BUSINESS
... to total assets, the ratio of C&I lending to total assets, the ratio of CRE lending to total assets, the ratio of residential real estate lending to total assets, the ratio of other real estate owned to total assets, the ratio of long-term deposits to total assets, and the ratio of thirty-day past-due loans to total assets. The state-level regional economic variables they used were payroll employment growth, residential house price appreciation, and total personal income growth. 6 See the papers by Jayaratne and Strahan (1996) and Morgan, Rime, and Strahan (2002) for research on the impact of banking deregulation on local economic variables. 6 and the amount of failed business liabilities, in order to predict bank failures and bank asset quality. Their basic result was that economic variables add little information to the forecasts of these bank-specific variables. Jordan and Rosengren (2001) investigated whether forecasts of regional economic variables had an impact on the supervisory CAMELS ratings assigned to banks.5 They found that contemporaneous measures of regional economic conditions did not add explanatory information over the bank-specific variables already used in bank surveillance models. However, one-year-ahead forecasts were found to be both economically and statistically significant in improving the predictive power of the surveillance models that assess bank performance out four quarters. Furthermore, they found that these effects were more pronounced during difficult regional economic periods. An important caveat to this analysis and to our data set is that bank deregulation in the U.S. has led to consolidation and to banks¡¯ geographic expansion.6 As noted by Morgan and Samolyk (2003) in their study of bank geographic diversification, U.S. banks have not only become bigger during the course of the 1990s, but they have also become wider by expanding their operations across multiple banking markets. Using a geographic diversification index based on deposits, the authors found important differences across bank size categories and 7Crone originally constructed his indices for states based on four monthly data series¡ªthe total number of jobs in nonagricultural establishments, real retail sales, average weekly hours in manufacturing, and the unemployment rate. 8These series have been made publicly available by the Federal Reserve Bank of Philadelphia. 7 over time. We address this concern only in terms of constructing the regional variables corresponding to the inter-state banks in our sample; see the discussion in the following section. III. Data and Measurement Issues Regional Composite Indicators Tracking the health of regional economies has long been a difficult job. Analogs to traditional measures of national economic performance such as GSP (GDP analog) are produced with a significant lag and only at an annual frequency, limiting their usefulness for most research endeavors. Other measures of performance including employment growth, commercial and residential building permits, or personal income growth, while more current, are less complete, capturing only a part of the picture of economic activity in a region. The incompleteness of single measures of economic activity is well recognized (Zarnowitz, 1992), prompting the NBER to look across many series to date business cycles (Rudebusch, 2001). Building on the work by Stock and Watson (1989) who develop a coincident index for the national economy, Crone (1994) began developing composite indices of regional economic activity.7 The indices produced proved useful for tracking regional economic trends and for dating regional business cycles (Crone (1999)). In 2002, Crone produced a set of consistent economic indexes for the 50 states.8 The indexes are produced at a monthly 8 frequency and cover the period from 1978 through 2002. As constructed by Crone (2002) the coincident indexes for the 50 states include three monthly indicators¡ªnonagricultural employment, the unemployment rate, and average hours worked in manufacturing¡ªand one quarterly indicator¡ªreal wage and salary disbursements¡ªof regional economic conditions. To ensure consistency, Crone applies the following criteria: (1) The indexes are constructed from the same set of indicators for each state (2) The timing of the index is benchmarked to employment in each state. (3) The trend for the index corresponds to (GSP) in each state. Figure 5 shows a scatter plot of average annual growth between 1986 and 2000 in GSP and the composite indicator growth for the 50 states. The composite indicators for each state are aggregated to the annual frequency for comparison with GSP. As the figure shows, the composite indicator tracks GSP quite well. Other Regional Indicators By and large, the literature on bank performance and regional economic conditions has relied on one or more regional indicators including house price appreciation, employment growth, personal income growth, and the unemployment rate. To tie this paper to previous work in the area and to evaluate the usefulness of composite measures relative to others, these variables are included in the analyses. House price appreciation by state is measured using data from the Office of Federal Housing Enterprise Oversight (OFHEO). OFHEO produces an index of home prices by state on a quarterly basis. These indices are used to compute yearover- year growth rates for house prices in each state. The personal income data come from 9Further information on and access to the