Monitoring the household sector with aggregate credit bureau data

new and promising tool has recently become available for tracking and forecasting consumer borrowing and payment behavior. For many years, credit grantors and insurance firms have used consumer credit reports to evaluate the repayment risk of individual applicants for loans and insurance. The rich detail in individual credit reports has supported the development of sophisticated statistical models that estimate an individual's repayment risk with remarkable accuracy. However, until very recently, this information was available only at the individual level. Timely, reliable data on consumer borrowing and payment activity, aggregated to the local, state, or regional level, have been largely unavailable except through proprietary marketing research surveys or infrequent government-sponsored household interviews.

Actual observations (as opposed to self-reported survey responses) on borrowing and payment behavior could have enormous value in calculating household debt burden, forecasting consumer spending behavior, estimating demand for consumer durables at the local or regional level, and benchmarking portfolio performance. The three major U.S. credit bureaus have recognized the value of aggregating the individual-level information in their archived files and are beginning to market such data products.

This article provides examples of how aggregated credit bureau data can be used for benchmarking and modeling to identify the factors that influence bankruptcy and delinquency trends at the county level. The following sections utilize a unique database assembled by Trans Union LLC. Dubbed TrenData1, this new tool is based on a series of large random samples of U.S. consumer credit histories drawn quarterly since 1992. Each quarterly sample contains approximately 30 million depersonalized credit reports. From this underlying database, variables have been built to describe consumer borrowing and payment behavior aggregated to the county, state, and national level. The Credit Research Center (CRC) at Georgetown University's McDonough School of Business is collaborating with Trans Union to explore the predictive value of TrenData variables.2

Advantages of Aggregated Credit Bureau Files The rich detail of individual credit file data supports

the creation of a host of aggregate variables. TrenData provides more than two hundred variables for analysis, all aggregated to the county level on a quarterly basis. The following brief list conveys the scope of what is available:

Average mortgage, installment, and revolving debt, per borrower

The percent of bank card holders thirty, sixty, or ninety days past due

Percent of revolving credit lines utilized, per borrower

Dollar amount of new automobile credit extended in the previous three-month period

Average monthly minimum debt payment (consumer + mortgage), per borrower

Number of new installment or revolving loan accounts opened in the previous three months.

Do aggregate credit bureau data provide additional insights into the current environment? Two examples illustrate the contribution of this new tool. First, consider the most widely-used measure of the rate of growth of consumer installment credit. The Federal Reserve Board (FRB) has reported that the total dollar amount of consumer (nonmortgage) credit grew 6.2 percent during the twelve months ending with the first quarter of 1999, up from a 4.1 percent growth pace during the twelve-month period ending in first quarter, 1998. Because TrenData is constructed from individual borrower files, it can provide additional insight regarding the composition of the aggregate growth. Interestingly, TrenData indicates that the average amount of nonmortgage debt per borrower actually fell slightly (3.2 percent) from the first quarter, 1998, to the first quarter, 1999. However, during the same period account openings accelerated. The average number of new accounts opened each quarter from March 1998 through March 1999 was 25.2 per hundred borrowers, as compared to an average quarterly rate of 23.0 per hundred borrowers during the twelve months from March 1997 to March 1998. The small decline in debt per borrower during a period of rapid account openings and rising aggregate consumer installment debt suggests that borrowers with little or no previous debt accounted for much of the aggregate growth. This insight can not be derived from the FRB aggregate statistics alone.

A second, major advantage of aggregated bureau data is in developing a more accurate picture of regional, state, and local differences in borrowing behavior. Delinquency statistics provide a good example. Traditional portfolio benchmarking tools which rely on the portfolio-wide experience of credit grantors are becoming less useful for identifying regional and local trends because the experience on which they are based is increasingly national, not local. The delinquency surveys conducted by associations such as the American Bankers Association and the Mortgage Bankers Association are dominated by institutions whose portfolios are nationwide. When these surveys publish state-level data, they reflect the experience of the institutions located within the state, not necessarily the customers who reside there.

TrenData reveals striking differences in delinquency behavior at the regional, state and county levels. Figure I reveals a wide divergence in loan performance across states during the first quarter, 1999. The chart shows 30+ day delinquency rates (in terms of borrowers, not accounts) on two types of bank loans-bank cards and closed-end nonmortgage loans (including automobile loans). To be precise, the delinquency rates are calculated as the percent of borrowers holding a particular type of credit who are delinquent 30+ days on one or more accounts. At the state level, the proportion of bank cardholders delinquent 30+ days ranged from a low of 1.95 percent (Wisconsin) to a high of 5.24 percent (Mississippi). For borrowers with closed-end (nonmortgage) accounts at banks, the proportion ranges from a low of 2.45 percent (Oregon) up to 6.28 percent (Mississippi).

