Do credit market shocks drive output fluctuations? Evidence from corporate spreads and defaults Roland Meeks Risk Assessment Division Financial Stability Directorate Bank of England1

NBB, Brussels 25th October 2010

1

The views expressed herein are not necessarily those of the Bank of England, the FPC or the MPC.

Motivation

Are there credit shocks? Not obvious that there are: Perhaps what look like credit shocks are really something else: there is an identification problem to solve Not obvious where we should look as credit market instruments vary in the degree of maturity/credit/liquidity transformation they embody: there is a measurement issue to confront

Motivation

Are there credit shocks? Not obvious that there are: Determining the impact of credit shocks is hampered by: Simultaneity, as the price/quantity of credit responds endogenously to the state of the economy ... ... and to expectations about the future state of the economy and lack of understanding about what these shocks are ... ... commonly thought to be liquidity or risk related; but these mechanisms have been absent from typical macro models (although recent work makes encouraging progress)

Some context

The pre-crisis literature was ambivalent about the relevance of credit shocks... Bernanke and Gertler (JEP, 1995) ‘Except in rare circumstances ... such as a spontaneous financial crisis whose effects spill over onto the real economy ... credit is not a primitive driving force [of macroeconomic fluctuations]. This scenario may be relevant to the Great Depression, for example, but is not relevant under “normal” conditions.’ Cochrane (Brookings, 1994) ‘[C]redit shocks do not seem to explain a large part of postwar U.S. output fluctuations ... Credit shocks may have been important in prewar recessions accompanied by banking panics’

Some context

More recent papers have produced surprising (in the light of previous work) results: Gilchrist, Yankov and Zakrajšek (JME, 2009) ‘[S]hocks to corporate bond spreads account, on average, for more than 30% of the variation in the growth of industrial production’

Motivation Why focus on corporate bonds? This paper focuses on long-term funding market for below investment grade (BB+ or lower rated) corporate borrowers Corporate bonds are a standard credit instrument, widely used, and with a long history Auction credit, not intermediated credit, so prices (and defaults) are observable Typical sub-IG firm is highly levered, and so interest sensitive Real effects will be evident if financial frictions matter for firms Shocks in secondary markets for long-term debt likely to have delayed effects

Contributions What’s new in the paper? Build a joint statistical model of macro and credit variables, that allows for feedbacks between real and financial sectors Show how to motivate minimal identifying restrictions using economic assumptions, and so how to disentangle macro and credit shocks Demonstrate that credit shocks have made a non-trivial contribution to post-82 U.S. recessions, but didn’t cause a large proportion of the average variation in output Relate estimated credit shocks to monetary policy, liquidity and risk Compare results to conventional recursive identification schemes, showing that they are crucially misspecified ...all to be covered in this presentation.

Non-Contributions

Issues not covered Don’t try to identify the effects of de-regulation e.g. removal of interest rate ceilings (Mertens, JME 2008), relaxation of inter-state banking rules (Strahan, St. Louis Fed 2003), legislative changes (Benk, Gillman and Kejak, RED 2005) etc. Don’t impose restrictions implying absence of arbitrage, as in the ‘macro-finance’ or finance literatures (Duffie and Singleton, 1999)

High yield spreads move closely with default rates 20%

15%

10%

5%

0% Nov-82 Mar-85 Jul-87 Nov-89 Mar-92 Jul-94 Nov-96 Mar-99 Jul-01 Nov-03 Mar-06 Jul-08 NBER Recession

High yield default rate

High yield bond spread

High yield spreads move inversely with output 15%

10%

5%

0%

-5% 5%

-10% 10%

15% -15% Nov-82 Mar-85 Jul-87 Nov-89 Mar-92 Jul-94 Nov-96 Mar-99 Jul-01 Nov-03 Mar-06 Jul-08 NBER Recession

High yield default rate

Industrial Production (Ann % Chg)

The price of a risky claim to $1 Risk-free zero coupon bond Time to maturity n periods. Denote price Pnt and yield to maturity Ynt . Define continuously compounded yield to maturity ynt := log[1 + Ynt ], and thus ynt = −n−1 log[Pnt ]. Risky (defaultable) zero coupon bond Denote price Qnt and the marginal likelihood of default in period s by δs . Then: Qnt =

(1 − δt+1 )(1 − δt+2 ) . . . (1 − δt+n ) (1 + Ynt )n

(1)

The risky (continuously compounded) yield to maturity rnt := −n−1 log[Qnt ], thus: rnt = −n−1 log{(1 − δt+1 )(1 − δt+2 ) . . . (1 − δt+n )} + ynt (2) P which for δ << 1 is well approximated by rnt = n−1 nj=1 δt+j + ynt .

