What explains the pre-crisis housing and credit boom in the US?

Transcription

1 What explains the pre-crisis housing and credit boom in the US? Esteban Prieto Sandra Eickmeier January 4 Abstract We analyze the contribution of credit supply, housing demand and monetary policy shocks to the most recent pre-crisis housing and credit boom in the US. We estimate a large time-varying parameter VAR model over the period and identify simultaneously the three structural shocks using sign restrictions. Our main findings are: (i) The housing and credit boom, which started in the early-99s and ended around 6, has initially been driven mainly by credit supply shocks. In the mid- s, however, monetary policy shocks were important drivers. (ii) The effects of monetary policy shocks on the housing and mortgage credit market and the effects of credit supply shocks on the housing market have increased over time. JEL classification:... Keywords: Housing, credit supply, monetary policy, large time-varying parameter VAR s of the authors: esteban.prieto@bundesbank.de, sandra.eickmeier@bundesbank.de. The paper has been presented at a seminar of the Bundesbank. The views expressed in this paper do not necessarily reflect the views of the Deutsche Bundesbank. Deutsche Bundesbank Deutsche Bundesbank

2 Introduction In the mid-s, prior to the Global Financial Crisis, housing investment, house prices and housing credit in the US rose very strongly. At the same time, risk spreads in the mortage credit market were are very low levels (Figure ). The housing and credit boom turned out to be unsustainable and its reversal is seen as one important trigger of the global financial crisis and the Great Recession. There still exists no consensus on the drivers of the pre-crisis housing and credit boom in the US. One explanation that is provided in the literature is loose monetary policy (Taylor (9a), Taylor (9b), Eickmeier and Hofmann (3), Iacoviello and Neri (), Jarocinski and Smets (8)). Possibly as a consequence of deflationary fears after the burst of the dotcom bubble the Federal Reserve lowered interest rates to levels below rates implied by the Taylor rule, and this fueled the housing and credit boom through various channels. Through the traditional portfolio rebalancing mechanism house prices rise after a monetary policy loosening, as agents substitute from securities to real estate (or other assets) (Bordo and Landon-Lane (3)). Moreover, low interest rates lead to lower cost of and increased demand for (mortgage) credit and housing. At the same time, the lenders risk appetite increases (the risk-taking channel of monetary policy, e.g. Borio and Zhu (8), Rajan (5), Altunbas, Gambacorta and Marques-Ibanez (), Jimenez and Ongena (forthcoming), Buch, Eickmeier and Prieto (3)), risk spreads fall, and the supply of (riskier) credit rises. Consequently, housing demand and house prices would go up as well. A second explanation for the housing and credit boom are regulatory changes and financial innovations in the housing sector. As an attempt to reduce rising income inequality and income stagnation, government administration and Congress already pushed in the 99s for affordable housing for low income families and allowed them to reduce their capital requirements (Bordo and Landon-Lane (3)). This led to riskier mortgages, especially in the subprime segment, which were securitized and bundled into highly rated mortgage backed securities (MBSs). Those MBSs were further repackaged into collateralized debt obligations (CDOs). Credit default swaps (CDSs) provided insurance on many of these products. Financial firms ramped up leverage and avoided regulatory oversight and statutory capital requirements. This led to a credit (supply) boom and raised the demand for assets and investment in housing, therefore house prices. As a third explanation, housing demand shocks, i.e. an increase in housing activity and house prices triggered, for instance, by an exogenous change in housing preferences, are found to be a driver of the boom (Iacoviello and Neri ()). The increase in house prices raises the return on housing investment and, hence, housing investment. Housing Iacoviello and Neri () provide two possible interpretations of a housing preference shock. The first interpretation is a change in the availability of resources needed to purchase housing relative to other goods or other social and institutional changes that shift preferences toward housing. The second interpretation is random changes in the factor mix required to provide home services for a given housing shock.

3 is important as collateral for credit. Therefore, the rise in house price relaxes collateral constraints (Iacoviello and Neri (), Campbell and Cocco (7)), and credit rises as well. Moreover, a permanent increase in housing wealth would lead to an increase in household spending and borrowing when homeowners attempt to smooth consumption over time (Goodhart and Hofmann (8)). Finally, international capital flows are considered to be an important driver of imbalances in housing and credit markets in the s in the US (e.g. Justianiano, Primiceri and Tambalotti (3) and references therein). According to the so called global saving glut hypothesis emerging (mainly Asian) and oil exporting countries accumulated international reserves which were then invested in riskless assets (Treasury securities) in the US (Bordo and Landon-Lane (3), Bernanke (5)). This allowed the US to run a persistent current account deficit and lowered long-term interest rates which fueled the boom. Another explanation (the so called global banking glut hypothesis) is that international (mainly European) banks invested increasingly in asset-backed securities which turned out to be risky in the crisis and became direct competitors of domestic financial institutions (Shin (), Bertaut, DeMarco, Kamin and Tyron ()). With a greater total intermediation the spreads betweeen safe funding rates and the returns on ABS and therefore ultimately mortgages fell. In this paper, we make an attempt to contribute to a better understanding of the drivers of the boom. There is no doubt (anymore) that housing and credit boom and bust cycles can be extremely costly, and therefore understanding what led to the boom is crucial for policy makers. The question whether and/or how central banks should react to asset price and credit booms is a very important one. Prior to the global financial crisis, there has been a consensus that central banks should mop up after the burst of a bubble by lowering interest rates rather than target asset prices or prick a bubble (Mishkin (7), Issing ()). After the crisis, however, the consensus shifted towards a view that central banks should counteract emerging bubbles, either by raising interest rates or using macroprudential tools. This is seen to be important not only to to avoid negative consequences for financial stability and, ultimately, the real economy. Strong increases in asset prices and credit can also lead to inflation, which central banks should avoid. Also, the low interest rate environment has, prior to the crisis and again in the last few years, challenged monetary policy. There are reasons to believe that the monetary transmission mechanism is different in sustained periods of low interest rates compared to normal times, when interest rates are at levels well above zero. For example, the risk-taking channel of monetary policy has found to be particularly effective in periods when interest rates are too low for too long (Altunbas et al. (), Buch et al. (3)), and, hence, the cost of holding interest rates low may be relatively large. Moreover, a low interest rate environment (and the zero lower bound) raises the question whether monetary policy can still credibly keep inflation stable. Also, we now observe that monetary policy