Contrast these TrenData statistics for the first quarter of 1999 with a similar measure available from the American Bankers Association's (ABA) Consumer Credit Delinquency Bulletin. Table I compares the ABA bank card delinquency rates (percent of accounts 30+ days past due) at the state level with the TrenData bank card delinquency rates (percent of bank card borrowers 30+ days past due). Although not directly comparable, it is interesting to note the divergence between each of the two measures across states. Moreover, the ABA series does not report a state-level delinquency rate if the number of reporting institutions is too small to protect the anonymity of the responding institutions. Consequently, twentyseven states had no bank card delinquency estimates in the ABA series.

Figure I emphasizes that drilling below national aggregates can yield a much different, multifaceted picture of consumer borrowing and payment trends. For example, we know that loan portfolio performance has regional/local drivers, including economic factors (income, unemployment, job growth, base/plant closings, casino gambling, etc.) and demographics (age, divorce, insurance coverage, education, homeownership). Regional and local credit market data give the portfolio analyst the raw material necessary to model the trends affecting all borrowers in the market area and to identify whether (and why) a single portfolio's experience differs from the broader population of customers. The following sections provide modeling examples at the local level for two distinct observations: personal bankruptcy and loan delinquencies.

Using Aggregate Bureau Data to Model Bankruptcy Filing Rates

Personal bankruptcies in the United States have soared over the past four years. During the twelve months ending December 31, 1998, nearly 1.4 million U.S. households filed for bankruptcy protection, about a half million more households than in 1995. Perhaps more striking is the fact that from 1992 through 1998, one out of every twenty U.S. households filed for bankruptcy, a total of over 5 million. Losses to creditors from personal bankruptcy (debts discharged through bankruptcy proceedings) in 1998 alone exceeded $45 billion. Yet, this dramatic growth in personal financial failures occurred against the backdrop of the most favorable economic conditions since World War 11. The apparent paradox of soaring bankruptcies coincident with extraordinarily low unemployment and strong income growth has puzzled researchers and spurred Congress to consider legislative remedies.

Debtor surveys consistently find that the majority of personal bankruptcies are triggered by "insolvency events." Bankrupt debtors cite events such as layoffs, loss of overtime hours, failure of a small business, divorce, extended illness and disability as creating either interruptions in income or unexpected expenses that result in repayment problems. However, an explanation for the rise in bankruptcies that is built around a comparable rise in insolvency events does not seem consistent with the marked improvement in the economic climate since 1992. In addition, a large minority of debtors (twenty-five to thirty percent) cites "buying too much on credit" as the primary cause of their financial woes.3

Consequently, two alternative explanations have gained popularity. Critics of the credit-granting industry argue that increasingly lax credit standards over the period have encouraged consumers to run up their debt loads. This is grounded in the observation that individuals have more credit available to them, and more individuals can get credit today than was the case a decade ago.4 Proponents of this view especially demonize credit card issuers as a driving force behind the escalation in bankruptcies.

In the philosophically opposite camp are those who argue that declining social and economic stigma to filing for bankruptcy has led to more filings in at least two ways. First, declining stigma causes borrowers to live "closer to the edge." If the apparent cost to using the bankruptcy safety net falls, households will be willing to take on more debt relative to their income. Doing so makes them more vulnerable to unexpected income or expense shocks, with a resulting higher frequency of financial problems, delinquency and bankruptcy. Note that the theorized link between rising debt-to-income ratios and a greater frequency of bankruptcy is essentially the same as in the "easy credit" explanation.

Additionally, as the perceived cost of using the bankruptcy safety net falls, borrowers with any given amount of debt become more likely to seek a discharge of their debts through bankruptcy, rather than incur the costs of repayment over time. In other words, the bankruptcy trigger is pulled sooner in the borrower's debt cycle, relative to when filing stigma was higher.5 Note that this second consequence of declining stigma does not require a concurrent rise in either insolvency events or debt/income, nor does it imply an increase in aggregate delinquencies or other conventional signs of financial problems.