From model to data: Default and non-default components of the spread

The default component Denote the theoretical spread with a tilde: S˜ nt := rnt − ynt . This depends on the expected future likelihood of default {δj }t+n . j=t+1 The non-default component (NDC) Denote the measured/observed spread (data, so no tilde) by Snt . The residual difference between measured spreads and their theoretical default component Snt − S˜ nt is the non-default component of the spread. The NDC is unobserved, since S˜ nt is unobserved.

From model to data: Estimating the NDC using a VAR A VAR with spreads and defaults Denote the measured default rates by Dj . Then let yt = [w0t , x0t ]0 where wt is a vector of standard macroeconomic variables (detailed later) and x0t = [Dt , Snt ]0 contains our credit market variables. The VAR B(L)yt = ut yields both unconditional and conditional forecasts (impulse-responses) for Dt+j , j = 1 . . . hd . Estimating credit shocks 1 2

3

0 An (orthogonal) shock widens the spread to Snt

The impulse-response function shows how the conditional forecast is revised, i.e. it tells us D0t+j − Dt+j 0 −S ˜ t+j is compared to the The change in the default component ∆S˜ nt = S˜ t+j change in the measured spread - giving an estimate of the change in the NDC

From model to data: Identifying credit shocks Sign restrictions If ∆S˜ nt > ∆Snt then ∆NDC < 0: a fundamental macroeconomic shock If ∆S˜ nt < ∆Snt then ∆NDC > 0: a credit shock Note that following a credit shock: The responses for variables of interest wt are left unrestricted Expected default cannot be the sole cause of higher spreads; but credit shocks are permitted to induce a rise in default rates Decompose spreads without requiring a particular parametric model or class of models Identification of macro shocks is permitted (e.g. via additional sign restrictions), requiring only that they do not raise the NDC of spreads

Review: The VAR identification problem

Vector moving average representation of the VAR may be written yt = C(L)ut . Wish to transform the ut to orthogonality to see the ‘distinct patterns of movement’ in the system. A Choleski factorization Σ = A0 A00 gives transformation 0 ut = A0 v∗t E[v∗t v∗t ] = I (3) But for any orthogonal matrix Q can form A = A0 Q0 and associated structural shocks vt = Qv∗t such that the transformation preserves the reduced form covariance as in: ut = Avt = A0 Q0 Qv∗t

E[Avt v0t A0 ] = AA0 = Σ

(4)

Review: Sign restrictions algorithm Candidate structural VMA representation is given by yt = C(L)A0 Q0 Qv∗t = C(L)Avt

(5)

Choose a number of iterations M. Then the mth loop is: 1

Draw reduced form parameters Σ(m) , B(m) from posterior distribution

2

Draw orthonormal matrix Q(1) (‘multivariate uniform’)

3

Given Q(1) , check if our sign restrictions are checked on the columns of the matrix: R(h) = [A00

4

A00 C01

A00 C02

... A00 C0h ]0 Q(1)

(6)

If every element of R(h) satisfies the sign restrictions, save draw Q(1) , Σ(m) and B(m) ; if not, draw another Q until a R(h) is found that does

Rubio-Ramírez et al. (2005).

Questions?

Data: High-yield default rates

Moody’s default rate calculation The data is a smoothed estimate of the marginal default likelihood P11 Dt =

s=0 dt−s Nt−11

where dj is the number of defaults in period j, and Nt−11 is the size of the cohort, which is adjusted for withdrawals over time. The rate at which issuers in the 12-months-lagged cohort have defaulted.

Data: High-yield bond spreads

Corporate bond yields are taken from the Merrill Lynch Corporate High Yield Master database. Cash pay only Average rating remains constant around B1/B2, ‘subject to high credit risk’ Macaulay duration between 6 years 3 months and 4 years 9 months Gap between effective yield and yield to maturity no larger than 15bps Interpolate Nov-1982 to Nov-1984 using quarterly index of Gertler and Lown and monthly Baa index from FRB Government zero coupon bond yields from the Gürkaynak, Sack and Wright database.