SOD data is available at http://www2.fdic.gov/sod/index.asp. 9 the Bureau of Economic Analysis and are released quarterly. Again, the quarterly levels are used to compute year-over-year growth rates. Data on employment and unemployment come from the Bureau of Labor Statistics. The data are released monthly. The analysis includes year-over-year changes in these variables at a quarterly frequency. Bank-Specific Regional Indicators A key challenge in this analysis is the problem of correctly identifying the regional economy for an individual bank. An obvious choice is to simply define the state of headquarters as the bank¡¯s region. Such a choice was accurate before the advent of cross-state banking laws in the early 1980s and the Riegle-Neal Act of 1994 that opened up the entire country to bank branches, and is probably still reasonable for the case of small banks (see Meyer and Yeager (2001)). For the sample used in this analysis, however, a more complete measure of banks¡¯ regions is required. For this paper, a bank¡¯s regional economy is defined as a weighted average of the states it operates in; the weight on any given state corresponds to the share of the bank¡¯s total deposits that originate from that state. The bank-specific weights are constructed based on the branch-level data available from the Summary of Deposits (SOD) data collected and maintained by the Federal Deposit Insurance Corporation (FDIC).9 The SOD database contains deposit data for more than 85,000 branches/offices of FDIC-insured institutions. SOD information is required for each insured office located in any state, the District of 10For SOD purposes, the FDIC collects deposit balances for commercial and savings banks as of June 30 each year. For insured commercial banks and FDIC-supervised savings banks, the definition of deposit is the same as in the Consolidated Report of Condition. The definition relates to domestic deposits held, or accepted, by the reporting bank in its main office and in any branch located in any State, the District of Columbia, the Commonwealth of Puerto Rico, or any U.S. territory or possession which include but are not limited to Guam and the U.S. Virgin Islands. 10 Columbia, the Commonwealth of Puerto Rico or any U.S. territory or possession such as Guam or the U.S. Virgin Islands, without regard to the location of the main office. For SOD purposes, a branch/office is any location, or facility, of a financial institution, including its main office, where deposit accounts are opened, deposits are accepted, checks paid, and loans granted. Some branches include, but are not limited to, brick and mortar locations, detached or attached drive-in facilities, seasonal offices, offices on military bases or government installations, paying/receiving stations or units, and Internet and PhoneBanking locations where a customer can open accounts, make deposits, and borrow money. This definition of a branch should very accurately gauge the cross-state activities of a bank.10 All the regional variables and the composite index are weighted by the deposit shares to define bank-specific regional economic conditions. Data are collected annually, hence the bank-specific weights are only updated once every four quarters. Banking Variables The banking data are collected from the quarterly Reports of Condition (the Call Reports) that commercial banks file with their bank regulators. The data set consists of all commercial banks with domestic charters between 1983.Q3 and 2002.Q3. As is welldocumented in the literature, changes in regulation and competitive pressures spurred a remarkable degree of consolidation in the banking industry. These changes are apparent in the 11 data set used in this paper (see Figure 5). At the beginning of the period there are 13,288 unique bank entities in the banking database; by the end of the sample period, this number had dwindled to 7,620 banks. Bank condition is measured here as the ratio of total nonperforming loans to total loans, where non-performing loans are defined as all loans past due thirty days or more but still accruing interest and non-accruing loans. Bank-level control variables include the natural log of assets to control for the many differences between large and small banks, the share of the loan portfolio assigned to commercial and industrial (C&I) lending, consumer lending, and residential and nonresidential real estate. The banking variables are summarized in Table 1. IV. Evidence of a Relationship Before specifying a model of the relationship between bank conditions and regional economic performance it is useful to pin down their association using simple Granger Causality tests. Specifically, for each state we test whether annual growth in the coincident indicator Granger-causes the state-level of nonperforming loans ratio, and vice versa. Aggregating across states, we also test whether the variance in states¡¯ annual growth rates of the coincident indicators Granger-causes the variance across states non-performing loan ratios. We use 8 lags of both the dependent and independent variables in the tests. The results of these tests are reported in Table 2. The results provide mixed evidence of the importance of regional indicators in predicting bank condition at the state level, with growth in the coincident indicator Grangercausing the nonperforming loan ratio in 12 states. Importantly, there is less evidence that the 11An alternative specification would be to do the estimation in two stages, in which the first stage regresses bank nonperforming loans on bank-specific variables and 50 state dummies for each quarter, and the second stage regresses the state dummies on regional economic indicators (Card and Krueger 1992; Hanushek, Rivkin, and Taylor 1996). The model in equation (1) was chosen to be consistent with the previous literature and to permit interactions between bank-specific variables and regional indicators. 