To summarize, there are essentially three competing explanations for the rise in personal bankruptcies during the latter half of the 1990s: (1) a rise in the frequency of external shocks to household income or expenses, (2) build-up of debt relative to income, or (3) increased willingness to seek discharge of debts through bankruptcy for any given debt-to-income level. The availability of aggregate credit bureau data affords an unprecedented opportunity to test these explanations via a regression model built to explain county-level bankruptcy filing rates from 1993 through 1998. The TrenData bureau database provides detailed controls for factors such as credit usage, borrower risk and credit supply responses, which are necessary for disentangling the competing hypotheses. Such detailed controls have not previously been available to researchers.

The Empirical Model

Figures 2 and 3 display the underlying trend that the model attempts to explain. The regression model is based on annual observations for over three thousand U.S. counties from 1993 through 1998 (18,313 observations). The model must account for two characteristics of personal bankruptcy filings during the period: a wide variance in filing rates across counties in any given year and a substantial increase in annual filing rates beginning in 1995. Figure 2 shows that filing rates per thousand households varied in 1993 from about 3.0 in the lowest decile of U.S. counties up to 18.2 in the highest decile of counties. Figure 3 displays the dramatic increase in the national filing rate from about 8.0 per thousand in 1994 to nearly 14.0 per thousand in 1998.

The empirical model estimates county-level bankruptcy rates per thousand households as a function of county-level variables that reflect: (1) household decisions to use more debt relative to income, (2) incidence of (and vulnerability to) unexpected declines in income or increases in expenses, and (3) social/economic stigma that accompanies filing for personal bankruptcy. All independent variables are lagged one year unless otherwise noted to reflect the common survey observation that bankruptcy petitioners struggle with financial problems for a year or more before filing.6

Independent variables in the first category include household income, household debt, and the number of revolving accounts per revolving user. Since bankruptcy can be triggered by unexpected income interruptions, household income enters the model through two variables, the prior year's average income and the change in average income between the prior and current year. Similarly, all debt variables enter the model with values for the prior year and the change between the prior year and the current year. Although household debt is not directly contained in the TrenData database, it can be decomposed into three TrenData components as follows:

Bankruptcy can be triggered by excessive debt relative to income. Holding income constant, counties with more debtors per household or higher amounts of mortgage or nonmortgage debt per debtor should have higher bankruptcy filing rates. Conversely, counties with higher income, holding the debt variables constant, should experience lower rates of filing.

The inclusion of the number of revolving accounts in addition to the total amount of debt is a proxy for the average level of risk of the population of debtors in the county, as reflected in creditor supply decisions. Creditors view individual credit files to assess individual risk and make their lending decisions accordingly. A decision to extend a revolving line with a lower limit signals a creditor's assessment that the borrower is riskier, relative to a second borrower who received a higher limit. Consequently, an increase in the number of accounts in an area, holding constant the total amount of household debt, implies a riskier population and higher likelihood of bankruptCy.7

Independent variables that capture household vulnerability to insolvency events include the state-level unemployment rate, change in unemployment rate from the prior year, the proportion of individuals divorced or separated, proportion of households with at least some health insurance, the value of housing, and the proportion of individuals over the age of fifty. Bankruptcy filing rates are hypothesized to rise with both unemployment and the divorce rate. Bankruptcies should fall as more of the population is covered by health insurance. The market value of housing, when coupled with average mortgage debt, reflects the average amount of home equity. This serves as a proxy for (1) the level of household assets available as a cushion against income interruptions or expense shocks, (2) how much equity value would be given up in a Chapter 7 bankruptcy (which requires liquidation of nonexempt assets in order to pay off creditors), and (3) the general level of risk of borrowers in the area. All three interpretations imply the same expectation: Higher average house values imply smaller likelihood of bankruptcy, other factors held constant. Finally, a higher proportion of borrowers over the age of fifty should reduce the bankruptcy filing rate. Asset holdings and net worth rise with age. Consequently, older borrowers are less vulnerable to external income and expense shocks because they tend to have more assets available for liquidation.

Variables that capture the effects of social and economic stigma include population density, the proportion of households over the age of fifty, a dummy variable for counties in states with an unlimited bankruptcy homestead exemption, a dummy variable for counties in states that exempt delinquent debtors from wage garnishment, and a set of time dummies for 1994-1998. County population density reflects the effect of anonymity in reducing the reputational costs of filing for bankruptcy in more densely populated areas. Consequently, counties with higher population density should experience higher filing rates. Conversely, social stigma is hypothesized to be higher for older borrowers, whose attitudes were formed decades earlier during a period when bankruptcies were far less common. Counties with older borrowers should experience lower filing rates.

