《国际清算银行-人口老龄化与银行风险承担（英）-2022.11-53正式版.docx》由会员分享，可在线阅读，更多相关《国际清算银行-人口老龄化与银行风险承担（英）-2022.11-53正式版.docx（57页珍藏版）》请在taowenge.com淘文阁网|工程机械CAD图纸|机械工程制图|CAD装配图下载|SolidWorks_CaTia_CAD_UG_PROE_设计图分享下载上搜索。
1、OBISBIS Working PapersNo 1050Population aging and bank risk-takingby Sebastian Doerr, Gazi Kaba and Steven OngenaMonetary and Economic DepartmentNovember 2022JEL classification: E51, G21.Keywords: Risk-taking, financial stability, low interest rates, population aging, demographics.the respective cha
2、nnels through which population aging a ects bank risk-taking di cult. The absence of major nancial regulatory changes during our sample period makes it well-suited to identify the e ects of population aging on bank risk-taking. Second, we can exploit the Great Recession to analyze whether higher ris
3、k-taking in the years leading up to the crisis manifested itself in higher nonperforming loans during the shock episode. That is, we can analyze whether laxer lending standards had an impact on nancial stability. And third, we avoid the zero lower bound on interest rates (Leahy et al., 2022).2.1 Mai
4、n variablesPopulation aging. Our main explanatory variable at the county level is the change in the log of the population of ages 65 and above from 1997 to 2007, denoted by oldc. We use the change in the number of seniors (i.e., in the level) rather than the change in the ratio of seniors to the tot
5、al population (i.e., in the share), because changes in the share could be driven by changes in the numerator or denominator. For example, a decline in the total population of a county would lead to an increase in the share of seniors, even though the number of seniors does not change. In such a case
6、, the relationship between e.g. deposits or loans and the change in the share of seniors would be driven by the decline in population.7 Detailed population data by age cohort are provided by the National Cancer Institute Surveillance, Epidemiology, and End Results (SEER) program. We use these data a
7、lso to construct changes in the size of other age cohorts.Bank exposure. To calculate banks* exposure to aging counties, we use data from the Federal Deposit Insurance Corporations (FDIC) Summary of Deposits (SOD), which provides information on the geographic distribution of bank deposits.8 We compu
8、te7ln regressions of the log of the share of seniors on the log of the total number of seniors (the numerator) and the log of the total population (the denominator), a Shapley decomposition of the R-squared shows that around two-thirds of the variation in the share is explained by the variation in t
9、he total population, and around one-third by the variation in the senior population. Using the share of seniors as an explanatory variable would hence lead to signi cant measurement issues, as most of the variation would be due to changes in the total population. Instead, we directly control for pop
10、ulationgrowth in our regressions.8An assumption underlying our analysis is that deposits raised by banks in a given county are raised from residents of that county. Amel et al. (2008) show that between 1992 and 2004, the median distance between a depositor and its bank was three miles and remained c
11、onstant over time. Nonetheless, the rise of online banking could mean that branchless deposits are recorded in the banks headquarters branch. To the extent that the amount of branchless deposits is uncorrelated with bank exposure, such measurement error would arguably not lead to a bias in the estim
12、ates. In line with this argument, we nd that excluding banks headquarter county from the analysis does not a ect our results (unreported).banks beginning-of-sample exposure asdepositsb；cX exposureb = c depositsb oldc；(1)where depositsb；c and depositsb denote bank bs deposits in county c and its tota
13、l deposits as of 1997. oldc is county cs change in the log of the population of ages 65 and above. High exposureb implies that a large share of banks, initial deposits is held in aging counties, while low exposure implies that deposits are held in counties with a small increase in the number of seni
14、ors. Higher exposure thus corresponds to an increase in deposit-weighted average aging in banks* borrower counties. Exposure is constructed from beginning-of-sample deposit shares, alleviating concerns about banks selectively opening branches in aging counties.9Instrumental variable strategy. Counti
15、es that experience a stronger increase in se-niors could di er along other dimensions that could matter for deposit growth. We thus predict old, i.e., the change in the population of ages 65 and above from 1997 to 2007, with the change in the population of age 45 to 65 from 1977 to 1987 in the same
16、county. In essence, we use the predetermined component of each countys age structure 20 years prior to our sample period as an instrumental variable for the actual change in the age structure. This approach builds on the assumption that the historical demographic structure is plausibly exogenous to