4 needs to complement interest rate policy with unconventional measures, the transmission of which is still quite unknown. Understanding the monetary transmission during the pre-crisis period certainly helps assessing the risks to financial and price stability we are facing today. Finally, if changes in the regulation or financial liberalization in housing and credit markets turn out to be the major drivers of the housing and credit boom, there are lessons to be learned for macroprudential policy. Clearly, it is very important for central banks to closely monitor developments in housing and credit markets. At the same time, building models that are able to account for the drivers and complex mechanisms of the pre-crisis housing and credit boom is not an easy task. Such models need to incorporate variables capturing the housing market, the credit market and monetary policy as well as macroeconomic variables. The variables need to be selected such that the relevant drivers of the boom can be identified and disentangled from each other and other potential influences. Those variables should also be allowed to mutually interact. Furthermore, a model suited to analyze the drivers of the boom should have time-varying parameters able to capture instable shock variances and relationships in the economy and in financial markets. Time variation in the parameters can be motivated in different ways. A large literature has documented the decline in the volatility of many variables between the early-98s and the mid-s (the Great Moderation ). Good luck (i.e. lower shock volatility) and better monetary policy were identified to be important factors which contributed to a smoothening of the business cycle (Stock and Watson (3), Gali and Gambetti (9), Clarida, Gali and Gertler ()). With the global financial crisis, the Great Moderation periods seems to have come to an end (refs). Moreover, regulatory changes, liberalization and innovation in housing and credit markets could have affected the link between housing markets and the financial sector. Finally, the relationship between monetary policy, credit and house prices can change over the business or the financial cycle not only because risk taking may be different in sustained periods of low interest rates, but also, for example, because of collateral constraints may become binding in economic recessions (?, Guerrieri and Iacoviello ()) or because greater information asymmetry between lenders and borrowers in crisis periods can drive up the cost of obtaining external funding (known as the financial accelerator ) (Bernanke, Gertler and Gilchrist (999)). We use a large quarterly VAR with time-varying parameters recently suggested by Koop and Korobilis (3a), which incorporates all these features. Koop and Korobilis (3a) suggest approximations for the parameter and the VAR residual covariance matrices, which allow to estimate a relatively large VAR model with time-varying autoregressive parameters and shock volatilities in a reasonable amount of time. We include nine US macroeconomic, housing and credit market variables. The model is estimated over The beginning of the sample marks the use of the Federal Funds rate as a policy instrument, while the end of the sample marks the binding of the zero lower bound 3

5 and unconventional monetary policy measures used by the Federal Reserve. We identify monetary policy shocks, credit supply shocks and housing demand shocks by imposing theoretically motivated sign restrictions. We focus explicitly on the three domestic boom explanations, and leave explicit accounting for the global saving glut or the global banking glut for future research. We discuss, however, whether or to what extent capital inflows may be reflected in the identified shocks. Based on an impulse response analysis, we assess the time-varying effects of those structural shocks on housing investment, house prices, mortgage credit and lending rates. We also investigate whether the volatility of the shocks has changed over time. The core of our paper is then, however, an assessment of the contributions of the three shocks to the housing and credit boom in the s using a historical decomposition. We make the following two main contributions. First, in a unified framework we consider simulteaneously three possible drivers (monetary policy, credit supply and housing demand), which we identify with sign restrictions, and, at the same time, allow for time variation in the parameters. Most previous papers have considered only one or two of the drivers. Only Goodhart and Hofmann (8) have considered money, credit and house price shocks simultaneously. They use a Panel VAR for industrialized countries, a recursive identification scheme and investigate time variation in the parameters based on a sample split and, alternatively, a dummy variable approach. Second, the large timevarying parameter VAR of Koop and Korobilis (3a) is a new tool. It has, so far, been applied in a forecasting context, but not yet in a structural context. We need to make a few modifications to the approach which we explain in the methodological section. We obtain the following main findings: (i) The housing and credit boom, which started in the early-99s and ended around 6, has initially been driven mainly by credit supply shocks. In the mid-s, however, monetary policy shocks were important drivers. (ii) The effects of monetary policy shocks on the housing and mortgage credit market and the effects of credit supply shocks on the housing market have increased over time. The remainder of the paper is organized as follows. In Section we outline the literature related to our analysis. In Section 3, we present the data we use for the analysis, in Section 4, we outline the large time-varying parameter VAR (TVP-VAR) methodology and explain our identifying assumptions. In Section 5 we present time-varying impulse response functions, document the temporal evolution of the sizes of the shocks and the historical decomposition. In that section, we furthermore discuss extensions and present robustness checks. We conclude in Section 6. Related literature Bordo and Landon-Lane (3) estimate a panel VAR for countries over 9-. Based on a historical decomposition they also analyze the drivers of the house price boom in the US and find that house price shocks explain the bulk. Also the low inflation envi- 4