An unlimited homestead exemption allows a debtor to protect the full value of home equity from liquidation through the bankruptcy process. Bankruptcy rates should be higher in these counties. Wage garnishment is a creditor collection tool that a delinquent debtor can escape by filing for bankruptcy. Consequently, a debtor's advantage to filing for bankruptcy is lower in states that prohibit garnishment, which should lead to lower filing rates. The time dummies for years 1994 through 1998 are included to detect any effects from an across-the-board decline in stigma over the past five years, independent of local effects related to population density. Table 2 displays the sources for all variables.

Results

To estimate the determinants of bankruptcy filing rates across counties and over time, we adopted the generalized estimating equation (GEE) extension to the generalized linear model (GLM) random-effects estimator.8 We assume a first-order autoregressive structure for filing rates within areas, as there is clear evidence that a given county's bankruptcy filing rate is directly correlated to prior filing rates. We assume the bankruptcy filing rates are distributed Poisson (also known as the log-linear model).

Table 3 reports the estimates. First notice that the coefficients on nine out of the ten debt variables (reflecting both prior year levels and change over the past year) are significant with the predicted (positive) sign. Holding household income and other factors constant, higher mortgage and nonmortgage debt levels per debtor are associated with higher bankruptcy rates, as is a larger average number of debtors per household. Clearly, household decisions to take on higher debt loads have contributed to the rise in bankruptcies over the past decade.

Interestingly, even after controlling for the amount of debt, the lagged number of revolving accounts per revolving user and the change in revolving accounts per user are positively related to the bankruptcy filing rate. This result supports the hypothesis that spreading a given amount of debt over a larger number of accounts signals a riskier population. Higher risk individuals (who ultimately overextend) apparently obtain smaller revolving lines from more lenders, behavior that can be explained if individual creditors limit the credit line on a given account when they perceive a higher risk applicant.

Counties with higher average household income had lower bankruptcy rates, as predicted. Similarly, changes in average household income were inversely associated with filing rates, consistent with the hypothesis that insolvency events in the form of income interruptions precipitate a bankruptcy decision.

Local economic and demographic factors contribute to bankruptcy filing rates. All of the local variables that capture household vulnerability to insolvency events were significant with the predicted sign. Higher unemployment rates, a rise in the unemployment rate, higher divorce/separation rates and less health insurance coverage all contribute to higher bankruptcy filing rates. Conversely, higher average housing values, and a higher proportion of residents over the age of fifty tend to lower bankruptcy filing rates. As a proxy for local-level social stigma, population density was positively associated with filing rates. Exemption from wage garnishment was a powerful predictor of lower bankruptcy filing rates, suggesting that debtors do take a calculating approach to handling financial problems and opt for bankruptcy when the advantages (in this case, escape from court-ordered garnishment) outweigh the costs. Of all the variables included in the model, only the dummy variable for the unlimited homestead exemption failed to add significant explanatory power to the model.

Lastly, the time dummies for years 1996, 1997, and 1998 reveal a significant increase in bankruptcy filing rates even after controlling for debt growth, number of accounts, a variety of insolvency events and local-level stigma effects. Recall the earlier discussion of two separate effects of declining stigma. To the extent that declining stigma has increased consumer willingness to take on more debt, those effects are captured in the debt and account growth variables. Consequently, the significance of the time dummies suggests that the second manifestation of declining stigma, i.e., an increased willingness to file for any given level of debt relative to income, may also have contributed to the dramatic surge in bankruptcies over the past three years.9

Table 4 illustrates the relative importance of the debt and income variables in driving the growth in bankruptcy filing rates. The table reports the model's predictions of the U.S. bankruptcy filing rate in 1998 under two simulation scenarios. The "stagnant economy" scenario assumes no growth in real income or decline in unemployment between 1994 and 1998. That is, if the 1994 values for income and unemployment had prevailed simultaneously with 1998 values of debt and other variables, the bankruptcy filing rate would have been 15.69 per thousand households, or thirteen percent higher than the actual 1998 rate. Alternatively, if real-debt growth were stagnant at the same time the income and employment picture improved, the bankruptcy filing rate in 1998 would have been 10.06 per thousand households, or 27.0 percent lower than the actual 1998 rate.