17、changes in contemporaneous confounding factors. For example, it purges old from changes in life expectancy, in- or out-migration, or incomes over the sample period.Similar to county aging, bank exposure could be correlated with other (unobservable) factors that a ect bank behaviour. For example, agi
18、ng counties could, for whatever reason, face better economic prospects in terms of their unemployment trajectory or income growth. Banks more exposed to aging counties would then also be more exposed to faster growing areas, which poses a threat for our identi cation strategy. To address this concer
19、n beyond the xed e ects strategy (explained below), we construct exposure in Equation (1) based on old predicted by the historical demographic structure. In addition, and to test whether exposure to aging counties is correlated with exposure to counties with better growth trajectories, we compute ex
20、posure to the changes in localQIn the Online Appendix, we show that neither local aging nor bank exposure predict in which counties banks open new branches prior to 1997, see Table OA3.unemployment rates and income per capita, analogous to Equation (1). We obtain a strong rst stage, with an F-statis
21、tic of almost 60 (see also Figure OA2).LTI ratios, loan and deposit growth. Home Mortgage Disclosure Act data provide detailed information on banks* residential mortgage lending.10 HMDA covers the vast majority of applications and approved mortgage loans in the U.S. The data include the application
22、outcome (granted or denied), loan amount, and borrower income for each loan.11 We measure bank risk-taking through the loan-to-income (LTI) ratio, de ned as loan volume over applicant income. The LTI ratio is a signi cant predictor of ex-post default (Fuster et al., 2021) and has been widely used in
23、 the literature to measure the riskiness of loans (DellAriccia and Marquez, 2006; DellAriccia et al., 2012; Duchin and Sosyura, 2014). We compute the change in the average LTI ratio, as well as at the 10th, 25th, 50th, 75th, and 90th percentile in each bank-county cell:LT I07 LT I97 b;cb;cLT lb；C =
24、(LT I07 + LT l97)=2 ：(2)b;c b;cAdditionally, we compute the change in the share of denied applications in each bank-county cell. For these measures of risk-taking, we are restricted to the Intensive margin1, in the sense that we can only take into account counties in which banks made loans in 1997 a
25、nd 2007.Further, we compute the change in loan amounts and deposits at the bankcounty level as y07y97_ b：c b;c yb；C =(yU，+产)=2；(o)b;c b;cwhere y is either HMDA loans or deposits. For our mortgage analysis, we use mortgage loans that were not sold in the respective calendar year. Since these loans ar
26、e mostly retained on banks* balance sheets, they are predominately funded by deposits (Han et aL, 2015; Cortes and Strahan, 2017). To account for entry into and exit out of counties over the long time horizon, we standardize the change in variables by their respective mid-10We follow the literature
27、and restrict the sample to conventional or Federal Housing Administration (FHA)-insured loans, exclude multi-family properties, and keep only originated, approved, and purchased loans. We also drop all observations with missing county Federal Information Processing Standards (FIPS) codes or missing
28、borrower income, as well as loans extended to borrowers residing outside of metropolitan statistical areas (MSAs).11 一In 2007, mortgage lending averaged around 30% of banks total lending, and 40% for the largest banks. Additionally, HMDA data represent the most detailed publicly available data on ba
29、nk lending disaggregated at the geographical level, which is why we focus on mortgage lending in our analysis.points. This de nition bounds growth rates to lie in the interval 2; 2, where 2 implies that a bank exited a county between 1997 and 2007, and 2 that it entered.12 Finally, we de ne the dumm
30、y no branch that takes on a value of one if bank b had no branch in county c in 1997 and zero otherwise.Bank and county data. The FDIC provides detailed bank balance sheet data in its Statistics on Depository Institutions (SDI). We collect 1997 second quarter data on banks* total assets, Tier 1 capi
31、tal ratio, nonperforming loans (NPL), return on assets, total deposits, total liabilities, share of non-interest out of total income, and overhead costs (e ciency ratio). We also include an indicator of an institutions primary specialization in terms of asset concentration that takes on ten distinct
32、 values.13 We also collect 1997 and 2007 data on banks total deposits, total liabilities, and total loans and compute the change in the logarithm of each variable.At the county level, we further collect public information on debt-to-income ratios (Federal Reserve Bank of New York Consumer Credit Pan
33、el, available from 1999 onward). We also collect 1997 data on the log of the population (NCI SEER), the unemployment rate (Bureau of Labor Statistics, Local Area Unemployment Statistics (BLS LAUS), log income per capita (Bureau of Economic Analysis, Local Area Personal Income (BEA LAPI), house price