6 ronment, which is associated with greater MP credibility and emphasis on price stability rather than financial stability and which has held down interest rates contributed to the house price boom, although to a smaller extent. Credit shocks (the authors do not distinguish between demand and supply) and monetary policy shocks, be contrast, were not found to have notably contributed to the boom. The study nicely illustrated that house prices, credit and interest rates already interacted in the past and in many other countries, but just the US. Taylor (9a) and Taylor (9b) have shown that deviations of monetary policy rates from the interest rate implied by the Taylor rule have contributed to the increase in housing starts in the mid-s in the US. Using a quantitative equilibrium model Justianiano et al. (3) find that foreign capital flows explain between /4 and /3 of the increase in US house prices and household debt prior to the crisis. The authors find that the global saving glut larger effects than the banking glut, as the former generates a fall in interest rates for both borrowers and savers, which stimulates housing demand, while the spread compresses in the case of the banking glut. Iacoviello and Neri () consider in their DSGE model featuring a housing sector monetary policy and housing preference shocks (as well as technology shocks). They find that housing preference shocks are the most important drivers, explaining the housing boom over the entire s. Monetary policy also matter, but to a smaller extent and at a later stage of the boom. The authors also investigate the role of collateral constraints based on a counterfactual experiment and find that the wealth effect on consumption increases with the fraction of households who use their home as collateral. Based on a BVAR fitted to housing variables, output and inflation as well as interest rates and the term spread, Jarocinski and Smets (8) find that housing demand and monetary policy shocks explain an important part of the increase in housing investment and house prices in the s in the US. The authors, however, do not include credit in their model. Eickmeier and Hofmann (3) apply a factor-augmented VAR to a large set of variables capturing balance sheet items of the non-financial price sector, asset prices (including house prices) and interest rates and spreads and macro variables in the US. They investigate the role of monetary policy for the housing and credit boom. They find that monetary policy shocks have contributed considerably at a late stage of the housing and credit boom, but were not the trigger, whereas monetary policy reactions to other (financial and macroeconomic) shocks played a notable role at an earlier stage. The authors, however, do not consider explicitly alternative drivers (such as housing demand or credit supply). Using a sample split, they investigate whether the monetary transmission to house prices has changed over time, but could not detect statistically significant changes. 5

7 Goodhart and Hofmann (8) analyze the transmission of monetary, house price and credit shocks in a Panel VAR for 7 industrialized countries over the period 97-6 (hence, the focus is not on the pre-crisis boom). The shocks are identified using a recursive identification scheme. They find that the effect of broad money shocks on house prices and credit has become stronger over compared to the entire sample and attribute this effect of financial system liberalization in the 97s and the early 98s. The also find that the effects of credit and money shocks on house prices are larger during housing booms, although not significantly (similarly to Eickmeier and Hofmann (3)). 3 Data Our VAR model is comprised of nine variables: the logarithm of real GDP, quarter-onquarter differences of the logarithm of the GDP deflator, the Federal Funds rate, the 3-year mortgage rate, the mortgage credit spread (defined as the 3-year mortgage rate minus the 3-year government bond rate), the logarithm of real mortgage credit (defined as nominal mortgage credit divided by the GDP deflator), the logarithm of real housing investment, the logarithm of the real Case-Shiller house price index (defined as the nominal house price divided by the GDP deflator), and the real net capital account (flows) (defined as the net capital account divided by the GDP deflator). All series are seasonally adjusted. They are all taken from the Fred database of the Federal Reserve Bank of St. Louis, except for the house price, which is taken from Robert Shiller s webpage. 4 Methodology 4. Large time-varying parameter VAR The analysis departs from an M-dimensional vector y t for t =,..., T, which includes the M(= 9) macroeconomic and financial variables in levels listed above. We assume that y t follows a time-varying parameter VAR(p) model: with y t = Z t β t + ε t β t+ = β t + u t where ε t is i.i.d. N(, Σ t ) and u t is i.i.d. N(, Q t ). Z t is a M k matrix comprised of a vector of intercepts and p lags of each of the M variables. The standard approach of estimating this model is via Bayesian methods. In a Bayesian setup, the researcher makes some distributional assumptions about the free parameters of the model and uses Marcov-Chain-Monte-Carlo methods to estimate the model (see Koop 6