Modeling Consumer Delinquency Rates Information about consumers' borrowing and payment

behavior is valuable for making economic forecasts and business decisions. Recent research indicates that delinquencies are significantly related to consumer spending and the performance of the U.S. economy.10 Creditors are concerned with the level of delinquencies because they affect cash flows and profitability. Rising delinquencies reduce cash inflow from debt payments and increase collection expenses. Understanding the determinants of delinquency is helpful for analyzing the performance of credit portfolios and strategic planning for household credit products.

Because of data limitations, past analyses of consumer delinquencies have generally been limited to studies of the determinants of aggregate delinquency rates (e.g., Sullivan, 1986; Peterson and Luckett, 1976; Moore and Klein, 1967)." However, as illustrated in Section 1, delinquency rates vary greatly by geographic location. Aggregate statistics mask much of the variation. Differences are visible even across large geographic areas such as Census divisions. For example, delinquency rates of sixty or more days for bank card borrowers were higher in the South than in the other Census regions through most of the 1990s (Figure 4). Delinquencies across all regions generally rose during this period, but the differences in delinquency rates for bank card borrowers in the South and the other regions widened during the 1990s. These findings suggest that the riskiest region, the South, became riskier during the 1990s. This example illustrates the potential value of studying determinants of delinquency at geographically disaggregated levels. However, as mentioned, geographically disaggregated data on borrowing and debt burdens, which are critical determinants of delinquency, have not previously been available.

The Empirical Model

The empirical model for explaining delinquencies is similar to the one for explaining bankruptcies. The model has been modified to reflect some of the conceptual differences between these two types of financial distress. Generally, bankruptcy is a consequence of serious financial distress that results in insolvency. Delinquency typically results from illiquidity, which normally is resolved over a relatively short period of time and need not lead to insolvency and bankruptcy. Delinquency also does not necessarily affect the entire balance sheet. A consumer can be delinquent on one account and pay other accounts on schedule.

We consider delinquencies of sixty or more days for three different types of accounts: (1) closed-end consumer installment credit, (2) retail accounts, and (3) revolving consumer credit accounts. Delinquencies are measured as the percentage of borrowers using a specific type of credit who are currently sixty or more days past due on at least one of those accounts. Delinquencies of sixty or more days are considered serious, but most delinquencies of this degree of seriousness do not lead ultimately to default (chargeoff). Delinquency rates are averages of quarterly delinquency rates for the year.

The independent variables in the delinquency models are the same as the independent variables in the bankruptcy model with the following exceptions:

The number of debtors per household is omitted from the delinquency models because the dependent variable in these models (delinquency per borrower) is based on the borrower, not the household.

The average mortgage debt per borrower, the proportion of borrowers with mortgage debt, and the average value of housing are omitted from the delinquency models. The value of housing assets is closely associated with solvency questions, which would affect the decision to default on mortgages or file for bankruptcy. These variables are less closely associated with either the cause or response to short-term liquidity problems, which might involve making late payments on nonmortgage, consumer credit obligations.

The set of variables on number of accounts and amount of debt was expanded to include variables specific to the type of delinquency being estimated.

Evidence on Determinants of Delinquency

The three models of county-level delinquency rates were estimated using the general estimating equation extension of the generalized linear model random-effects estimator. Again, we assume a Poisson distribution for delinquency rates and first-order autoregressive structure within areas.

Table 5 reports the estimates. The results suggest, not surprisingly, that local measures of credit use are important determinants of county-level delinquency rates. Credit variables are significant in the delinquency regressions for each of the three types of credit. Consistent with the hypothesis that the spreading of a given amount of debt over a larger number of accounts signals a riskier population, both the average number of accounts per borrower and the change in accounts from the prior year were significant and positive for all three types of credit. Decisions to take on greater amounts of debt also contribute to higher delinquency rates. The level and change in category debt per borrower is significant and positive in the closedend installment and retail delinquency regressions.

Contrary to expectations, both a greater amount of total consumer debt per borrower and an increase in that amount from the prior year are associated with lower, not higher, delinquency rates in the closed-end installment delinquency regression. However, an increase in total consumer debt increases revolving delinquencies. This result may reflect a supply effect as well as reveal something about consumers' preferred method of adjusting their debt holdings. First, lower-risk consumers are able to use more consumer debt of all types, hence the coincidence of higher debt levels and lower delinquencies in the installment delinquency equation. However, in the revolving delinquency equation, notice that the variables capturing a change in total debt per borrower and a change in revolving debt per borrower are both positively associated with revolving delinquencies, but only the total debt variable is significant. It may be simply that changes in total debt holdings are more likely to be in revolving accounts rather than installment accounts, and in this particular equation the two variables are capturing the same effect. Of course, we have already noted that a larger number of accounts, holding the amount of debt per borrower constant, clearly signals risk for all three types of credit.

