Has the credit supply shock asymmetric effects on macroeconomic variables?

Colombo, Valentina; Paccagnini, Alessia Working Paper Has the credit supply shock asymmetric effects on macroeconomic variables? Quaderni - Working Paper DSE, No. 1140 Provided in Cooperation with: University of Bologna, Department of Economics Suggested Citation: Colombo, Valentina; Paccagnini, Alessia (2020) : Has the credit supply shock asymmetric effects on macroeconomic variables?, Quaderni - Working Paper DSE, No. 1140, Alma Mater Studiorum - Università di Bologna, Dipartimento di Scienze Economiche (DSE), Bologna, https://doi.org/10.6092/unibo/amsacta/6308 This Version is available at: https://hdl.handle.net/10419/213537 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc/3.0/ ISSN 2282-6483 Has the credit supply shock asymmetric effects on macroeconomic variables? Valentina Colombo Alessia Paccagnini Quaderni - Working Paper DSE N°1140 Has the credit supply shock asymmetric eects on macroeconomic variables? ∗ Valentina Colombo † Alessia Paccagnini ‡ Abstract We investigate the role played by the credit supply shock across the business cycle in the U.S. over the period 1973 - 2018. We estimate a nonlinear VAR including nominal, real, monetary, and nancial variables. According to our results, a credit supply shock triggers asymmetric and negative eects on macroeconomic variables. We nd that the state-dependent forecast error variance decomposition of industrial production, employment, and ination due to the shock is from six to eight times larger in recessions than in normal times. JEL classication: C32, E32, E52 Keywords: Credit supply shock, Smooth Transition VAR, Nonlinearities. ∗ We thank Giovanni Caggiano, Efrem Castelnuovo, Raaele Giuliana, and Massimiliano Marzo for useful comments. All remaining errors are ours. † Department of Economics, University of Bologna. Email: [email protected]. ‡ Michael Smurt Business Graduate Business School, University College Dublin, and Center for Applied Macroeconomic Analysis. Email: [email protected]. Non - Technical Summary Financial shocks have been recognized as important drivers for explaining macroeconomic dynamics via the well-known nancial acceleration mechanism. Since the onset of the Great Recession, the literature has renewed interest in the interaction between credit supply shock and macroeconomic activities. Most empirical contributions study this interaction in a linear set-up. They nd that credit supply shocks have contractionary eects on macroeconomic variables. However, empirical evidences highlight that the macro-nancial linkages may be nonlinear. Such nonlinear relation has been scrutinized in studying the role of credit markets in the transmission mechanisms of economic shocks rather than studying the asymmetric eect of a credit supply shock per se . We contribute to the state of art studying whether credit supply shocks aect asymmetrically macroeconomic variables over the business cycle. We model variables with a Smooth Transition VAR (STVAR) in which an exogenous credit supply shock is allowed to aect macroeconomic variables conditional on the states of the economy ("Recessionary Periods" vs "Normal Times"). The credit supply shock is identied by appealing to the excess bond premium indicator (EBP). Fitting the post-WWII U.S. monthly data in a Smooth Transition VAR, we nd systematic asymmetries across business cycle phases in the response to a credit supply shock. We conjecture that the nancial-accelerator mechanism may play a larger role in severely deepen the macroeconomic activity, depending on which phase of the business cycle the economy is when the credit shock occurs. Findings reveal that during normal times, the EBP shock impacts negatively only industrial production. Dierently, there is an amplication eect when the economy is already in recessions: an exogenous contraction in the supply of credit aects negatively not only industrial production, but also ination and employment. As our results suggest, negative business cycle eects due to nancial shocks get magnied when the economy is already in a bust phase. Our impulse responses conrm the role of credit supply shocks in driving macroeconomic uctuation quantifying a more than double drop in the macroeconomic variables in recessions than in normal times. Moreover, we show that contractions in the supply of credit in recessions (but not in normal times) work as a demand shock, in the sense of being associated with a fall in output and prices at the same time. Interesting, the shock seems to explain a fraction of the variance of real variables (industrial production and employment) and ination that is from six to eight times larger in recessions than in normal times. Moreover, the credit supply shock appears to be the rst source of uctuation of employment in recessions but not in normal times on which the contribution of macroeconomic shocks prevail. The EBP is more important than monetary shocks in explaining macroeconomic uctuation in recessions. 