8 and Korobilis for a textbook level treatment of Bayesian methods for multivariate time-series models). A drawback of approaching the model with Bayesian methods is the high computational costs involved with the posterior simulation of the joint distribution of the parameters. Indeed, so far this Bayesian approach can be applied only to small sized VARs. In order to make the estimation of TVP-VARs with more than a handful of variables computationally feasible, Koop and Korobilis (3b) propose replacing Σ t and Q t with approximations. The essence of the approach of Koop and Korobilis (3b) lies in adapting the original Kalman filter for the states β t N(, V t ): b t t = b t t V t t = V t t + Q b t t = b t t + V t t Z t(σ t + Z t V t t Z t) (Y t Z t β t t ) V t t = V t t V t t Z t(σ t + Z t V t t Z t) (Z t V t t ). with the following approximation b t t = b t t V t t = λ V t t ˆΣ t = ( κ)(y t Z t β t t ) (Y t Z t β t t ) + κˆσ t b t t = b t t + V t t Z t(ˆσ t + Z t V t t Z t) (Y t Z t β t t ) V t t = V t t V t t Z t(ˆσ t + Z t V t t Z t) (Z t V t t ) with λ [, ) and κ (, ) being forgetting factors which need to be set. If the aim is to obtain estimates of the time-varying autoregressive parameters β t and the time-varying variance-covariance of the VAR residuals Σ t the procedure proposed by Koop and Korobilis (3b) completely avoids the need for posterior simulation. Indeed, only a single forward recursion of the adapted Kalman filter is required to obtain these estimates. Koop and Korobilis (3b) use the filtered estimates obtained from the large TVP- VAR to conduct some forecasting exercises. We however aim at a structural interpretation of the model. In order to obtain optimal estimates of the time-varying parameters while taking into account as much information as possible we apply a standard fixed-interval Up to 6 variables in the baseline and 7 variables in the robustness checks have been used so far by us in Prieto, Eickmeier and Marcellino (3). In this paper, we apply, however, a recursive identification scheme. Papers using sign restrictions use up to 4-5 variables (e.g. Gambetti and Musso, Baumeister and Benati 3). While the former paper identifies only shock, the latter identifies 4 shocks simultaneously. 7

9 Kalman smoother using the output from the Kalman filter. Specifically, starting at the last time step we proceed backwards in time using the following recursive equations b t T = b t t + V t t (V t+ t ) (b t+ T b t t ) (4.) V t T = V t t + V t t (V t+ t ) (V t+ T V t t )(V t t (V t+ t ) ). (4.) 4. Inference In the Bayesian approach of estimating TVP-VARs the posterior simulation provides the entire (joint) distribution of the model parameters. Inference about the parameters of the models, and other non-linear function of the model parameters, e.g. impulse response functions, can be based on this joint posterior distribution. The approach described above by contrast only provides the (optimal) smoothed point estimates of the parameters of the TVP-VAR. Our approach to draw valid inference in the above model is based on re-sampling methhods. We propose to use a re-sampling algorithm which is based on a wild bootstrap. In the context of a TVP-VAR with (potentially) non-stationary volatility a wild bootstrap scheme is required, rather than a standard residual based re-sampling scheme often used in the VAR literature (e.g. Christiano, Eichenbaum and Evans (999)). The reason is that different from standard residual re-sampling methods the wild bootstrap can replicate the pattern of heteroscedasticity in the original shocks (see also Goncalves and Kilian 4 and Cavaliere, Rahbek and Taylor ). The following steps constitute the wild bootstrap re-sampling algorithm: Step Generate T bootstrap residual according to the device ε b t = ε t ω t, where ω T tt= denotes a sequence from the Rademacher two-point distribution: P (ω t = ) = P (ω t = ) = / Step Construct bootstrap sample recursively from yt b = Zt b ˆβ t + ε b t, t =... T Step 3 Using the Y b t obtain bootstrap estimates of ˆβ b t t and ˆΣ b t from the large TVP-VAR We also experimented with replacing the Rademacher distribution with the Gaussian distribution and Mammen s two-point distribution given by P (ω t =.5( (5) )) =.5( (5) + )/ (5) := p; P (ω t =.5( (5) + )) = p. The choice of the distribution does however not make any dramatic differences. Finally, note that we require the VAR system to be non-explosive at each point in time. This is achieved by requiring that the roots of the VAR polynomial associated with the time t bootstrap parameters are on or outside the unit circle for all t =... T. If this condition is violated we discard the entire bootstrap sample and go back to Step of the re-sampling algorithm. 8

