use subjective probabilities of job loss for a representative panel of the population of the Netherlands, as measured in a survey. They link the survey data with job loss expectations to integral administrative data on a (i) worker-firm panel and (ii) administrative data on car acquisitions and household income and wealth. The first linking allows us to investigate the empirical content of job loss expectations. The second linking allows us to say something about economic behavior following stated job loss expectations. The new data part is combining the strengths of survey data (measuring subjective expectations) and administrative data (tracking outcomes in the months after the survey is performed). This allows us to circumvent issues of panel attrition and non-response for sensitive outcomes (i.e. unemployment)

This project uses an administrative data set containing millions of transaction-level records from a number of the largest indirect auto lenders. These data come from the Consumer Financial Protection Bureau supervisory efforts to evaluate lenders' compliance with federal consumer financial laws, and have generally not been available to researchers. Though the identities of the lenders are masked, the institutions in the data set comprise more than 20% of the indirect market in any given year, and include both traditional banks and finance companies that specialize in auto loans (both manufacturer-specific “captive” lenders, and non-captives). These data include all the objective variables used by the institutions to underwrite and price the loan, the financial characteristics of borrower observable to the dealer (e.g. FICO, risk-based interest, income), vehicle information (e.g. make, model, year, new/used), and negotiated terms of the transaction (e.g. price paid, add-ons, etc.). Importantly, they also separately report the buy rate and markup for each deal. The markup is the amount of additional interest a dealer discretionary adds to the loan over and above the lender’s risk-based buy rate. Dealers receive additional compensation for adding markup, which creates a number of potentially interesting incentives. Pervious work on markup and auto loans has typically relied on data from sources such as the Consumer Expenditure Survey or the Survey of Consumer Finances, which make it difficult to parse out markup, and which may not capture the specific income, credit, and risk characteristics that the lenders use to underwrite and price the loans. By using administrative data collected directly from lender Lanning is able to perform a more comprehensive and complete assessment of this under-explored market, and conduct credible tests for specific forms of economic discrimination that may be driving disparity in the market.

Recent local price growth explains differences in search behavior across prospective homebuyers. Those experiencing higher growth in their postcode of residence search more broadly across locations and house characteristics, without changing attention devoted to
individual sales listings. Effects are stronger for homeowners, in particular those living in less wealthy areas and looking for a new primary residence. These findings are not consistent with local price growth influencing behavior through extrapolative expectations,
and rather line up with the predictions of a collateral constraints channel. The expansion of search breadth leads to widespread spillovers onto house sales within a metropolitan area.

This paper studies the value of privacy, for individuals, using data from large-scale field experiments that vary disclosure requirements for loan applicants and loan terms on an online peer-to-peer lending platform in China. Tang finds that loan applicants attach positive value to personal data: Lower disclosure requirements significantly increase the rate at which applications are completed. The researcher quantifies the monetary value of personal data--and the welfare effect of various disclosure policies--by developing a structural model that links individuals' disclosure, borrowing, and repayment decisions. Using detailed application-level data, Tang estimates that social network ID and employer contact are valued at 230 RMB (i.e., $33, or 70% of the average daily salary in China); for successful borrowers, this accounts for 8% of the average net present value of a loan. Requiring answers to these application questions reduces borrower welfare by 13% and costs the platform $0.50 in expected revenue per applicant.

How do inexperienced consumers learn to use a new financial technology? Breza, Kanz, and Klapper present results from a field experiment that introduced payroll accounts in a population of largely unbanked factory workers in Bangladesh. In the experiment, workers in a treatment group receive monthly wage payments into a bank or mobile money account while workers in a control group continue to receive wages in cash, with a subset also receiving an account without automatic wage payments. The researchers find that exposure to payroll accounts leads to increased account use and consumer learning. Those receiving accounts with automatic wage payments learn to use the account without assistance, begin to use a wider set of account features, and learn to avoid illicit fees, which are common in emerging markets for consumer finance. The treatments have real effects, leading to increased savings and improvements in the ability to cope with unanticipated economic shocks. Breza, Kanz, and Klapper conduct an additional audit study and find suggestive evidence of market externalities from consumer learning: mobile money agents are less likely to overcharge inexperienced customers in areas with high payroll account penetration. This suggests potentially important equilibrium effects of introducing accounts at scale.

Using data from the United States and Canada, Felt, Hayashi, Stavins, and Welte quantify consumers' net pecuniary cost of using cash, credit cards, and debit cards for purchases across income cohorts. The net cost includes fees paid to financial institutions, rewards received from credit or debit card issuers, and the merchant cost of accepting payments that is passed on to consumers as higher retail prices. Even though credit cards are more expensive for merchants to accept compared with other payment methods, merchants typically do not differentiate prices at checkout, but instead pass through their costs to all consumers. As a result, credit card transactions are cross-subsidized by cheaper debit and cash payments. Card rewards and consumer fees paid to financial institutions are additional sources of cross-subsidies. The researchers find that consumers in the lowest-income cohort pay the highest net pecuniary cost as a percentage of transaction value, while consumers in the highest-income cohort pay the lowest. This result is robust under various scenarios and assumptions, suggesting payment card pricing and merchant cost pass-through have regressive distributional effects in the United States and Canada.