2 1 Introduction Financial shocks have been recognized as important drivers for explaining macroeconomic dynamics via the well-known nancial acceleration mechanism (see e.g., Bernanke and Blinder, 1988; Gilchrist and Zakraj²ek, 2012). Since the onset of the Great Recession, the literature has renewed interest in the interaction between credit supply shock and macroeconomic activities (Gertler and Gilchrist, 2018). Most empirical contributions study this interaction in a linear set-up (see e.g., Gilchrist and Zakraj²ek, 2012; López- Salido, Stein, and Zakraj²ek, 2017; Caldara, Fuentes-Albero, Gilchrist, and Zakraj²ek, 2016; Faust, Gilchrist, Wright, and Zakraj²ek, 2013; Stock and Watson, 2012). They nd that credit supply shocks have contractionary eects on macroeconomic variables. However, empirical evidences highlight that the macro-nancial linkages may be nonlinear. Such nonlinear relation has been scrutinized in studying the role of credit markets in the transmission mechanisms of economic shocks (see e.g., Alessandri and Muntaz, 2019; Rüth, 2017; Alessandri, Conti, and Venditti, 2017) rather than studying the asymmetric eect of a credit supply shock per se . As for the literature dealing with asymmetries of credit supply shocks, Barnichon, Matthes, and Ziegenbein (2019) nd that the eects of such shock depend on its size and sign. It highlights that some asymmetries may be at work. Do credit supply shocks aect asymmetrically macroeconomic variables over the business cycle? To answer our question, we model variables with a Smooth Transition VAR (STVAR) in which an exogenous credit supply shock is allowed to aect macroeconomic variables conditional on the states of the economy ("Recessionary Periods" vs "Normal Times"). The credit supply shock is identied by appealing to the excess bond premium indicator (EBP) constructed by Gilchrist and Zakraj²ek (2012) and plotted in gure 1. Fitting the post-WWII U.S. monthly data in a Smooth Transition VAR, we nd the answer to our question to be positive: (i) a one standard deviation shock leads to systematic asymmetries across business cycle phases in the responses to a credit supply shock; (ii) the variance of real and nominal variables explained by the shock is from six to eight times larger in recessionary periods than in normal times; (iii) the shock triggers eects of demand-type in recessions but not in normal times. Our paper is structured as follows. Section 2 describes the data and the identication strategy implemented in the Smooth Transition VAR. Section 3 discusses the results. Section 4 concludes. 2 Data and Methodology We study the asymmetric eects of a credit supply shock relying on a Smooth Transition VAR (STVAR) which is dened as follows: 3 Xt=F(zt−1)ΠR(L)Xt+ (1 −F(zt−1))ΠNT (L)Xt+εt, (1) εt∼N(0,Ωt), (2) Ωt=F(zt)ΩR+ (1 −F(zt))ΩNT , (3) F(zt) = exp(−γzt)/(1 + exp(−γzt)), γ > 0, zt∼N(0,1). (4) where Xt is a set of endogenous variables, Π(L)R and Π(L)NT are the polynomial matrices capturing the dynamics of the system during recession and normal times, respectively. The vector of reduced-form residuals ( εt ) has zero-mean and heteroskedastic variance-covariance matrix Ωt . The function F(zt−1) is the logistic function capturing the probability of being in a recession. It depends on the state variable zt and on the smoothness parameter γ which dictates how smooth is the transition from one regime to another (i.e. lower value a higher smooth, higher value lower smooth). The transition variable zt is the standardized backward-looking 12-month moving average growth rate of industrial production. As in Auerbach and Gorodnichenko (2012) and Caggiano, Castelnuovo, Colombo, and Nodari (2015), we calibrate the smoothness parameter γ to match the probability of being in recessions as identied by the NBER business cycle dates (15% in our sample). The recessionary phase is dened as a period in which Pr(F(zt)⩾0.85) ≈15% . It means that the economy spends about 15% of time in recession and 85% in normal times. This implies setting γ= 2.3 . 