Data: Spread measurement error 2.5%

Differencce due to M Maturity Missmatch

2.0%

1.5%

1.0%

0 5% 0.5%

0 0% 0.0%

-0.5% -0 5% Nov-84 Mar-87 Jul-89 Nov-91 Mar-94 Jul-96 Nov-98 Mar-01 Jul-03 Nov-05 Mar-08 NBER Recession

High yield bond spread less constant maturity spread

Data: Macro variables Include a standard set of variables for a (monthly) monetary US VAR: industrial production, core CPI, effective federal funds rate. Perhaps less ‘standard’ but important in the current context: Money although frequently omitted, actually contains summary information about broad credit market conditions (Nelson) and is typical in exercises aimed at identifying monetary policy shocks (Bernanke and Mihov, Uhlig et cetera) Equity prices are known to be closely associated with below-investment grade yields, and carry important information concerning macro state variables All variables enter in levels, or log-levels as appropriate. Where relevant, measured at month-end.

Credit shock impulse-response 0.5

Ind. Production (%)

0.1 0.0 −0.1 −0.2

0 −0.5 0

10 20 30 Fed Funds Rate (ppt)

40

50

60

0

0.1

0.5

−0.1

0

0 10 20 2.0 Equity Prices (%)

30

40

50

60

30

40

50

60

10 20 Default Rate (ppt)

30

40

50

60

30

40

50

60

0.0

−2.0

−0.2 10 20 30 40 High−Yield Bond Spread (ppt)

50

60

50

60

0.25 0.00 0

10 20 Real M1 (%)

0.2

0

0

0

Core CPI (%)

10

20

30

40

0

10

20

Interpretation Output response Financial frictions do appear to matter for firms Consistent with standard theory that says higher cost of funds raises effective input costs, lowers input demand, and so output Sluggish recovery in output consistent with theories of debt overhang (Myers, 1977) Default response Default rates roughly unchanged on impact, but do respond to credit shocks Financial stress leads to balance sheet reorganization: asset sales, mergers, renegotiation of private debt ... before move to default (Asquith, Gertner and Scharfstein, QJE 1994) Delayed issuance could account for lower aggregate default rate in the medium-run (the ‘aging effect’, Helwege and Kleiman, 1996)

Historical decomposition: output 25%

(a) Industrial production

20% 15% 10% 5% 0% -5% -10% -15% Nov-83 Jan-87

Mar-90 May-93

NBER Recession

16

Jul-96

Sep-99 Nov-02 Jan-06

Deviation from Base Projection ((b)) High-yield g y bond spread p

Mar-09

Shock Contribution

Nov-83 Jan-87

Mar-90 May-93

Jul-96

Sep-99 Nov-02 Jan-06

Historical decomposition: spreads NBER Recession Deviation from Base Projection 16

Mar-09

Shock Contribution

((b)) High-yield g y bond spread p

14

Percentage Points

12 10 8 6 4 2 0 -2 -4 -6 Nov-83 Jan-87

Mar-90 May-93

NBER Recession

Jul-96

Sep-99 Nov-02 Jan-06

Deviation from Base Projection

Mar-09

Shock Contribution

Historical decomposition: summary The cumulative effect of credit shocks is negative for output in every recession. Can ask a counterfactual question: What proportion of the fall in output (or the increase in spreads) would not have occurred if credit shocks had been zero during each recession, but other (independent) shocks had taken the same values? 1990-1 Peak-to-trough output: 1/3

spreads: 2/5

2001 Peak-to-trough output: 3/4

spreads: 2/3

2007-9 The Great Recession June-December (peak) 2007; spreads: 1/3 2007 Peak-to-end 2009Q1; output: 2/5 2007 Peak-to-November 2008 (widest spread); spreads: 1/5

Forecast error variance decomposition Contribution of credit shocks to output fluctuations at one year similar to estimated contributions of monetary policy shocks (Christiano, Eichenbaum and Evans, 2005). About a 10-15% average contribution (imprecisely estimated) at 1-2 year horizon, so credit shocks ‘matter’ for ‘normal’ business cycle fluctuations, but probably not that much. Estimates are way below estimates based on a recursively-identified VAR, or estimates from some quantitative-theoretic models. h IP D S

3 15 (4-38) 9 (1-32) 8 (3-22)

6 16 (4-36) 8 (2-26) 7 (2-20)

Note: 16-84 percentiles in parentheses.