12 relationship works the other way; nonperforming loan ratios Granger-caused growth in just 5 states. A much clearer picture emerges for the relationship between the cross-state variance in growth in the coincident indicator and the non-performing loan ratio (final row of Table 2). The variance in annual growth of the coincident indicator across states Granger-causes the variance in non-performing loan ratios across states, but the reverse does not hold. There is no evidence of feedback from nonperforming loans to regional economic growth. These results suggest a role for regional indicators in models of bank condition, especially in modeling differences across states over time. Based on these findings we turn to estimating a reduced form model of bank performance that includes bank specific variables and regional economic conditions. V. In-Sample Importance of Regional Variables The Model By and large, the literature on the impact of regional economic performance on bank conditions relies on a basic model specification that regresses individual bank condition variable on a set of bank-specific variables and regional indicators. This type of model is followed in this paper.11 The estimated model takes the following form: (1) 4 4 , ijt it jt ijt y x z ¦Á ¦Â ¦È ¦Å − − = + + + 13 where yijt is the nonperforming loan ratio for bank I operating in region j at time t, ¦Ái is the intercept term, xi is a vector of bank-specific variables, and zj is a vector of region-specific variables, including the composite index. The explanatory variables are lagged four quarters under the assumption that changes in bank characteristics and regional economic condition will take several quarters to appear in the performance or asset quality variables. All regressions are estimated with state dummies and robust standard errors to account for the non-independence of multiple observations on the same bank over time. The regression results are reported in Tables 3-5. Note that the five specifications of the model used throughout the paper are: (I) bank-specific variables only; (ii) bank-specific variables plus regional variables excluding the composite index; (iii) bank-specific variables and all regional variables; (iv) bank-specific variables and just the composite index; and (v) all the prior variables plus the composite index interacted with the bank¡¯s portfolio shares. All Banks The results for all banks are reported in Table 3. The findings point to a strong impact associated with the inclusion of the lagged value of the nonperforming loan ratio in all of the specifications. Invariably, the coefficient on the lagged dependent variable is approximately 0.66 and is statistically significant. In addition, including this variable in the specification increases the regression R2 from approximately 4% to 50% in most of our regressions. Also, consistent with many other studies in banking, the size control is statistically significant. Larger banks tend to have lower nonperforming loan ratios than smaller banks. The inclusion of economic variables yields somewhat promising results (column 2 of Table 3). However, the counterintuitive sign on personal income growth is an example of 14 some of the difficulties researchers have had in pinning down the influence of regional economic conditions on bank condition. We would expect positive changes in this variable to be associated with decreasing problem loan ratios, all else being held equal. Overall, the coefficients on the change in house prices and the employment variables all have the predicted signs and are statistically significant. In column 3 of Table 3, the composite index is added to the model along with the other regional variables. All coefficients on the observable economic variables remain statistically significant, although their magnitudes are somewhat diminished when the composite index is included. This is particularly true for the employment growth variable, which declines in importance by an order of magnitude. The coefficient on the year-over-year change in the composite index, meanwhile, is estimated to be -0.046, which has the largest magnitude of any of the coefficients on economic variables. A more parsimonious specification that drops all the observable economic series and includes only the control variables and the composite index is shown in column 4 of the table. Again, the estimated coefficient on the composite index is sizeable (-0.044) and statistically significant at the conventional levels. The final column of Table 3 shows a specification of the model that includes interactions with the bank¡¯s loan portfolio. The interaction terms are meant to examine whether banks with particular loan concentrations are more/less susceptible to regional shocks. The findings indicate that the largest impact of the composite index is for banks with large concentrations of non-residential real estate and C&I loans. This result seems consistent with both the greater volatility in the commercial sectors and the ability of banks to more 12The Table 3 analysis was repeated including bank fixed effects. Except the growth in personal income, all coefficients on the economic variables have the same expected signs and are statistically significant at conventional levels. Results are available from the authors upon request. 