1 . Following Caldara, Fuentes-Albero, Gilchrist, and Zakraj²ek (2016), Xt includes (from the top to the bottom): (i) the CPI ination; (ii) the manufacturing industrial production growth; (iii) the employment rate; (iv) the EBP; (v) the cumulated value-weighted total stock market (log) return; (vi) the nominal 10-year Treasury yield, and (vii) the nominal 1-year Treasury yield 2 . The EBP captures the risk-bearing capacity of the nancial intermediate sector. Based on market prices of individual corporate bonds traded in the secondary market, Gilchrist and Zakraj²ek (2012) construct a credit spread of the U.S. nonnancial corporations over Treasury bond yields with identical cash ow and maturity characteristics. Then, they decompose such spread into a component reecting the countercyclical move- 1 The choice is consistent with the threshold value of z=−0.95% discriminating recessions and normal times. In particular, if the realizations of the standardized transition variable zt is lower (higher) than the threshold value z , it will be associated with recessions (normal times). The transition variable zt has been standardized to be comparable to those employed in the literature. Following Auerbach and Gorodnichenko (2012), we rely on the lagged value of z in Equation (1) to avoid contemporaneous feedbacks from the shock into the state of the economy. The online appendix, Section A, reports the gure of the F(zt) . 2 To overcome the fact that the Federal Funds rate was at the zero lower bound, we rely on the one-year Treasury maturity since it accounts for term structure eects due, for instance, to forward guidance (see e.g., Gertler and Karadi, 2015; Alessandri, Conti, and Venditti, 2017) 4 ment in default risks and a residual, the so-call Excess Bond Premium (EBP) which is linked to the nancial conditions of the issuer. Thus, the EBP captures the extra return that investors demand to hold corporate bonds over and above the compensation for the credit risk (credit market sentiment). Figure 1: Excess Bond Premium vs Business cycle Notes: The shaded area indicate the U.S. recessionary phases (1973:M1-2018:M12), whereas the blue line refers to the Excess Bond Premium indicator by Gilchrist and Zakraj²ek (2012). The credit supply shock is identied via the Cholesky-decomposition with the assumptions provided by Gilchrist and Zakraj²ek (2012). In other words, the slow-moving variables (CPI, Industrial Production, and Employment) are ordered before the shock, whereas the fast-moving variables (risk-stock return and free rates) are ordered after that. It means that we "purge" our credit supply indicator from the contemporaneous movements of our macroeconomic variables, therefore sharpening the identication of the credit supply shock. Hence, an unexpected change in EBP will be orthogonal to the business cycle at time t. We estimate the STVAR in (1) via the Markov-Chain Monte Carlo simulation (Chernozhukov and Hong, 2014) and we model the endogeneity of the transition from one state to another one after a credit supply shock occurs computing the Generalized Impulse Response Functions (GIRFs) proposed by Koop, Pesaran, and Potter (1996). Since the GIRFs depend on the initial condition, we study the evolution of the GIRFs over histories (i.e., recession versus normal times). Our data are monthly and span the period 1973M1-2018M12. The beginning of the period depends on the availability of the EBP indicator. We estimate a nonlinear 5 VAR including ve lags, as indicated by the Akaike information criterion. The data are retrieved from the Federal Reserve Bank of St. Louis, apart from the EBP and the stock return downloaded by the Board of Governors of the Federal Reserve System's website and the Center for Research in Security Prices (CRSP), respectively. Before estimating the STVAR in (1), we test the linearity of our VAR and the LM test suggests a strong rejection of the linearity for the system as a whole in favor of a particular nonlinear model, the STVAR. 