BIS Working Papers
The macroeconomic effects of macroprudential policy
by Björn Richter, Moritz Schularick and Ilhyock Shim
Monetary and Economic Department
JEL classification: E58, G28
Keywords: macroprudential policy, loan-to-value ratios, local projections, narrative approach
This publication is available on the BIS website (www.bis.org).
© Bank for International Settlements 2017. All rights reserved. Brief excerpts may be reproduced or translated provided the source is stated.
ISSN 1020-0959 (print)
ISSN 1682-7678 (online)
The macroeconomic effects of macroprudential policy
Moritz Schularick and Ilhyock Shim
Central banks increasingly rely on macroprudential measures to manage the financial cycle. However, the effects of such policies on the core objectives of monetary policy to stabilise output and inflation are largely unknown. In this paper we quantify the effects of changes in maximum loan-to-value (LTV) ratios on output and inflation. We rely on a narrative identification approach based on detailed reading of policy-makers’ objectives when implementing the measures. We find that over a four year horizon, a 10 percentage point decrease in the maximum LTV ratio leads to a 1.1% reduction in output. As a rule of thumb, the impact of a 10 percentage point LTV tightening can be viewed as roughly comparable to that of a 25 basis point increase in the policy rate. However, the effects are imprecisely estimated and the effect is only present in emerging market economies. We also find that tightening LTV limits has larger economic effects than loosening them. At the same time, we show that changes in maximum LTV ratios have substantial effects on credit and house price growth. Using inverse propensity weights to rerandomise LTV actions, we show that these effects are likely causal.
Keywords: macroprudential policy, loan-to-value ratios, local projections, narrative approach.
JEL classification codes: E58, G28.
How do macroprudential policies interact with the core objectives of monetary policy to stabilise prices and output? As a response to the global financial crisis, central banks and regulators across the world have increasingly relied on macroprudential policies to maintain financial stability. A recent literature has shown that policy-makers can moderate credit and asset price cycles using macroprudential instruments (e.g. Akinci and Olmstead-Rumsey (2018), Bruno et al. (2017) and Kuttner and Shim (2016)). This way, they can reduce negative output tail risks emanating from the link between excess credit and costly financial crises (Schularick and Taylor (2012) and Jorda et al. (2013)). However, there is very little empirical evidence on how the use of such instruments affects the traditional objectives of monetary policy, that is, output and inflation.
In this paper, we explore the effects of macroprudential measures on output and inflation for a large cross-country panel of 56 countries over more than two decades. Building on a new dataset, we use a narrative approach in order to address identification challenges. Our results show that macroprudential measures, specifically changes in the maximum loan-to- value (LTV) ratio, do have modest and imprecisely estimated spillover effects on output and inflation. In particular, we find that a 10 percentage point reduction in the maximum LTV ratio lowers output by about 1.1% after four years. However, this effect is rather imprecisely estimated. The near-term impact on the price level is in most specifications even smaller and close to zero. The effect is more pronounced in emerging market economies (EMEs), while the path of output is almost unaffected by LTV limit changes in the set of advanced economies (AEs). In a back-of-the-envelope calculation, we compare the magnitude of this effect to estimates of GDP responses to monetary policy in Jorda et al. (2017), which are also based on a broad sample of countries. We find that the two-year response of GDP to a 10 percentage point reduction in the maximum LTV ratio can be compared to that of a 26 basis points increase in interest rates.
Importantly we also test for potential asymmetries and find that the output effect is mainly driven by the negative effects of tightenings in maximum LTV ratios, and not by higher output generated through loosening actions. We also assess the treatment effects of LTV limit tightenings on financial variables using inverse propensity weighting and find that credit and house prices fall after a tightening. Overall, these results imply that central banks might be able to use macroprudential policies to dampen the financial cycle without risking major interference with their core monetary policy objectives.
Macroprudential measures are not randomly assigned. In the ideal environment to measure the effects of changes in LTV limits on output and inflation, the following three conditions are satisfied: (i) LTV policy actions are exogenous with respect to current and lagged real variables; (ii) such actions are uncorrelated with other shocks (e.g., monetary policy acting at the same time); and (iii) they are unexpected. While the unsystematic nature of macroprudential policies means they are typically unexpected, we clearly need to worry about the exogeneity of the policy action. To address the exogeneity condition, we rely in this paper on a novel hand-collected dataset detailing the intentions or stated objectives of policy-makers when they change LTV limits. This approach is in the tradition of the narrative approach pioneered in Friedman and Schwartz (1963) and Romer and Romer (1989). In a similar spirit to the narrative identification of monetary policy shocks, we argue that macroprudential actions taken without reference to current or expected trends in real output and inflation can be seen as exogenous with respect to price and output stabilisation objectives of monetary policy. This new narrative measure therefore allows us to establish the causal effects of macroprudential actions on economic activity and inflation. Using a battery of tests, we confirm that there is indeed no systematic relationship between changes in our narrative measure and real economic variables. To address the second condition, we control for monetary policy shocks in all our specifications. To trace the dynamic propagation of such exogenous policy interventions, we rely on local projections (Jorda (2005)).
Our identification approach requires detailed reading and understanding of the underlying motivations for macroprudential measures and the information set of policy-makers. To keep the required information manageable, we focused on one specific tool that is frequently used to tackle boom-bust cycles in credit and housing markets, namely, changes in maximum LTV ratios. We compiled a comprehensive new dataset consisting of quarterly observations of LTV actions in 56 economies, building on the database developed by Shim et al. (2013). Our quarterly dataset contains 92 changes in maximum LTV ratios and loan prohibitions.
