U.S. households have increasingly large amounts of debt. From 1945 to the second quarter of 2009, household liabilities grew from 20% of disposable personal income to 129%. The increase in liabilities has been driven by both increasing mortgage debt (as a percent of real estate assets and as a percent of disposable personal income) and an increase in consumer credit. At the end of the second quarter of 2009, home mortgages stood at $10.4 Trillion, with consumer credit adding another $2.5 Trillion to household liabilities.
A central issue in the context of household debt is default. Compared to the large literature on determinants of household portfolio choice, relatively little is known about what drives households' choices of debt levels and their decision of whether and when to default. Industry models for predicting default emphasizes past borrowing and repayment behavior. For example, in the calculation of an individual's FICO score (a commonly used measure of credit risk in the U.S.), a weight of 35% is given to on time payment of past debt, 30% weight is given to the current amount of debt of various types, how many accounts the individual has, and how large the debt is relative to the total available credit, 15% weight is given to the length of time of credit history, 10% to the number of new accounts and recent requests for credit, while 10% is given to the mix of credit (credit cards, installment loans, finance company loans, and mortgages) used in the past (Fair Isaac Corporation (2005)). FICO scores do have predictive power for default (e.g. Keys, Mukherjee, Seru and Vig (2010) in the context of mortgage delinquencies). However, from the perspective of understanding the underlying economic drivers of debt and default, predicting default based on past repayment behavior is not informative. Furthermore, models predicting default tend to have a modest statistical fit. For example, Gross and Souleles (2002) find a pseudo-R of about 0.14 in a probit model predicting credit card delinquency using time effects, account age, measures of account risk (including credit scores), and local economic conditions as explanatory variables.
This paper seeks to improve our understanding of household default from both an economic and a statistical perspective. Using a new proprietory data set from one of the largest retail chains in Mexico I document that information about which products a customer buys provides substantial information about potential default losses on a given loan. The data set is large both in terms of the number of borrowers covered -- about half a million -- and in its panel dimension, with monthly data available from January 2005 to August 2009. The unique feature of the data which enables me to study the link between which goods are purchased and subsequent loan losses for the lender is that each good purchased at this retailer has its own loan associated with it. For example, a customer may first take out a loan to buy a washing machine and then later take out a separate loan to buy new tires for her car.
I find that lender loss rates (measured as the ratio of the amount not repaid to the size of the loan), are dramatically higher for certain types of products than others. Lender losses are low (below 12%) on loans for appliances, kitchen equipment and furniture, while they are about 21% on loans for electronics (cell phones, stereos, TVs etc.), and almost 40% on loans for jewelry purchases. These differences in default losses across product categories are robust to controlling for characteristics of the loan contract (e.g. the size of the loan and interest rate on the loan), demographics, and more standard measures of credit risk based on past repayment behavior, and the differences do not diminish substantially with how long the borrower has been a customer. From a statistical perspective this implies that which products people borrow money to buy is a useful additional predictor of subsequent default, above and beyond known predictors documented in past work.
To begin assessing the economic forces underlying differential lender losses across product categories, I estimate models of loss rates that include customer fixed effects. With these fixed effects included differences in loss rates across product categories are economically small. This indicates that differential loss rates across product categories are driven mainly by which types of individuals buy particular products, as opposed to being product-specific features. In other words, customers who tend to buy electronics generate high lender loss rates both when they buy electronics and when they buy other products.
This raises the question of why people with preferences for particular goods on average are worse risks. I find that high loss products tend to be luxuries and that they tend to be purchased by individuals who (in their spending at the retail chain analyzed) consumer abnormally large fractions of luxuries given their income. A one standard deviation increase in the fraction spent on luxuries (controlling for income), increases the predicted lender loss rate by about 3.4 percentage points. I confirm that the link between preferences for luxuries and getting into repayment difficulties is not specific to the particular retail chain studied or to Mexico. Using U.S. data from the 2003-2007 Consumer Expenditure Survey I find that households who consume abnormally large fractions of luxuries given their overall consumption have larger amounts of consumer credit (an extra $300 for a one standard deviation increase in the fraction spent on luxuries) and incur larger amounts of finance charges on their consumer credit.
While many aspects of budget constraints or preferences (income risk, discount rates, the elasticity of intertemporal substitution, or risk aversion) could happen to be correlated with people’s preferences for luxuries, I discuss whether the existing literature on consumer choice includes theories that have predictions about both which products particular people tend to buy and why particular people would have higher default rates on consumer credit. The marketing literature (e.g. Dhar and Wertenbroch (2000)) distinguishes between hedonic goods which provide more experiential consumption, fun, pleasure, and excitement, and utilitarian goods which are primarily functional. Utilitarian goods tend to be necessities while hedonic goods tend to be luxuries. Furthermore, Hoch and Loewenstein (1991) model consumption outcomes as the result of an inner struggle between a desire for immediate gratification (i.e. hedonic pleasure, or indulgence) and self-control. I argue that high spending relative to available resources (and thus high default risk) and spending an abnormally high fraction on luxuries may be two dimensions of having a high preference for indulgence or low self-control to overcome impulses to indulge. I discuss possible field surveys and experiments at the Mexican retailer that could further shed light on this interpretation of the evidence.
The paper's findings have immediate implications for lenders. The fact that people who tend to buy certain products generate larger loss rates implies that providers of consumer credit could benefit from adjusting credit terms (down-payment requirements, interest rates, or credit limits) as a function of product mix purchased to date, and thus that product mix should be an important component in credit scoring. This requires lenders to estimate the borrower fixed effects in ``real time'', i.e. using only information available up to the data of purchase. I propose that lenders use the average (product level) default rate of products purchased by a particular customer to date as an additional indicator of the customer's creditworthiness, above and beyond standard predictors of default such as credit scores and demographics.
Since an increasing fraction of consumer debt is securitized, my findings also suggest that information about which products (and in turn which types of buyers) a given security was issued to finance could help price the security more accurately. More broadly, understanding what drives heterogeneity in borrowing and default behavior across households in the market for consumer credit is likely to be informative for other loan markets and for understanding
consumption and savings behavior more generally. For example, the same factors that make some individuals buy more luxurious consumer goods than would be predicted by their income may be relevant for understanding the behavior of home buyers some of whom may buy more luxuriously houses than suggested by their income, with corresponding implications for mortgage default. Furthermore, if substantial preference heterogeneity can be documented within a sample of households who all have consumer debt, one would expect an even larger degree of preference heterogeneity in the full set of households, with correspondingly broader implications for explaining heterogeneity in household net worth.
The paper is organized as follows. Section II describes the data set and how lending works at this retailer. Section III documents differences in loss rates across product categories. Section IV documents that these differences are driven mainly by high-risk borrowers disproportionately buying certain products as opposed to being product-specific characteristics. It also discusses how lenders can construct an estimate of a given borrower's type using real-time data. Following this, section V turns to the link between borrowers' risk and their preference for luxury goods as well as to what more fundamental heterogeneity may drive this link. Section VI concludes.
Download
PDF Ebook Consumer Credit: Learning Your Customer's Default Risk from What (S)he Buys