3 3 Results Figure 2 depicts the GIRFs of an EBP shock that is orthogonal to the state of the economy. An unexpected increase of one standard deviation in EBP generates asymmetric eects on the economy and nancial markets. A higher EBP triggers negative macroeconomic uctuations both in recessions (rst column of gures 2) and in normal times (second column). However, the responses of our variables are larger (in absolute value) and more persistent in recessions than in normal times. Indeed, the shock causes a trough response of industrial production that is more than twice larger in recessions than in normal times (-1.3% versus -0.6%). The industrial production goes back to its steady-state two years after the shock occurs in normal times but it takes one year more to turn to its pre-shock level during a recessionary period. Meanwhile, the fall in employment is four times larger in recessions than in normal times (-1.2% versus -0.3%). The deationary impact and the reduction in employment of the shock are statistically signicant only during recessionary periods but not in normal times. The Federal Reserve lowers the interest rate in both states by adopting an expansionary monetary policy. Despite an easing monetary policy environment, the stock market returns drop. The statistical test based on the empirical density of the dierence between the reactions of macroeconomic and nancial variables in recessions and normal times (third column of gure 2) conrms that quantitatively the responses are dierent across regimes from a statistical point of view. Overall, our results highlight the systematic asymmetries across business cycle phases in the response to a credit supply shock. A possible interpretation of our ndings is provided by Gilchrist and Zakraj²ek (2011). Working with a DSGE framework, they show that an adverse nancial shock conceptually in line with an increase in EBP reduces the risk-bearing capacity of the nancial sector and, consequently, the supply of credit available to potential borrowers. Such a reduction in credit supply is associated to a drop in rms' cash ows and in the value of asset prices, and a slowdown in economic activity. We conjecture this nancial-accelerator mechanism may play a larger role in severely 3 See Section C, D and E of the online Appendix for further details related to the linearity test, the STVAR estimation and the computation of the GIRFs. 6 deepen the macroeconomic activity, depending on which phase of the business cycle the economy is when the credit shock occurs. Findings reveal that during normal times, the EBP shock impacts negatively only industrial production. Dierently, there is an amplication eect when the economy is already in recessions: an exogenous contraction in the supply of credit aects negatively not only industrial production, but also ination and employment. As our results suggest, negative business cycle eects due to nancial shocks get magnied when the economy is already in a bust phase. Figure 2: Generalised impulse responses (GIRFs) to credit supply shocks Notes: The gure reports the generalized impulse responses (GIRFs) to an unanticipated U.S. credit supply shock in recessions (rst column), in normal times (second columns), and the median realizations of the dierences between generalized impulse responses in recessions and normal times (third column). The red and blue lines denote the median GIRFs in recessions and in normal times, respectively. The magenta lines refer to the median of the dierence realizations between the two states of the world. Shaded bands denote condence intervals at 68% levels. The responses of ination, industrial production ad employment are accumulated. The horizontal axis identies months, whereas the vertical axis is expressed in percentage points. Relying on time-varying parameters VAR, Gambetti and Musso (2019) nd that credit supply shocks are particularly important during recessions. Our impulse responses con- rm the role of credit supply shocks in driving macroeconomic uctuation quantifying a more than double drop in the macroeconomic variables in recessions than in normal times. Moreover, we show that contractions in the supply of credit in recessions (but not in normal times) work as a demand shock, in the sense of being associated with a fall in 7 ( k x p + q ) vector of exogenous variables which includes lagged variables ( k ) and a vector of constants. The transition variable is zt , while Θ0 and Θi are matrices of parameters. In our empirical assessment, we have p = 7 as number of endogenous variables, q = 1 as number of exogenous variables, and k = 5 as number of lags. Under the null hypothesis of linearity, we assume Ho: Θi =0 ∀i . The Teräsvirta and Yang (2014) test features the following four steps: 1) We estimate the restricted model ( Ho: Θi =0 ∀i ) by regressing Xt on Yt . We collect the residual ˜ E calculating the matrix for the residual sum of squares RSS0 = ˜ E ' ˜ E . 