34、 indices (Federal Housing Finance Agency (FHFA), as well as employment shares in manufacturing (SIC code 20), retail trade (SIC code 52), and services (SIC code 70), provided in the County Business Patterns (CBP). CBP also provide information on employment in tradable and nontradable industries. Fol
35、lowing Adelino et al. (2015), we classify two-digit NAICS code 23 as construction; codes 44, 45, and 72 as nontradable, and all others as tradable industries.To remove outliers, we winsorize all variables at the 0.5th and 99.5th percentile. We then trim all remaining extreme values that lie at least
36、 ve standard deviations above or below the mean.1 2While the (og di erence is symmetric around zero, it is unbounded above and below, and does not easily a ord an integrated treatment of entry and exit. The growth rale used in this paper is divided by the simple average in t j and t It is symmetric
37、around zero, lies in the closed interval -2,2,Jacilitates an integrated treatment of entry and exit, and is identical to the log di erence up to a second order Taylor scries expansion (Davis and Halhwanger, 1999).13Tlic indicator recds.Jbr example, whether a bank has an agricultural or a mortgage le
38、nding specialization.Survey of Consumer Finances. The Survey of Consumer Finances (SCF) of the Federal Reserve provides detailed infoimalion on ihe allocation of households1 nancial assets. The SCF is a triennial cross-sectional survey on household assets and demograpli-ics. We combine the survey wa
39、vesfom 1998 and 2007 (41,366 observations). We collect iriformation on respondents total nancial assets, deposits, and debt, as well as dummy indicators Jor whether they borrowed in the past year, had any outstanding debt, or whether they have been turned down for credit, or feared being denied cred
40、it in the past 5 years. As control variables, we Ju r( her use data on the education level, number of kids, gender, race, marital status, home ownership, and a dummyJor business ownership.1 .2 Descriptive statistics4In the average county, the number of seniors ( Oldc) increased by 12%, with a standa
41、rd deviation of 15%. These numbers suggest signi cant variation in population aging across U.S. counties.Table 1 provides descriptive statisticsJbr the main variables. In total, our sample includes 1,843 banksjbr which we have dala on loans and deposits over (he sample period. Panel (a) summarizes b
42、ank exposure and other balance sheet charactcristicsjrom the SDL For the average bank, exposure equals 0.10, with a standard deviation of 0.11. As exposure re ects deposit-weighted aging, a mean of 0.10 implies that the number of seniors increased by 10% in the average county 15 where a bank takes d
43、eposits.Panel (b) reports summaiy statistics at the bank-county level. The average bank saw tittle change in its LTl ratio, but (here is sizeable dispersion in (he change across banks, (he standard deviation equals 85 pp. While LTl ratios at the 25th percentile declined by 19 pp, LTl ratios increase
44、d by more than 5 pp on average at the 75th percentile, and by more than 20 pp at the 90th percentile. The share of denied loans increased by 7 pp on average. For the average bank-county pair, deposits increased by 114% over the time period (13,086 observations); mortgage tending (one- to Jour-Jamily
45、 residences) 16 increased by 103% (53,197 observations). Along the intensive margin (disregarding bank entry and exit across counties), 17 lending increased by 84% (17,643 observations).14 rhe ims疝gh虱 county average of n.6% in the dunge in the number ofseniors compares to a US-uxde increase bi the s
46、enior population of arouttd 10% over the sample period.15 Figure OAi $A(nvs the distribution of aging and cxposuir to aging.16 The di erence in the number of observations reects that the average bank lends to counties in which it does not raise deposits.17The unweiglited average enchase in mortgage
47、lending in our sample k louvr (han thegutr basedFinally, panel (c) repoj-ts summary statisticsJbr county-level variablesJbr the set 0/ 2,163 counties.BalanCedneSS. To examine the balancedness in beginning-of-sample covariatcs among our sample of banks, in Tabic 2 we split banks into those below (low
48、 exposure) and above (high exposure) the median of the exposure distribution. High-exposure banks are slightly smaller (the di erence in size is signi cant at the 10% level). They are statistically similar in terms of the share o/nonperfo-ming loans, return on assets, capital ratio, ratio of deposit
49、s to liabilities, or e ciency ratio (re18，ecting overhead costs). There is also no signi cant di crcnce in the share of C&l loans or the share of loans extended to no-branch counties.A potential concernJbr identi cation is that banks strategically opened branches to bene tjromJiilure deposit in ows in anticipation of demographic (rends. In the Online Appendix, we show that neither local aging nor bank exposure predict in which counties banks open new branches prior to