Bank leverage shocks and the macroeconomy: a new look in a data-rich environment

Transcription

1 Bank leverage shocks and the macroeconomy: a new look in a data-rich environment Jean-Stéphane Mesonnier and Dalibor Stevanovic Banque de France EUI and UQAM October

2 Motivation and overview Literature Overview of modeling strategy Macroeconomic shocks Macroeconomic data set 3 Bank-level CAR shocks Dynamic model of bank capital Database of large US banks Panel regression results Aggregation 4 IRFs using FADL FADL Results Overview Checking the factor structure IRFs Robustness Comparison with VAR

3 Motivation Overstretched leverage of large US nancial institutions a major culprit of accumulation of risks leading to the subprime crisis As result, Basel 3 package includes a substantial increase in quantity and quality of core capital relative to risk-weighted assets: potential large shock on bank leverage Empirical question: what are the macroeconomic consequences of shocks reducing bank leverage? Vivid debate: IIF () vs Admati et al. () Macroeconomic Assessment Group (Basel): negative e ects in the short-run (transition) but positive in the long-run? Severe methodological shortcomings in the attempt to reconcile facts observed at the micro level and assessment of macro e ects.

4 In this paper... A new approach to assess the macroeconomic consequences of a shock to the capital-to-asset ratio (CAR) of large US bank holding companies (BHC)... relying on a rich database of both bank balance sheet informations (micro-level) and macroeconomic aggregates: Construct measures of exogenous CAR shocks at bank level, using a dynamic model of bank capital ratios Estimate the dynamic responses of a large set of macroeconomic variables to the previously aggregated bank leverage shock using the new FADL methodology from Ng and Stevanovic (). Aims to bene t from combined advantages of existing approaches, while overcoming limitations: bank-level panel regressions: potentially good identi cation of capital shock monetary VARs with bank leverage: potentially useful to catch general equilibrium e ects

5 Main results Leverage shocks matter for understanding credit aggregates variations as well as the US business cycle. A basis points unexpected rise in CAR causes: Signi cant and persistent fall in the growth of loans Total commercial bank credit contracts by some % on impact and 3% over 6 quarters, but rates on C&I loans shoot up Short-run fall of investment, consumption and GDP Asymmetric e ects of leverage reducing and leverage increasing shocks; the former matter much more. Compare our results with IRFs obtained with alternative measure of leverage shock from a small-scale macro VAR (Berrospide and Edge, ) highlights bene ts from use of micro data for identi cation

6 Related literature Identi cation of credit supply e ects in macro VARs using relevant "non-credit" banking indicators: share of assets in impaired banks (Peek, Rosengren and Tootel, 999, 3), lending standards (Lown and Morgan, 6, Ciccarelli et al.,, Basset et al., ) Papers using bank-level regressions to gauge the e ects of bank capital (more precisely bank leverage) on lending (Hancock and Wilcox, 994, Kashyap and Stein, or Berrospide and Edge, ) FAVAR models combining micro (bank-level) and macro datasets (Buch, Eckmeier et al.,, Dave, Dressler, Zhang, 9, Jimborean and Mesonnier,, )

7 Overview of modeling strategy: four steps Extraction of "structural" macroeconomic "non-bank related" shocks from a large macro dataset X using a dynamic factor model: vector of "non-bank" macro shocks bη t

8 Overview of modeling strategy: four steps Extraction of "structural" macroeconomic "non-bank related" shocks from a large macro dataset X using a dynamic factor model: vector of "non-bank" macro shocks bη t Estimation of bank-level shocks to the CAR, bε i,t, using a standard dynamic model of bank capital k i,t that includes the macro shocks bη t as regressors

9 Overview of modeling strategy: four steps Extraction of "structural" macroeconomic "non-bank related" shocks from a large macro dataset X using a dynamic factor model: vector of "non-bank" macro shocks bη t Estimation of bank-level shocks to the CAR, bε i,t, using a standard dynamic model of bank capital k i,t that includes the macro shocks bη t as regressors 3 Aggregation of the bank bε i,t into a macro measure of CAR shocks bε t

10 Overview of modeling strategy: four steps Extraction of "structural" macroeconomic "non-bank related" shocks from a large macro dataset X using a dynamic factor model: vector of "non-bank" macro shocks bη t Estimation of bank-level shocks to the CAR, bε i,t, using a standard dynamic model of bank capital k i,t that includes the macro shocks bη t as regressors 3 Aggregation of the bank bε i,t into a macro measure of CAR shocks bε t 4 Estimation of IRF of variables in main macro database X or ancillary credit-banking macro database Y using FADL methodology (Ng & Stevanovic, ).