2) We run an auxiliary regression of ˜ E on ( Yt , Zn ) where the subscript n indicates the n-order Taylor expansion of the transition function. We save the residuals ˜ Ξ computing the matrix for the residual sum of squares RSS1 = ˜ Ξ ' ˜ Ξ . 3) We compute the test-statistic: LM =Ttr[RSS−1 0(RSS0−RSS1)] = T[p−tr(RSS−1 0RSS1)]. (6) Under the null hypothesis, the test statistic is distributed as a χ2 with a number of degree of freedoms equals the number of restrictions, p ( kp + q) . We compute two LM-type linearity tests xing the value of the n-order of the Taylor expansion equal to n= 1 and n= 3 (as proposed by Luukkonen, Saikkonen, and Teräsvirta, 1988). In our estimation, LM=503.4 and LM=1254.6 when n= 1 and n= 3 , respectively. The corresponding p-value in both tests is zero. In other words, our model presents non-linear dynamics. D Appendix: Estimation of the Non-linear VARs Our STVAR model (1)-(4) is estimated via maximum likelihood. The log - likelihood function is as follows: logL =const −1 2 T X t=1 log|Ωt| − 1 2 T X t=1 ε0 tΩ−1εt, (7) where the vector of residuals εt=Xt−(1 −F(zt))ΠNT Xt−1−F(zt)ΠRXt−1 . Our purpose is to estimate the parameters Ψ = {ΩR,ΩNT ,ΠR(L),ΠNT (L)} , where Πj(L) = [Πj,1, ..., Πj,p] , j∈ {R, NT} . Due to the high non-linearity of the model its estimation is problematic using standard optimisation procedures. Hence, as in Auberbach and Gorodnichenko (2012), we employ the procedure as described as follows. Conditional on γ , ΩR , ΩNT , where γ is the slope parameter calibrated as described in section 2, the model is linear in ΠR , ΠNT . Hence, for a given guess on γ , ΩR , ΩNT , the coecients ΠR , ΠNT can be estimated by minimizing 1 2PT t=1ε0 tΩ−1εt . Hence, we can re-write the regressors as below. 14 Let Wt= [F(zt)Xt−1(1−F(zt))Xt−1...F(zt)Xt−p(1−F(zt))Xt−p] be the extended vector of regressors, and Π= [ΠR(L)ΠNT (L)] . Consequently, we can write εt=Xt−ΠW0 t . In this case, the objective function becomes: 1 2 T X t=1 (Xt−ΠW0 t)0Ω−1 t(Xt−ΠW0 t). (8) We can show that the rst order condition with respect to Π is given by: vecΠ0= ( T X t=1 [Ω−1 t⊗W0 tWt])−1vec( T X t=1 W0 tXtΩ−1 t). (9) We iterate this procedure over dierent sets of values for {ΩR , ΩNT } (conditional on a given value for γ ). For each set of values, Π is obtained and the logL (7) is calculated. Due to the high non-linearity of the model in its parameters, we might get several local optima. Then, it is recommended to try dierent starting values of γ . To guarantee positive deniteness of the matrices ΩR and ΩNT , we focus on the alternative vector of parameters Ψ = {chol( ΩR ), chol( ΩNT ), ΠR (L), ΠNT (L) }, where chol means the Cholesky decomposition. We compute the condence intervals using a Markov Chain Monte Carlo (MCMC) algorithm developed by Chernozhukov and Hong (2003) (CH hereafter). This methodology gives us both a global optimum and densities for the parameter estimates. We implement the CH estimation via a Metropolis-Hastings algorithm. Given a starting value Ψ0 , the procedure constructs chains of length N of the parameters of the estimated model following two steps: Step 1: Draw a candidate vector of parameter values Θ(n)=Ψ(n)+ψ(n) for the chain's n+ 1 state, where Ψ(n) is the current state and ψ(n) is a vector of i.i.d. shocks drawn from N(0,ΩΨ) , and ΩΨ is a diagonal matrix. Step 2: Set the n+1 state of the chain Ψ(n+1) =Θ(n) with probability min{1, L(Θ(n))/L(Ψ(n))} , where L(Θ(n)) is the value of the likelihood function conditional on the candidate vector of parameter values, and L(Ψ(n)) is the value of the likelihood function conditional on the current state of the chain. Otherwise, set Ψ(n+1) =Ψ(n) . The starting value Θ(0) is calculated using the second-order Taylor approximation of the model described from (1) to (4) in the section 2, hence the model can be written as regressing Xt , Xtzt , and Xtz2 t . We employ the residuals from this regression to t the expression for the reduced-form time-varying variance-covariance matrix of the VAR (as explained in the main text) using maximum likelihood to estimate ΩR and ΩNT . We can construct Ωt , conditional on these estimates and given the calibration for γ . Conditional on Ωt , we can compute the starting values for ΠR(L) and ΠNT (L) using equation (9). 