Local economic and demographic variables also help explain county-level delinquency rates. Of the local variables that reflect vulnerability to financial distress, all but the proportion of adults who are divorced or separated are significant in each delinquency regression. All significant variables have the predicted sign. Higher income, increases in income, lower unemployment, decreases in unemployment, and greater health insurance coverage are associated with lower delinquency rates. The proportion of adults over age fifty is significant and negative in all three regression, and population density, a proxy for locallevel social stigma, is significant and positive. Reflecting effects of the legal environment, the dummy variable for unlimited home exemption is significant in the closed-end installment credit regression, and the dummy variable for exemption of wages for garnishment is significant in the revolving credit regression. The latter result is noteworthy for its consistency with theoretical models of creditor remedies (for example, Jaffee and Russell 1976; Barro 1976), which predict that lowering the cost of default reduces borrowers incentives to repay, thereby increasing the probability of default. Because much of revolving credit is unsecured, the boost to revolving delinquencies associated with the exemption from garnishment is not surprising.

Finally, the time dummies for years 1996, 1997, and 1998 indicate significant increases in delinquency rates after controlling for other factors. These results are similar to results in the bankruptcy regression and are consistent with declines in the second type of stigma--that is, an increase in the willingness to default on payments for any level of debt relative to income.

Benefits of Geographic Disaggregation

The preceding section indicates that the county-level data on credit use are important explanatory variables for bankruptcy and delinquency rates. To evaluate the predictions of models estimated from the county-level data, we computed predicted values of each observation using the estimated coefficients of each regression model. These predicted bankruptcy or delinquency rates for each county/time period were regressed on the actual county/time period bankruptcy or delinquency rates. The adjusted R2 values of these regressions presented provide a measure of the "goodness of fit" between the two sets of values. The results are presented in the first column of Table 6. Estimated models explained 42.63 percent of the variation in county bankruptcy rates, 42.14 percent of the variation in county delinquency rates for closed-end consumer installment credit, 60.77 percent of the variation in county delinquency rates for retail credit, and 45.41 percent of the variation in county delinquency rates for revolving credit.

To simulate a situation where only state-level data were available, we aggregated the county-level data to the state level, and estimated the same model using the statelevel data. The predicted state delinquency rate then serves as a proxy for the county-level rate. To measure the effectiveness of the proxy, the predicted state delinquency rates were regressed on the actual county rate. The adjusted R2 values of the resulting equations are shown in the second column of Table 6. The state-level predictions were moderately less accurate than county-level predictions of bankruptcy rates (37.66 percent versus 42.63 percent) and slightly lower than county-level predictions of revolving credit delinquency rates (43.78 percent versus 45.41 percent). The state-level predictions were considerably less accurate than county-level predictions for closed-end consumer credit delinquencies (26.72 percent versus 42.14 percent) and retail credit delinquencies (42.74 percent versus 60.77 percent). These results indicate that analysts interested in predicting bankruptcies or delinquencies in local geographic markets will explain more, sometimes substantially more, variation using county-level data than if they use state level-data. These results provide strong evidence of the utility of disaggregated data from TrenData for local-level analysis of consumers' borrowing and payment behavior.

Conclusion

Research is currently underway at the Credit Research Center to explore the relationships among the aggregated credit bureau variables (e.g., Do bank card delinquencies lead installment loan delinquencies? Does a surge in new account openings foreshadow rising delinquencies?) and identify those with predictive power. Perhaps more intriguing is the prospect that the bureau variables themselves may explain other economic variables, such as consumer spending or durable goods purchases. For example, economists have shown that consumer debt payment burden appears to influence future consumer spending (Murphy, 1998). However, virtually all such studies in the past have been based on estimates of debt payment burden rather than the more precise variables now available from aggregated bureau files.

The analytical challenge is to merge local economic/demographic data with TrenData variables to develop new forecasting tools. Since TrenData's coverage begins in 1992, it does not yet span an entire business cycle. Consequently, its usefulness as a tool for business cycle forecasting is still limited. However, as experience accumulates over time, through the inevitable downturn, the aggregated bureau variables should provide the foundation for developing predictive indicators that link borrowing behavior with other economic variables such as employment and consumer spending. N