11 Estimation of macroeconomic shocks Assume that X t has a dynamic factor representation: X t = λ(l)f t + u t () u t = D(L)u t + v Xt () f t = Γ (L)f t + Γ v ft (3) f t : q common factors that evolve as a vector autoregressive (VAR) process of order h, λ(l) is a polynomial matrix of factor loadings of order s, D(L) is a diagonal polynomial matrix, v Xt is a vector white noise process, v ft is a vector of q structural shocks such as demand, supply or monetary policy, characteristic roots of Γ (L) are strictly less than one, E (v Xit v Xjt ) =, and E (v Xit v fkt ) = for all i 6= j and for all k =,... q.

12 Estimation of macroeconomic shocks Static factor representation of x t = (I D(L)L)X t : x t = ΛF t + v Xt (4) F t = Φ F F t + ɛ Ft () where ɛ Ft = G η t, Λ be the N r matrix of loadings, F t is r = q(s + ) Step : Estimate F t by IPC as in Stock and Watson (). Step : Estimate η t = v ft : Let ˆɛ Xit = x it ˆΛ i ˆΦ F ˆF t. The estimate of η t consists of rst q principal components of ˆɛ Xit. Use information criteria from Amengual and Watson (7), and Bai and Ng (7) to estimate q.

13 Macroeconomic data A rich macroeconomic dataset for two joint purposes: to uncover real and nominal shocks spanning macro uctuations, aside from bank leverage shocks. to compute IRFs for a large set of macro variables. Two separate macroeconomic datasets: A selection of 3 real, nominal and nancial series representing the US economy, with the exclusion of any banking/credit/money indicator, stacked in X. A complementary set of 4 macro credit and banking indicators, stacked in Y.

14 Macroeconomic data 3 real, nominal and nancial series in X : close to the selection in Gilchrist et al. (9). Limited number of series in line with recommendation in Boivin and Ng (6) and Onatski (9), in order to avoid problem of weak factor structure. 4 blocks: real economic activity ( series), prices (7), term structure of interest rates (7), assets prices (NEER, S&P, AAA & BAA spreads, FHFA housing price index). 4 credit and banking indicators in Y : total US commercial bank credit and its components (H8 Release of the Fed) interest rates on loans (personal, car, C&I) (SCC and STBL) Credit standards (SLOOS) Aggregate leverage of US commercial banks (as in Berrospide and Edge, ).

15 A standard dynamic model of bank leverage targeting Following on Hancock and Wilcox (994) among others, we assume: k i,t k i,t k i,t = λ k i,t k i,t + ei,t (6) where k i,t : actual capital ratio at period t for institution i, : target capital ratio, λ drives adjustment speed. Bankers assumed to set their target in light of individual info and macro outlook: k i,t = θ Z.Z i,t + θ M.M i,t. Problem: bank-speci c innovations to leverage e i,t may not be orthogonal to macro shocks occurring between t and t, like exogenous credit demand or monetary policy shocks.

16 Extracting structural bank leverage shocks Suppose that observed macroeconomic uctuations can be subsumed to the propagation of a small number of unobserved common shocks, η t = v ft (cf. macro factor model above). Then, innovations e i,t may be partly driven by these shocks: e i,t = θ η.η t + ε i,t Replacing in equation (6), rearranging and adding bank FE, we nally get: k i,t = α i + ( λ).k i,t + λ.θ Z.Z i,t + λ.θ M.M t + θ η.η t + ε i,t (7) Now, the ε i,t can be interpreted as structural shocks to individual banks capital ratios Possible causes are manyfold (changes in regulatory environment, in bank speci c requirements, in the business model or risk strategy of the bank, windfall P&L etc.)

17 Estimation of bank capital model: regressors Usual bank-speci c determinants of target capital ratio: Size, ROA, net charge-o s, asset composition (mortgage and C&I loan shares). Measures of global risk and macroeconomic expectations: Stock market volatility (VIXX), expected changes in short term rate and GDP growth from Phil Fed SPF. 3 unobserved macro shocks extracted from "non-banking" macro dataset X. Dummies for M&As, FHC status, seasonality. Bank FE.

18 Bank balance sheet data Consolidated balance sheet information for all US Bank holding corporations (BHC) from the Federal Reserve s "Call reports". Highly unbalanced panel (89 banks present, but only 66 throughout). Period of 986 Q to Q, potentially including large shocks to US banks leverage. Focus on large BHCs (assets always above $ 3 bns). Trade-o stability vs selection bias: keep only banks present more than 8 consecutive years. We end up with an unbalanced panel of 4 banks (only present over whole period), representing 7% of total banking assets on average.