10 4.3 Sign restrictions We aim at identifying simultaneously a housing demand shock, a credit supply shock and a monetary policy shock imposing sign restrictions on impulse responses in the spirit of Faust (998), Canova and de Nicolo (3) and Uhlig (5). All restrictions are imposed on impact only. The identifying assumptions imposed on impact on the impulse responses are summarized in Table. The restrictions we impose disentangle the three shocks from each other and other structural shocks (such as aggregate supply and demand shocks or a saving glut shock ). We normalize the GDP response to be non-negative after all shocks, i.e. we identify expansionary shocks. After a monetary policy loosening shock, moreover, the Federal Funds rate and the capital account decline and the GDP deflator rise. Moreover, the Federal Funds rate changes by more than the mortgage rate. The restrictions on GDP, the Federal Funds rate and the GDP deflator are standard and in line with many theoretical and empirical models (e.g. Peersman 5, Smets and Wouters 3). The monetary policy shock is disentangled from aggregate supply and demand shocks. After aggregate supply shocks, GDP and the GDP deflator would move in opposite directions, and the policy rate, GDP and the GDP deflator would all rise after an aggregate demand shocks. The restriction on the Federal Funds rate reaction relative to the mortgage rate reaction helps disentangling the monetary policy from credit supply shocks. The underlying assumption is that lending rates move in response to monetary policy shocks in a sticky way, which can be rationalized by the presence of menu costs in loan rate adjustment or relationship banking and which is consistent with empirical evidence (see the discussion in Gerali, Neri, Sessa and Signoretti and for empirical evidence on the stickiness assumption after monetary policy shocks Hofmann and Mizen 4, Eickmeier and Hofmann 3). Finally, the restriction in the capital account balance can be justified as follows. After a monetary policy shock interest rates in the US decline relative to interest rates in other countries, which leads to net capital outflows (and hence, a current account surplus). This restriction disentangles the monetary policy shock from a savings glut shock, after which net capital inflows should increase (i.e. the current account should decline) and interest rates should decline as well. We note that nothing prevents net capital inflows to increase after a monetary policy loosening in the medium run, as the shock can be expected to reduce asset prices and, via negative wealth effects, consumption and, hence, improve the trade balance (e.g. Fratzscher, Saborowski and Straub 9). But these effects should take time to materialize, and we impose restrictions on impact only. After an expansionary housing demand shock, most variables are required to rise (GDP, the GDP deflator, house prices, mortgage credit, housing investment, the Federal Funds rate and the mortgage rate). Hence, housing demand shocks are disentangled from monetary policy shocks by the differential restrictions on the monetary policy rate. Furthermore, by requiring the mortgage rate and credit to rise, we impose that credit demand 9

11 effects (due to increased housing wealth, due to increased capacity to borrow because of higher collateral, due to shifts in borrower s preferences or because the mortgage lending rate follows the Federal Funds rate which increases after the housing demand shock to counteract inflationary pressures) are larger than credit supply effects (due to an improvement of the banks balance sheets as a consequence of higher collateral). This is meant to capture a shock we have observed in the mid-s. Housing demand shocks ressemble, so far, standard aggregate demand shocks. Therefore, we restrict, in addition, housing investment to increase by more than GDP in the short run, which disentangles housing demand from aggregate demand shocks and is in line with Jarocinski and Smets 8. GDP and interest rates move in the same direction after the shock, whereare after an unexpected rise of capital inflows (a saving glut shock), we would expect interest rates to go down, and output to go up. Hence, the housing demand shock is also separated from a saving glut shock. Finally, after a credit supply shock, GDP, mortgage credit, housing investment and the capital account balance are restricted to increase. Credit is required to increase by more than GDP, which disentangles the credit supply shock from aggregate supply or demand shocks. 3 Finally, the mortgage rate declines, and it moves by more in the short run than the government bond yield (i.e. the credit spread declines) and the Federal Funds rate. Hence, the price and the quantity of credit move in opposite directions, which disentangles credit supply from credit demand shocks. After credit demand shocks they would move in the same direction. The restriction on the mortgage rate relative to the government bond yield reflects that the risk premium declines. Moreover, the restriction on the mortgage rate relative to the Federal Funds rate helps disentanging credit supply from monetary policy shocks. Those restrictions to identify the credit supply shock are also imposed and discussed in length in Eickmeier and Ng (). Finally, we would expect riskless long-term interest rates to decline by more than (risky) mortgage rates and, hence, the credit spread to increase after a saving glut shock. Hence, our credit supply shock is also disentangled from such a shock. We note that our credit supply shock likely contains a global banking glut shock, i.e. capital inflows by European banks, which needs to be kept in mind when interpreting the results. We implement the sign restrictions along the lines of the procedure suggested by Rubio- Ramirez, Waggoner and Zha (). Specifically, let Σ t = D t D t be the Cholesky decomposition of the time-varying variance-covariance matrix of the reduced from VAR residuals. Further, let Ω be a M M random matrix drawn from an independent standard normal distribution. The QR decomposition of Ω delivers Ω = QR. The time-varying impact matrix of the structural shocks is then computed as Āt = D tq. If for all t =... T the impulse responses generated by the impact matrix Āt satisfy the sign restrictions we keep 3 This restriction is consistent with structural models with a banking sector (Curdia and Woodford (), Kollmann ()). It is less restrictive than previous papers which have imposed output not to react on impact to credit supply shocks (e.g. Peersman (), Ciccarelli et al. (), Buch et al. (3), Eickmeier and Hofmann (3)).