15 easily diversify their risk on residential and consumer loans.12 Intrastate and Interstate Banks Given the national deregulation of interstate banking that occurred during the 1990s, we chose to extend the analysis by examining inter- and intrastate banks separately. The results are presented in Tables 4 and 5. The separate analysis focuses on whether banks that eventually become interstate banks differ from those that remain intrastate banks. The composite index as the sole economic variable, the coefficient on the composite index is virtually the same across interstate and intrastate banks. That said, there is an interesting difference between the intrastate and interstate regressions. The interactions between loan shares and the composite index are insignificant in the interstate regressions. This contrasts with the intrastate regressions where the effect of the composite index on nonperforming loans was significantly greater for banks with relatively large commercial real estate and commercial lending portfolios. This finding suggests that the geographic diversification by interstate banks may translate into diversification of risk on all loans, not just residential and consumer. In-Sample Results Summary The results from our reduced form models point to a clear link between regional indicators and bank performance. The linkage appears most robust for the composite index measure of regional economic performance. This is consistent with other work showing positive and consistent results for GSP variables and more mixed results for other regional 16 indicators. The results from the inter- and intrastate banks suggest that regional conditions are important explanatory factors for both types of entities, although interstate banks appear more able to diversify away portfolio risk more than intrastate banks. VI. Out-of-Sample Importance of Regional Variables The in-sample results point to the usefulness of regional indicators for explaining bank condition. A next logical step is to see whether such regional indicators could help in forecasting bank conditions out-of-sample. Previous attempts to forecast bank performance using regional economic variables have been unsuccessful, as discussed in the literature review. Similar calculations based on our dataset also provide little evidence that including regional variables improves the forecasts of individual bank condition. However, this outcome need not imply that regional variables cannot be used to understand broader trends in banking sector conditions. To examine this alternative perspective, we focus on forecasting the relative rankings of bank risk, measured as nonperforming loan ratios, by state. Such rankings are useful to both bankers, who are potentially managing loans to borrowers across the country, and to bank supervisors, who monitor the condition of banks nationwide. Although these rankings abstract from the absolute level of bank risk, they retain a useful amount of relative risk information at a point in time and across time. Our forecasting exercise proceeds as follows. First, we modify the modeling framework from the earlier part of the paper and estimate a linear probability model of 13The new dependent variable is a binary variable equal to one if the total nonperforming loan ratio is greater than 5.4%. This threshold corresponds to the 80th percentile of the empirical distribution of nonperforming loans in our entire data. sample. 14Note that is neither an indicator variable, like our dependent variable, nor is it a true ijt 1 y + probability, since its support is not limited to the unit interval. Instead, is a relative value that ijt 1 y + indicates proximity to the 5.4% nonperforming loan ratio that we selected; that is, higher values of ijt 1 y + indicate that bank i is more likely to be above the threshold, while lower values of indicate that the ijt 1 y + bank is more likely to be below the threshold. 17 whether or not a bank¡¯s problem loan category exceeds a predetermined level¡ª5.4%.13 Second, we break the full sample into three subperiods and estimate the models for each subperiod absent the last year. We use the fitted models for each period to forecast the value of the linear probability model for each bank in each period¡¯s out-of-sample period; that is, the four quarters of 1989, the four quarters of 1995, and the three quarters of 2002, respectively. We denote the forecasted value for bank I headquartered in state j at time t+1, conditional on the information available at time t, as 14 Since we are estimating a linear probability ijt 1 y . + model, the actual numerical values of are not especially meaningful to the analysis. ijt 1 y + However, the cardinal ordering of these numerical values allows us to generate state-level risk rankings based on nonperforming loan ratios. To generate state-level ratios of nonperforming loans to total loans, we aggregate across banks in each state at each point in our forecasting periods. That is, the nonperforming loan ratio for state j at time t+1, denoted as is calculated as jt 1, NPL + ijt 1 jt 1 ijt 1 i j jt 1 L NPL y , L + + + ¡Ê + = ¡Æ where Lijt+1 is the total loans for bank I headquartered in state j at time t+1 and Ljt+1 is the total 18 loans across all banks headquartered in state j at time t+1. We construct the actual