19 4 6 8 in % Nb of banks Share of our sample of banks in total US BHC assets and number of banks in sample through time 98q 99q 99q q q q

20 Bank balance sheet data: institutional and statistical issues Consolidation wave of the US nancial banking sector since the late 98s: M&A identi ed using Fed of Chicago database (36 M&A in our sample). Impact of the Gramm-Leach-Bliley Financial Modernization Act of 999: identi cation of switches from BHC to FHC status using reports from the US National Information Center of Financial institutions. Reporting break in 6 Q: large subsidiaries (assets above $ bn) do not have to ll separate reporting anymore.

21 Summary stats for bank variables N mean sd p Assets ($ mns) 6,37 76, ,79. Capital-to-assets 6, ROA 6, Loans-to-assets 6, Mortgage loans-to-assets 6, C&I loans-to-assets 6, Net chargeo s-to-assets 6,

22 Regression results MICRO EXP-SWIPC Lagged Capital ratio.933 (98.48).93 (94.7) Size.94 (3.73).99 (3.9) ROA -.6 (-.4) -.6 (-.6) Net chargeo.6 (.3).4 (.8) Real estate loans -. (-.7) -. (-.9) C&I loans -. (-.9) -.4 (-.46) FHC status. (.). (.39) Mergers. (.).6 (.4) SP volatility -.39 (-.4) Exp. change policy rate. (.7) Exp. change GDP growth -.33 (-.64) SWIPC.3 (.86) SWIPC.9 (.7) SWIPC3 -. (-.6) Observations R Robust t-stats betw. ().

23 Aggregation We aggregate individual capital ratio shocks as a weighted average over the banks present at each period: bε t = en t a i,t i=.bε i,t where en t = min(n t, N t ) and a i,t is the share of bank i in the sample at t. See Figure below. Estimate of bε t robust to smaller sample of banks (- biggest) How to interpret a positive shock ε t? "Good" vs "bad" deleveraging, depending on the signs of K and A Hints from correlations at bank level, see Table 4

24 in percent Estimated aggregate bank capital ratio shocks 98q 99q 99q q q q

25 percent Estimated aggregate bank capital ratio shocks, di erent sample of banks (all 4 vs - largest) 98q 99q 99q q q q date

26 Capital ratio shocks and bank assets growth at bank level (Assets i,t ) < > ε i,t < (Equity i,t ) < ε i,t (Equity i,t ) < Table: Breakdown of individual capital ratio shocks according to the sign of contemporaneous changes in bank equity and total assets (in percent of total number of individual shocks).

27 FADL model and Impulse response analysis Main idea of FADL methodology is to augment an autoregression of a variable of interest, y t, with current and lagged values of ˆη t and ˆε t : y t = α y (L)y t + α η (L) ˆη t + α ε (L)ˆε t + v yt. (8) If y t X t, its FADL representation is derived from dynamic factor model. If the leverage shock is important for (idiosyncratic part of) y t, the corresponding coe cients should be signi cant. To construct the impulse responses, estimate (8) by OLS. The impulse response function of y t to a unit increase in ˆε t is de ned by ˆψ ε y (L) = ˆα ε (L) ˆα y (L)L.

28 Advantages of FADL methodology Identi cation of the shock of interest is done here at micro-level and hence separated from the estimation of common shocks. y t does not need to be in X t. Just test ex post whether it indeed has a factor structure. Restrictions on the responses of variables in X t or Y t, if any, are imposed equation by equation. In principle, any linear regression restriction can be imposed to shape the impulse response functions of interest. For example, to impose that y t does not react on impact to ˆε t, it is su cient to restrict α ε () =.

29 Overview of results Factor structure of data broadly validated Recessionary impact of a positive capital ratio shock Outstanding credit contracts while (C&I) loan interest rates rise: avour of a supply shock Assymetric e ects of shocks reducing vs augmenting bank leverage Robustness checks: lag selection in FADL (baseline: [,,4]) exclusion of subprime crisis period

30 Testing the factor structure of selected macroeconomic series F-test p-value R-squared REAL ECONOMY MEASURES α(l) α ε (L) Total Marginal Capacity utilization GDP Investment Industrial production Consumption CPI Housing starts House prices FFR B-spread Table: Note: The F-test test the null that the coe cients in the FADL regression of a given macro series are jointly zero (col. ). In the second column, we test the null that only coe cients of the bank leverage shocks are

31 Testing the factor structure of individual macroeconomic series. F-test p-value R-squared CREDIT MEASURES α(l) α ε (L) Total Marginal Commercial and industrial loans Consumer loans Bank credit Deposits RE loans Spread All CIL Spread Large CIL Spread Small CIL Standard Equity / Assets Table: Note: The F-test test the null that the coe cients in the FADL regression of a given macro series are jointly zero (col. ). In the second column, we test the null that only coe cients of the bank leverage shocks are

32 IRFs of macro series ( X t ) to a shock reducing bank leverage 3 Capital Utilization GDP Investment Industrial P roduction PMI Consumption Cons: durables Cons: nondurables UR CPI.3 Housing 4 House prices FFR. T Bill Y. B spread....