12 the matrix, otherwise we discard it. We keep drawing from the random matrix Ω until we obtain impact matrices which satisfy all sign restrictions at each point in time simultaneously. We use the Median Target method by Fry and Pagan (7) and Fry and Pagan () to choose among those models the one which provides a good representation of the central tendency. For each point in time we then apply the Median Target approach. Ultimately we use the one rotation matrix which is most frequently selected to compute impulse response functions. It turns out that for 88 out of points in time, the same rotation matrix is selected. Hence, unlike the Bayesian TVP-VAR literature which applies sign restrictions, we require the rotation matrix used to generate the structural impact matrix to be time constant. Hence, we do not induce additional time variation into our structural model due to identification uncertainty over time. 5 Results 5. Time varying impulse responses and shock sizes Time varying shock sizes Figure shows the time-varying shock sizes, which we measure as impact effects of standard deviation monetary policy shocks on the Federal Funds rate, of standard deviation credit supply shocks on mortgage credit and standard deviation housing demand shocks on the real house price. 4 The monetary policy shocks and the credit supply shocks (since the mid-98s) have become smaller over time, consistent with the Great Moderation. The size of the credit supply shock also temporarily increases in the early-98s, the early- 99s and the mid-s, which coincides with major mortgage credit booms (Figure ). The size of the housing demand shock also varies notably over time. It strongly increases in the early 99s, which marks the beginning of the most recent housing boom (Figure ). It is relatively large between then and the end of the 99s, and increases again after a short fall since. Time varying impulse responses In Figure 3 we present the evolution of impulse responses to the three shocks for selected horizons (impact effect and effects after and 5 years) over the sample period. The transmission of monetary policy tightening shocks / credit supply shocks / housing demand shocks are shown in panels (a) / (b) / (c). The shocks are normalized to raise the Federal Funds rate by 5 basis points / mortgage credit by percent / the real house price by percent at each point in time. The normalization allows us to focus on possible changes 4 At the beginning of the sample period, the estimates still depend on the starting values. We therefore show results not from 977, but from 983 onwards.

13 in the transmission of the shocks. We comment on variables capturing the housing and financial variables, which are the focus of this paper, as well as on the capital account. Monetary policy shocks Figure 3(a) shows that there is a substantial amount of time variation in the propagation of monetary policy shocks to the housing and credit market. Over the entire estimation period the effects of monetary policy shocks on house prices and mortgage credit increased over time. The higher sensitivity of mortgage credit and house prices to monetary policy shocks might be the result of the general increase in overall indebtedness of household in the US. Indeed, as Carroll and Dunn (997) argued, precautionary motives make the spending of households with high debt levels more sensitive to uncertainty about income than households with less debt. Highly indebted households are therefore more likely to pull back their spending in the face of an adverse shock. This interpretation is also supported by the reaction of the mortgage rate which drops on impact. This, together with the reduction in aggregate mortgage credit implies that negative mortgage demand effects after monetary policy shocks have become stronger over the last three decades. 5 Interestingly, the stronger reaction of house prices to monetary policy shocks was not a gradual phenomenon. Instead, there are two abrupt changes in the reaction of house prices on impact. Incidentally, the timing of the changes coincide with the 99/99 recession and the recession. In the period between these recessions and during the 98s this impact effect is approximately constant. However, starting with the recession the effect increases steadily. A similar discrete change at the beginning of the new millennium is a also present in the reaction of the mortgage rate. The increase in the capital account on impact (which is implied by the sign restrictions) becomes larger over time with the effect accelerating in the late 99s. Note however that the positive impact effect is only temporary. After one year the sign of the impulse response changes and we observe capital outflows. This pattern in the response of capital flows to monetary policy shocks is consistent with the results presented by Fratzscher, Saborowski and Straub (9) and Bruno and Shin (3). The tightening of monetary policy, associated with an increase in the risk free short term interest rate first leads to capital inflows due an increase in the interest rate differentials. After some time the short term interest rate returns to its initial value. The effects of the initial monetary policy shocks on the housing and credit market are however more persistent with negative wealth effects and negative credit supply effects leading to capital outflows. Note that Bruno and Shin (3) attribute the delayed capital outflow to a deleveraging of financial institutions after the monetary policy shock. Note that our findings are consistent with this interpretation: after approximately one year the mortgage rate becomes positive while mortgage credit remains negative, indicating that negative credit supply effect seem to become dominating. The negative credit supply effects in the medium run uncovered 5 Eickmeier and Hofmann (3), using a sample split approach, also find stronger effects of monetary policy shocks on house prices in more recent periods.