12 13 (3-35) 7 (2-18) 6 (2-10)

24 10 (2-30) 5 (1-13) 6 (2-18)

60 8 (2-26) 7 (2-18) 7 (3-20)

Validation: Estimated credit shocks 6

4

2

0

-22

-44 Nov-83

Jan-87

Mar-90

May-93

Jul-96

NBER Recession

Sep-99 Credit shock

Nov-02

Jan-06

Mar-09

Validation: Monetary policy shocks MP shocks an obvious confounding influence on estimates: Source of high-frequency variation in interest rates Affect measurement error from mismatched maturities Regression of credit shocks on MP shocks Use a MP shock measure derived from external data on Fed Funds Futures (Faust, Swanson and Wright, JME 2004). Then run a regression (as in Rudebusch, IER 1998): vˆ t = −.153 + 1.99 εˆFFF , t (1.37) (.942)

R2 = .006, N = 131

No statistical relationship between credit and MP shocks. Greater confidence (a) in identification scheme, and (b) in proper measurement of the spread.

Validation: Liquidity shocks Why liquidity matters: Average trading volumes for corporate bonds roughly 5% of Treasuries (Bessembinder and Maxwell, 2008) Many studies confirm empirically relevant bond-specific liquidity effects in prices (Chen, Lesmond and Wei, 2007) Government bonds are ‘special’ (Duffee, 1996) Regress the on-the-run/off-the-run spread on own lags and lags of all other variables to produce estimate of surprise εˆONOFF t Regression of credit shocks on liquidity shocks vˆ t = −

.108 − 1.97 εˆONOFF , R2 = .003 t (1.56) (.998) 1984 : 5 − 2009 : 4

N.B. Sample May 1984 - Apr. 2009

Validation: Risk premium Why risk matters: Theoretical considerations point to importance of risk premiums for pricing, even when investors can diversify (Duffie and Singleton, 2003) Evidence that a significant portion of time series variation in spreads is accounted for by changes in risk premiums (Elton, Gruber, Agrawal and Mann, 2001) ... but risk premiums are hard to measure VIX is a measure of 30 day expected volatility in the S&P; trading is dominated by deep out-of-the-money index puts, so can interpret as cost of insurance =⇒ likely related to risk compensation. Construct ‘surprise’ as before. Regressions of shocks on VIX surprises vˆ t = −

.182 + 2.68 εˆVIX R2 = .070 t (2.33) (3.72) 1990 : 7 − 2009 : 4

N.B. Sample Jul. 1990 - Apr. 2009

Conventional restrictions: Variance decomposition h IP D S

3 11 (7-16) 3 (1-6) 83 (77-88)

6 23 (16-30) 15 (9-22) 81 (73-87)

12 34 (24-44) 46 (36-56) 75 (66-84)

24 38 (25-52) 57 (46-69) 70 (56-80)

60 33 (18-51) 50 (37-61) 60 (46-71)

Why the difference? Strong default channel for credit shocks - amplifies the response of output Most variation in spreads due to credit shocks - little role for macro shocks Estimated shocks confound monetary policy disturbances vˆ rec 3.99 εˆFFF , t t = −.118 + (1.12) (1.89)

R2 = .029, N = 131

Caveats

Several caveats to take into account Identify a subset of shocks - estimated effects are a lower bound Peso problems - alternative approach could use Bekaert, Hodrik and Marshall’s (JME 2001) approach on corporate bonds Segmented markets - endogenous changes in quantities could drive spreads independent of default; Korajczyk and Levy (JFE 2003) find evidence against for constrained firms Linear VAR is an approximation - and underestimates the effect of a shock if there are ‘regimes’ (Balke, REStat 2000)

Summary and conclusion Are credit markets a source of shocks? Focus on long-term funding markets Model macro and credit variables jointly, allowing for two-way interactions Use simple bond pricing model to derive identifying restrictions, without resorting to recursive causal scheme Good evidence for financial frictions/financial accelerator effects Exogenous shocks on average make a modest contribution to average output fluctuations... ...although they did knock 6.4% off industrial production during the Great Recession Results robust to various types of misspecification To generate large effects, need to assume a recursive VAR, which is shown to be misspecified

End of main presentation.