state problem loan ratios, denoted as NPLjt+1 using the same formula as above, but replacing the values with the actually observed nonperforming loan ratios for the individual banks. ijt 1 y + We generate four sets of forecasted state rankings based on the variables specification used in Section V. We assess the accuracy of these forecasted state rankings by examining the extent to which they reproduce the actual rankings of state-level nonperforming loan ratios. The degree of forecast accuracy is measured using the Spearman rank correlation test, which tests the null hypothesis that the actual and the forecasted rankings are independent. If we reject the null hypothesis, then the model generating the forecasts is accurately characterizing the underlying economic relationships. The values of these correlation coefficients are presented in Table 6; note that they are all significant at the 5% level. To consider the importance of the regional indicators, we look to see whether the forecasted rankings are different, depending on the inclusion of the composite index in the forecasting model. The findings show that regional variables do improve forecasts of statelevel bank risk, i.e., the coefficients in the fourth column are larger than those in the first column for all 11 out-of-sample forecast quarters. However, the results also highlight the potential difficulties of using a collection of individual regional indicators¡ªemployment growth, personal income growth, and home price appreciation. As shown in the second column, in some of the time periods and relative to the baseline model in the first column, the inclusion of these individual regional indicators hinder rather than help the forecast. This is where the composite index clearly dominates the other regional variables in the analysis. By collecting the trend that is common to all included regional indicators it improves the signal to 15We do not show results from these regressions, but they are available from the authors upon request. 16Another possibility is the relatively short sample period. However, regressions for sample periods of equal length during the 1980s and early 1990s produce statistically significant coefficients on 19 noise ratio and aids the forecast accuracy of the state rankings, as shown in the third column. In all the time periods forecast, the composite indicator either improves or fails to change the forecast with the bank variables alone. VI. Regional Economic Conditions Under Interstate Banking Our findings point to a clear link between regional economic conditions and bank condition for both intrastate and interstate banks. That link, however, largely results from variance in state regional growth and loan performance that predates the full expansion of interstate banking; data from the 1980s and early 1990s. This has led some to argue that these linkages may have dissolved. To get at this issue, we ran regressions based on equation (1) on a subsample of the data ranging from 1995.Q1 to 2002.Q3 for both interstate and intrastate banks.15 The idea was to examine whether the economic and/or statistical significance of the regional indicators has dissipated for the interstate banks, while remaining important to intrastate banks. The results show that during the latter half of the 1990s none of the regional economic variables (including the composite index) had a statistically significant effect on either intrastate or interstate bank nonperforming loan ratios. The reason for this non-result can be seen in Figure 7, which shows the variance in regional economic growth (measured by the composite index) and state nonperforming loan ratios.16 As the figure indicates, during the latter half of the 1990s, the variance of these two the regional variables, suggesting the short sample is not driving our 1995-2002 findings. 20 series dropped dramatically and stayed virtually unchanged for the remainder of the decade. The lack of movement in either series makes hinders our ability to track their underlying relationship. What can we expect going forward? Without additional data we cannot know whether the 1995-2002 period was an aberration or per the discussion in Stock and Watson (2002) a permanent change in the relationship between bank condition and regional economic performance. Given very recent data, the convergence of regional growth that characterized the late 1990s already is fading as states recover from the recession at varying speeds. Moreover, the variance of the nonperforming loan values has begun to increase, suggesting that its period of stability also may be nearing an end. Finally, earlier results from our Granger causality tests provide little support for the idea that the convergence of banking outcomes portends convergence in regional economic growth. VII. Conclusion In summary, our work points to a clear link between bank conditions, measured by nonperforming loan ratios, and state-level composite indexes and other regional variables. The statistical significance and economic relevance of the composite index remains stable across bank sub-samples (interstate and intrastate). In general, the composite index performs at least as well as employment growth and outperforms other measures of regional conditions. Moreover, it appears to capture important and tractable interactions among the included regional variables that would otherwise be relegated to the state fixed effect or the residual. 21 We find that for interstate banks, the composite index is the singl...