14 in our analysis might be driven by the deleveraging process uncovered by Bruno and Shin (3). Credit supply shocks Figure 3(b) uncovers some interesting patterns in the response of the credit and housing market and the capital account over time. The impact effects of credit supply shocks on house prices and housing investment have risen over time. By contrast, the response inflation and GDP to credit supply shocks has remained approximately constant over time. A tentative interpretation of this finding is that changing preferences for housing related consumption increased the sensitivity of the housing market to shocks related to the availability of housing credit. We want to emphasize that although the sign restrictions imply that housing investment increase following the credit supply shock the reaction of house prices is left unconstrained. In this respect we leave it open to the data to determine whether credit supply shocks lead to housing demand or supply effects. Apparently, credit supply shocks induced housing demand effects. Interestingly, even though these housing demand effects have become stronger over time the response of the mortgage rate and the risk premium has become smaller. This effect might be due to the deepening of the mortgage market in the last decades. The positive credit supply shock also induces net capital inflows, which might indeed reflect that a global banking glut is embedded in the credit supply shock. The effect of the credit supply shock on mortgage credit in the medium run has also become stronger over time. This is accompanied by a stronger medium run house price response over time. This suggests that with the a stronger positive reaction of house prices over time the collateral value increases which allows households to increase debt. Consistent with these patterns medium run credit inflows have also become stronger over time. Housing demand shocks Turning to Figure 3(c), the effects of housing demand shocks on most variables (including residential investment) have declined over time. Interestingly, the risk premium declines temporarily, but this effect becomes smaller and ultimately disappears over time. One explanation could be that the riskless government bond yield closely follows the Federal Funds rate. The effects of the shock on the latter interest rate have also strongly declined over time, while the effects on the mortgage rate are fairly stable. Another interpretation is that the housing demand shock has positive effects on the housing market and the rest of the US economy, which implies a decline in general risk. In the short run, those effects seem to have dominated balance sheet or risk-taking effects as a consequence of increases in house prices (Buch, Eickmeier and Prieto 3) or the policy rate. Over time, the latter effects seem to have become more important, and at the one-year horizon they seem to dominate over the sample period. 3

15 5. Historical decomposition In Figure 4, we present the historical decomposition for the house price, residential investment and housing credit since 993Q. The bars reflect contributions by the three identified shocks, the black line represents the contributions of the sum of all (identified and unidentified) shocks. The housing and credit boom which started in the late 99s was initially mainly driven by positive credit supply shocks, in line with Bordo and Landon-Lane (3), and - to a much smaller extent - housing demand shocks. Monetary policy shocks did not contribute or contributed even negatively. All three identified shocks temporarily made negative contributions to the evolution of house prices and residential investment after the burst of the dotcom bubble. Since 4, all three shocks contributed significantly to the housing boom. Monetary policy shocks were - by far - the main drivers out of the three shocks, explaining about /4-/3 of the housing boom, followed by credit supply and housing demand shocks. The latter finding is in line with Taylor (9a) and Taylor (9b). 5.3 Robustness and extensions 6 Conclusions... 4

16 References Altunbas, Y., Gambacorta, L. and Marques-Ibanez, D. (), Does monetary policy affect bank risk-taking?, ECB Working Paper 66. Baumeister, C. and Benati (3), Unconventional monetary policy and the great recession: Estimating the macroeconomic effects of a spread compression at the zero lower bound, International Journal of Central Banking 9(), 65. Bernanke, B. (5), The global savings glut and the us current account deficit, Remarks by Governor Ben S. Bernanke at the Sandridge Lecture, Virginia Association of Economics, Richmond, Vorginia, The Federal Reserve Board of Governors. Bernanke, B., Gertler, M. and Gilchrist, S. (999), The financial accelerator in a quantitative business cycle framework, in John Taylor and Michael Woodford (eds.), the Handbook and Macroeconomics, Amsterdam: North Holland. Bertaut, C., DeMarco, L., Kamin, S. and Tyron, R. (), Abs inflows to the united states and the global financial crisis, International Finance Discussion Papers 8. Bordo, M. and Landon-Lane, J. (3), What explains house price booms? history and empirical evidence, NBER Working Paper Borio, C. and Zhu, H. (8), Capital regulation, risk taking and monetary policy: a missing link in the transmission mechanism?, BIS Working Paper 68. Bruno, V. and Shin, H. S. (3), Capital flows and the risk-taking channel of monetary policy, Working Paper 894, National Bureau of Economic Research. Buch, C., Eickmeier, S. and Prieto, E. (3), In search for yield? survey-based evidence on bank risk taking, Journal of Economic Dynamics and Control forthcoming. Campbell, J. and Cocco, J. (7), How do house prices affect consumption? evidence from micro data, Journal of Monetary Economics 54(3), Canova, F. and de Nicolo, G. (3), On the Sources of Business Cycles in the G-7, Journal of International Economics 59(), 77. Cavaliere, G., Rahbek, A. and Taylor, A. R. (), Testing for co-integration in vector autoregressions with non-stationary volatility, Journal of Econometrics 58(), 7 4. Christiano, L. J., Eichenbaum, M. and Evans, C. L. (999), Monetary policy shocks: What have we learned and to what end?, in J. B. Taylor and M. Woodford, eds, Handbook of Macroeconomics, Vol. of Handbook of Macroeconomics, Elsevier, chapter, pp Clarida, R., Gali, J. and Gertler (), Monetary policy rules and macroeconomic stability: Evidence and some theory, The Quarterly Journal of Economics 5(),