33 IRFs of credit series (not in X t ) to a shock reducing bank leverage Commercial and Industrial loans Consumer loans Bank credit Deposits Loans& Leases in bank credit RE loans Total A ssets Loans Banks + Thrift Spread Personal Loans Spread Car Loans.4 S pread All CIL.4 Spread Large CIL Spread Small CIL S tandard. Equity/Asset

34 IRFs of macro variables to a shock reducing bank leverage: various speci cation of FADL. Capital Utilization GDP 4 Investment Industrial Production 3 P MI. E mployment Consumption Cons: durables. Cons: nondurables CPI. Housing House prices FFR T Bill Y. B spread

35 IRFs of credit variables to a shock reducing bank leverage: various speci cation of FADL Commercial and Industrial loans Consumer loans Bank credit Deposits Loans& Leases in bank credit RE loans Total Assets Loans Banks + Thrift Spread Personal Loans Spread Car Loans. Spread All CIL. Spread Large CIL Spread Small CIL S tandard E quity/a sset

36 IRFs of macro series to a shock reducing bank leverage: with data up to 8 Q 4 Capital Utilization GDP Investment Industrial Production P MI E mployment Consumption Cons: durables Cons: nondurables CPI.4 Housing 6 House prices FFR. T Bill Y.4 B spread

37 IRFs of credit series to a shock reducing bank leverage: with data up to 8 Q Commercial and Industrial loans Consumer loans Bank credit Deposits Loans& Leases in bank credit RE loans Total A ssets Loans Banks + Thrift. Spread Personal Loans Spread Car Loans.6 S pread All CIL.6 Spread Large CIL Spread Small CIL S tandard. Equity/Asset

38 IRFs of macro series to an asymmetric shock reducing bank leverage 6 Capital Utilization GDP Investment Industrial Production P MI 3 E mployment Consumption Cons: durables Cons: nondurables CPI.4 Housing House prices FFR T Bill Y B spread

39 IRFs of credit series to an asymmetric shock reducing bank leverage Commercial and Industrial loans Consumer loans Bank credit Deposits Loans& Leases in bank credit RE loans Total A ssets Loans Banks + Thrift Spread Personal Loans Spread Car Loans.8 S pread All CIL Spread Large CIL Spread Small CIL 4 S tandard Equity/Asset

40 IRFs of macro series to an asymmetric shock augmenting bank leverage 3 Capital Utilization GDP Investment Industrial Production P MI E mployment Consumption Cons: durables Cons: nondurables CPI Housing House prices FFR T Bill Y B spread

41 IRFs of credit series to an asymmetric shock augmenting bank leverage Commercial and Industrial loans Consumer loans Bank credit Deposits Loans& Leases in bank credit RE loans Total A ssets Loans Banks + Thrift Spread Personal Loans Spread Car Loans S pread All CIL. Spread Large CIL Spread Small CIL 4 S tandard Equity/Asset

42 4 in percent MS C/A shock vs shocks from a small-scale macro VAR (BE, ) 98q 99q 99q q q q

43 Comparison to shocks on aggregate bank capital ratio (BE) and credit standards (BCDZ).3 Capital Utilization GDP Investment Industrial P roduction.4 P MI. E mployment Consumption.4 Cons: durables Cons: nondurables CPI.4 Housing.6 House prices FFR T B ill Y. B spread MS BE

44 Comparison to shocks on aggregate bank capital ratio (BE) and credit standards (BCDZ) Commercial and Industrial loans Consumer loans. B ank credit x 3 Deposits x Loans& 3 Leases in bank credit x 3 RE loans x 3 Total A ssets x 3 Loans B anks + Thrift. S pread P ersonal Loans S pread Car Loans.4 S pread A ll CIL.4 S pread Large CIL S pread S mall CIL S tandard. E quity/a sset MS BE

45 Conclusion We propose a new integrated two-step framework to identify economy-wide bank leverage shocks and their consequences, relying on a rich database that encompasses both bank balance sheet information (micro-level) and macroeconomic aggregates. We nd that leverage shocks matter for understanding uctuations in credit aggregates as well as the US business cycle. Policy implications: Does not support necessarily claim that bank capital requirements should not be raised (Basel III): motives of the CAR shock unknown here But gradual adjustment through retained earnings should be preferred.