Actual default rates and structural model distance to default 14

12

10

8

6

4

2

0 Nov-82

Jun-84

Jan-86

NBER Recession

A Aug-87

Mar-89

Oct-90

May-92

Jan-94

Default Likelihood (Vassalou-Xing)

A Aug-95

Mar-97

Oct-98

High yield default rate

Loss and default rates 14

9 8

12

7 10

8

5

6

4 3

4 2 2

1

0 l SSep-90 Nov-92 Jan-95 Mar-97 May-99 Jul-01 l SSep-03 Nov-05 Jan-82 Mar-84 May-86 Jul-88 NBER Recession

Default rate

Loss rate

0

Percent

Percent

6

Example: The agency cost model

Main assumptions Firm with risky project seeks leverage to raise expected returns Costly state verification leads to ‘debt-like’ contracts Investor and firm are risk neutral Static, one shot setting/no aggregate uncertainty Main implications Macro shocks cause spreads ↑ and defaults ↑: the default channel Credit shocks cause spreads ↑ and defaults ↓: the non-default channel

Agency cost model Firm operates linear technology, production function θωy Firms repayments are Rd (y − n); if idiosyncratic shock is ω, ¯ firms break even θωy ¯ = Rd (y − n). Firm’s expected profit: Z θf (ω)y ¯ :=

θωydΦ(ω) − [1 − Φ(ω)]θ ¯ ωy ¯

ω>ω ¯

(7)

Investor’s expected profit: Z θg(ω)y ¯ := (1 − µ)

ω<ω ¯

θωydΦ(ω) + [1 − Φ(ω)]θ ¯ ωy ¯

where µ default costs, θ reduced form macro shock, ω idiosyncratic shock, Φ(ω) ¯ likelihood of default.

(8)

Agency cost model Contracting problem solves max θf (ω)y ¯ s.t. θg(ω)y ¯ =y−n {y, ω} ¯

(9)

The first order conditions include f$ (ω) ¯ g$ (ω) ¯ λ(ω) ¯ f (ω) ¯ + λ(ω)g( ¯ ω) ¯

λ(ω) ¯ = −

(10)

θ =

(11)

The spread Rd = ω/g( ¯ ω) ¯ is an increasing function of the default threshold ω ¯ R$ (1 − µ) 0 ωdΦ(ω) + µφ(ω) ¯ ω ¯ ∂Rd = >0 (12) ∂ω ¯ [g(ω)] ¯ 2 Macro shocks work through the default channel since dω/dθ ¯ > 0 and d dR /dθ > 0.

Credit shocks: The non-default channel Figure: The effect of an increase in µ on y and ω ¯

a

y 6 b b0

a0

q

b0 n b

q

ω(µ ¯ 0 ) ω(µ) ¯

-

ω ¯

Note: The investor’s initial break-even constraint is labeled na, and the borrower’s initial iso-profit line is bb.

Credit shocks: The non-default channel

What does an increase in information frictions µ do to spreads? dRd dω ¯ ∂Rd = − ωg ¯ µ (ω) ¯ dµ dµ ∂ω ¯ We know that investor income is directly reduced gµ < 0, so the sign is indeterminate... ... so calibrate the model by matching moments.

Credit shocks: The non-default channel Three parameters to calibrate: {Φ, σω , µ}: The default rate Φ(ω) ¯ is matched to average defaults on speculative grade bonds. Estimate remaining parameters ψ = (σω , µ)0 from sample orthogonality conditions for the expected recovery rate and the external finance premium: $(ψ)

Z (1 − µ) 0

T 1X ω RRt = 0 dΦ(ω; σω ) − Φ(ω(ψ); ¯ σω ) T t=1 T X

ω(ψ) ¯ 1 − g(ω(ψ), ¯ ψ) T

t=1

St = 0

Credit shocks: The non-default channel

Table: Sensitivity of Rd to a change in µ

Recovery Rate % 10 20 30 40 50

1.5 0.52 2. 4.6 8.5 13.

2.0 0.62 2.6 6 10. 13.