17 Eickmeier, S. and Hofmann, B. (3), Monetary policy, housing booms and financial (im)balances, Macroeconomic Dynamics 7, Eickmeier, S. and Ng, T. (), How do credit supply shocks propagate internationally? A GVAR approach, CEPR Discussion Paper 87. Faust, J. (998), The robustness of identified VAR conclusions about money, Carnegie- Rochester Conference Series on Public Policy 49(), Fratzscher, M., Saborowski, C. and Straub, R. (9), Monetary policy shocks and portfolio choice, ECB Working Paper. Fry, R. and Pagan, A. (7), Some issues in using sign restrictions for identifying structural VARs, NCER Working Paper 4. Fry, R. and Pagan, A. (), Sign restrictions in structural vector autoregressions: a critical review, Journal of Economic Literature 49(4), Gali, J. and Gambetti (9), On the sources of the great moderation, American Economic Journal: Macroeconomics (), Gambetti, L. and Musso, A. (), Loan supply shocks and the business cycle, ECB Working Paper 469. Gerali, A., Neri, S., Sessa, L. and Signoretti, F. (), Credit and banking in a DSGE model of the euro area, Journal of Money, Credit and Banking 4(6), 8 4. Goncalves, S. and Kilian, L. (4), Bootstrapping autoregressions with conditional heteroskedasticity of unknown form, Journal of Econometrics 3(), 89. Goodhart, C. and Hofmann, B. (8), House prices, money, credit and the macroeconomy, Oxford Review of Economic Policy 4(), 8 5. Guerrieri, L. and Iacoviello, M. (), Collateral constraints and macroeconomic asymmetries, Mimeo, Federal Reserve Board. Hofmann, B. and Mizen, P. (4), Interest rate pass through and monetary transmission: Evidence from individual financial institutions retail rates, Economica 7, Iacoviello, M. and Neri, S. (), Housing market spillovers: evidence from an estimated dsge model, American Economic Journal: Macroeconomics, Issing, O. (), Lessons for monetary policy: What should the consensus be?, IMF Working Paper WP//97. Jarocinski, M. and Smets, F. (8), House prices and the stance of monetary policy, Federal Reserve Bank of St. Louis Review 9(4),

18 Jimenez, G. and Ongena, S., P.-J. L. S. J. (forthcoming), Hazardous times for monetary policy: What do twenty-three million banks say about the effects of monetary policy on credit risk-taking?, Econometrica. Justianiano, A., Primiceri, G. and Tambalotti, A. (3), The effects of the saving and banking glut on the u.s. economy, NBER Working Paper Koop, G. and Korobilis, D. (), Bayesian multivariate time series methods for empirical macroeconomics, Foundations and Trends(R) in Econometrics 3(4), Koop, G. and Korobilis, D. (3a), Large time-varying parameter vars, Journal of Econometrics 77(), Koop, G. and Korobilis, D. (3b), Large time-varying parameter vars, Journal of Econometrics 77(), Mishkin, F. (7), Housing and the monetary transmission mechanism, Federal Reserve Bank of Kansas City, Housing, Housing Finance, and Monetary Policy, 7 Jackson Hole Symposium (Federal Reserve Bank of Kansas City, Kansas City, 7) pp Peersman, G. (5), What caused the early millenium slowdown? Evidence based on vector autoregressions, Journal of Applied Econometrics, Prieto, E., Eickmeier, S. and Marcellino, M. (3), Time variation in macro-financial linkages, Deutsche Bundesbank Discussion Paper 3/3. Rajan, R. (5), Has financial development made the world riskier?, NBER Working Paper 78. Rubio-Ramirez, J. F., Waggoner, D. F. and Zha, T. (), Structural Vector Autoregressions: Theory of Identification and Algorithms for Inference, Review of Economic Studies 77(), Shin, H. (), Global banking glut and loan risk premium, IMF Economic Review 6(), Smets, F. and Wouters, R. (3), An estimated dynamics stochastic general equilibrium model of the euro area, Journal of the European Economic Association (5), Stock, J. and Watson, M. (3), Has the business cycle changed? evidence and explanations, FRB Kansas City symposium, Jackson Hole, Wyoming, August 8-3, 3. Taylor, J. B. (9a), The financial crisis and the policy response: an empirical analysis of what went wrong, NBER Working Paper

19 Taylor, J. B. (9b), Housing and monetary policy, In: Housing, housing finance, and monetary policy. Proceedings of the Federal Reserve Bank of Kansas City Symposium in Jackson Hole, Wyoming pp Uhlig, H. (5), What are the Effects of Monetary Policy on Output? Results from an Agnostic Identification Procedure, Journal of Monetary Economics 5(),

20 7 Tables and Figures Figure : Time-varying shock size x -3 MP shock x -3 CS shock x -3 HD shock

21 impact four quarters Figure : Time-varying shock transmission - Monetary Policy Shock GDP Inflation House Prices Res. Investment Mortg. Credit Mortg. Rate FFR Risk Premium Capital Account quarters

22 impact.. GDP Figure 3: Time-varying shock transmission - Credit Supply Shock.6.4. Inflation House Prices Res. Investment Mortg. Credit Mortg. Rate FFR Risk Premium Capital Account four quarters quarters

23 impact.5 GDP Figure 4: Time-varying shock transmission - Housing Demand Shock.5..5 Inflation House Prices Res. Investment Mortg. Credit Mortg. Rate FFR Risk Premium Capital Account four quarters quarters

24 Figure 5: Historical Decomposition - House Price House Prices Figure 6: Historical Decomposition - Mortgage Credit Mortg. Credit Monetary Policy Credit Supp