15 Given the calibration for the initial (diagonal matrix) ΩΨ , a scale factor is adjusted to generate an acceptance rate close to 0.3, the typical value for this computational methods as pointed out by Canova (2007). The estimation accounts for N= 50,000 draws and we use the last 20% for inference. As described by CH, Ψ∗=1 NPT t=1Ψ(n) is consistent estimate of Ψ under standard regularity assumptions on maximum likelihood estimators. The covariance matrix of Ψ is given by V=1 NPT t=1(Ψ(n)−Ψ∗)2=var(Ψ(n)) , which is the variance of the estimates in the generated chain. E Appendix: Generalized Impulse Response Functions The Impulse Response Functions for the STVAR model are computed following the approach introduced by Koop, Pesaran, and Potter (1996) which propose an algorithm to calculate the Generalized Impulse Response Functions (GIRFs). The implementation of their procedure is composed of the following steps. 1) We construct the set of all possible histories Λ of length p= 12 : {λi∈Λ} , where Λ contain T−p+ 1 histories λi and T is the sample size ( T =551). 2) We separate the set of all recessionary histories from that of all normal times histories. We calculate the transition variable zλi for each λi . If zλi≤z∗ =-0.95 % , then λi∈ΛR , where ΛR refers to all recessionary histories; if zλi > z∗=−0.95% , then λi∈ΛNT , where ΛNT refers to all normal times histories. 3) We select at random one history λi from the set ΛR , taking ˆ Ωλi obtained as follows: ˆ Ωλi=F(zλi)ˆ ΩR+ (1 −F(zλi)) ˆ ΩNT , (10) where zλi is the transition variable computed for the selected history λi . ˆ ΩR and ˆ ΩNT are calculated from the generated MCMC chain of the parameter values during the estimation step. As in Koop, Pesaran, and Potter (1996), we consider the distribution of parameters rather than their mean values to allow for parameter uncertainty. 4) We estimate the variance-covariance matrix ˆ Ωλi using the Cholesky-decomposition: ˆ Ωλi=ˆ Cλiˆ C0 λi, (11) we orthogonalize the estimated residuals to get the structural shocks as: e(j) λi=ˆ C−1 λiˆε. (12) 5) From eλi draw with replacement h nine-dimensional shocks and get the vector of 16 bootstrapped shocks e(j)∗ λi={e∗ λi,t ,e∗∗ λi,t+1 , ..., e∗∗ λi,t+h}, (13) where h is the number of horizons for the IRFs we compute. 6) We form another set of bootstrapped shocks which are equal to (13) except for the kth shock in e(j)∗ λi which is the shock we perturb by a δ amount. We call the vector of bootstrapped perturbed shocks as e(j)δ λi . 7) We transform back e(j)∗ λi and e(j)δ λi as follows: ˆε(j)∗ λi=ˆ Cλie(j)∗ λi, (14) and ˆε(j)δ λi=ˆ Cλie(j)δ λi. (15) 8) We use (14) and (15) to simulate the evolution of X(j)∗ λi and X(j)δ λi and we construct the GIRF(j)(h, δ, λi) as X(j)∗ λi - X(j)δ λi . 9) Conditional on history λi , repeat for j =1,..., B vectors of bootstrapped residuals and get GIRF1(h, δ, λi) , GIRF2(h, δ, λi) , ..., GIRFB(h, δ, λi) . We set B =500. 10) We calculate the GIRF conditional on history λi as: ˆ GIRF(i)(h, δ, λi) = B−1 B X j=1 GIRF(i,j)(h, δ, λi). (16) 11) We repeat all previous steps for i =1,...,500 histories belonging to the set of recessionary histories, λi∈ΛR , and we get ˆ GIRF(1,R)(h, δ, λ1,R) , ˆ GIRF(2,R)(h, δ, λ2,R) , ..., ˆ GIRF(500,R)(h, δ, λ500,R) where the subscript R means that we are conditioning upon recessionary histories. 12) We take the average and we get ˆ GIRF(R)(h, δ, ΛR) , which is the average GIRF under recessions. 13) We repeat all the previous steps from 3 to 12 for 500 histories belonging to the set of all normal times and we get ˆ GIRF(NT )(h, δ, ΛNT ) . 14) We compute the 68% condence bands for the IR by picking up for each horizon of each state, the 16th and 84th percentile of the densities ˆ GIRF([1:500],R) and ˆ GIRF([1:500],NT ) . 17 References Gilchrist, S. and E. Zakraj²ek (2012). Credit Spreads and Business Cycle Fluctuations. Econometrica, 102(4), pp. 16921720. Hansen, B. (1999). Testing for Linearity. Journal of Economic Surveys, 13(5), pp. 551 574. Koop, G., M. Pesaran, and S. Potter (1996). Impulse Response Analysis in Nonlinear Multivariate Models. Journal of Econometrics, 74(1), pp. 119148. McCracken, M.W. and Ng (2016). Fred-MD: A Monthly Database for Macroeconomic research. Journal of Business and Economic Statistics, 34(4), pp. 574589. 18 