2.5 0.83 3.4 7.6 10. −50

Spread (Ann. %) 3.0 3.5 4.0 1.16 1.7 2.7 4.7 6.1 −0.14 8.2 −17. −6K −39. −12K -

4.5 4.2 −18K -

5.0 −173 -

Note: All entries are ×104 . Premium is in annual percentage points, with taxation assumed to contribute .75ppts. The assumed annual default rate is 4.9%. Entries marked as missing require inadmissible parameter values or are not attainable.

Fry-Pagan critique

Critique applies to any innovation accounting method etc. Let R(h) be the impulse-response matrix for n variables and m shocks up to horizon h. Define: i  h ¯ (k) /std(R(h)) φ(k) = vec R(h)(k) − R(h) Pick a single model k that is ‘closest’ in one of the following senses to the median impulse response: k(∞) = arg min{||φ(k) ||∞ } k

Estimation

Stack y to a T × n matrix, etc. and write the system as Y = XB + u ˆ Σ). ˆ Then the prior is given by Let the MLEs be denoted (B, −1 Σ−1 ∼ Wn (Σˆ /T, T),

ˆ Σ ⊗ (X0 X)−1 ) D(vec(B)|Σ) ∝ J × N(vec(B),

where J = 0 if any element of the spectrum of I − B is greater than 1 + η in absolute value; degree of permitted explosiveness η = 0.005.

Which channel matters? Friedman and Kuttner (REStat 1998) Offer four possible channels through which the paper-bill spread might signal a recession, but attribute a limited role to default risk Stock and Watson (JEL 2003) Spreads ‘have the potential to provide useful forecasts of real activity, and at times they have, but the obvious default risk channel appears not to be the relevant channel by which [they] have their predictive content’ Gilchrist, Yankov and Zakrajšek (JME 2009) The information content of credit spreads for real activity is concentrated at lowto-medium categories of credit risk

Which channel matters? Collin-Dufresne, Goldstein and Martin (JoF 2001) As much as 3/4 of the change in credit spreads unexplained after proxies for structural determinants of default risk, and recovery rates are accounted for. Longstaff, Mithal and Neis (JoF 2005) Use data on credit default swaps, put the size of total default compensation at between 1/2 and 4/5 of the spread, but also find a significant time-varying nondefault component. Chen, Lesmond and Wei (JoF 2007) Find a significant association between corporate bond liquidity and yield spreads using three different liquidity measures.

What are credit shocks? Nolan and Thoenissen (JME, 2009) Costly state verification. Financial shocks cause exogenous changes in net worth. An adverse shock worsens agency problems, raising the finance premium. Jermann and Quadrini (NBER WP, 2009) Limited contract enforceability. Financial shocks affect the contract enforcement constraint; adverse shock reduces the recovery value of firm assets in the case of renegotiation This paper: Credit Market Shock A credit market shock is any innovation to the non-default component of spreads. Thus, an adverse credit market shock causes wider spreads, without an increase in the likelihood of future defaults

Short term funding to shadow banks collapsed Per centage Change in Balance Sheet, ABS Issuer s

0. 4

0. 4

0. 3

0. 3

0. 2

0. 2

0. 1

0. 1

0. 0

0. 0

-0. 1

-0. 1

-0. 2

-0. 2 90

95

Sour ce: Haver Anal yt i cs

00

05

Long term non-financial corporate funding remained robust Net Nonfinancial Bond Issuance, Pr opor tion of Stock

0. 20

0. 20

0. 16

0. 16

0. 12

0. 12

0. 08

0. 08

0. 04

0. 04

0. 00

0. 00

-0. 04

-0. 04 90

95

Sour ce: Haver Anal yt i cs

00

05

Validation: Conventional restrictions 0.5

Ind. Production (%)

0.1

0

0

−0.5

−0.1

−1.0 0 10 20 30 0.2 Fed Funds Rate (ppt)

40

50

60

−0.2 0 10 20 1.5 Real M1 (%)

0.0

1.0

−0.2

0.5

0

10 20 Equity Prices (%)

30

40

50

60

0

0

0.5

−2.0

0.0

0 10 20 30 40 1.0 High−Yield Bond Spread (ppt)

50

60

50

60

0.5 0.0 0

10

20

30

40

Core CPI (%)

0

10 20 Default Rate (ppt)

10

20

30

40

50

60

30

40

50

60

30

40

50

60