Almost all papers in the literature on measuring the effectiveness of macroprudential actions employ dummy variables to measure macroprudential policies. Such variables however do not capture the intensity of policy actions of the same type. For instance, a decrease in the maximum LTV ratio by 10 percentage points and a decrease in the ratio by 20 percentage points are treated equally. We use instead a numerical variable quantifying the quarterly changes in the maximum LTV ratio. To our knowledge, this paper is the first that constructs an intensity-adjusted LTV change variable, which considers not only the change in the maximum LTV ratio in percentage points, but also accounts for the scope of loans to which such a change is applied. This allows us to estimate the effect of a one percentage point change in LTV limits as described above. We also assess the differences between tightenings and loosenings and find that the negative long-run effects are driven mainly by tightenings.
Do LTV limits help dampen the financial cycle? To answer this question, we turn to inverse propensity weighting to mimic random allocation. As in Jorda et al. (2015), we employ a two-stage procedure, where the probability of an economy being treated with a macroprudential action, here a tightening in maximum LTV ratios, is estimated in a separate first stage regression. This purges the data of observable sources of endogeneity. In the second stage regression, observations are weighted inversely to the estimated probability of receiving treatment, thus giving greater weight to an action that comes closer to the random allocation ideal. We find indeed that real household credit and real mortgage credit are reduced when LTV limits are decreased. At the same time, house prices fall. Macroprudential policies seem to achieve the desired targets. Our results on real variables indicate furthermore that they do so at a relatively small cost. Therefore, for central banks such macroprudential measures may serve as a complementary policy tool that does not interfere with other objectives in a major way.
This paper is organised as follows. Section 2 provides a literature review. Section 3 describes data and empirical strategies focusing on how we identify the stated objectives of LTV measures, how we construct the intensity-adjusted quantifiable LTV change variable, and whether LTV measures are exogenous to the real cycle. Section 4 presents the empirical results on the response of real variables to changes in LTV limits. Section 5 considers the response of financial variables to LTV actions. Finally, section 6 concludes.
A large amount of literature has considered the effects of monetary policy on output and price levels. For an overview of the literature on this relationship see Ramey (2016). While the global financial crisis has renewed interest in the ability of macroprudential policies to help manage the financial cycle, the responses of output and inflation to macroprudential policies have rarely been addressed. The few investigations into the effects of these policies often use historical data on credit controls that were common in many Western European countries in the decades following World War II (Kelber and Monnet (2014)).
As one of the few studies, Aikman et al. (2016) evaluate in a joint framework the impact of monetary and macroprudential policies in the United Kingdom from the 1950s to the early 1980s. They rely on local projection methods to estimate impulse response functions to the two policy shocks augmented by forecasts (in line with Romer and Romer (1989)) and factors (in line with Bernanke et al. (2005)). They find that an analogue to macroprudential measures, credit controls, were quite effective in taming the credit cycle and had a dampening effect on industrial output. Furthermore, they find little evidence for an effect of credit controls on the price level. Monnet (2014) studies the effect of quantitative controls on money and credit during France's golden age between 1948 and 1973 and finds strong effects on output and prices. Similar to our approach, identification builds on narrative evidence on the intentions associated with enacted policies extracted from archival records.
There are also a few contributions building on recently collected cross-country databases on macroprudential policy actions. Kim and Mehrotra (2017) analyse the responses of credit, output and inflation to changes in macroprudential and monetary policies based on data for four countries in the Asia-Pacific region using a VAR framework. They find a negative effect of changes in macroprudential policies on output as well as inflation. Their macroprudential policy measures are based on data from the Shim et al. (2013) database. Also based on this database, Boar et al. (2017) analyse the relationship between a country's propensity to use macroprudential measures and output outcomes. Dividing countries into two groups depending on their use of macroprudential policies, they find countries which use macroprudential policies more frequently, experience higher growth rates in the cross-section, while the use of macroprudential policies reduces output volatility. Finally, Sanchez and Rohn (2016) analyse the effects of various policies on economic growth using quantile regressions. They find that macroprudential policies reduce output growth, but also reduce the tail risk of output growth. Among other policy variables, they specifically analyse the role of LTV policies (based on a dummy variable) and find consistently negative effects of these measures on output growth. However, the significance of this result depends on the quantiles analysed.
More generally, many studies have examined the impact of macroprudential policies on the financial cycle, particularly on measures of credit and house price cycles. For example, Kuttner and Shim (2016) find that introductions or reductions of the maximum debt-service-to-income (DSTI) ratio and increases in housing-related taxes have significant negative effects on real housing credit growth for 55 economies over the period of 1980 to 2012. They find that a typical DSTI tightening action lowers the real credit growth rate by 4-6 percentage points over four quarters. Using data on total credit to households and non-profit institutions serving households (NPISHs) from the BIS total credit database, Cerutti et al. (2015) find that borrower-based measures such as the maximum LTV ratio and the maximum DSTI ratio are associated with lower growth in credit to households over 2001-13. Dell'Ariccia et al. (2012) find that macroprudential policies can contain the incidence of credit booms and limit the costs of busts associated with credit booms.
Claessens et al. (2013) investigate how macroprudential policies affect individual banks and find that maximum LTV and DSTI ratios reduce asset and leverage growth. Policies implemented in adverse times, however, do not help to stop declines. In a single country study, Wong et al. (2011) find that higher LTV caps lead to a lower level of the mortgage debt-to-GDP ratio in Hong Kong SAR in the 1990s and 2000s. Tillmann (2015) considers the impact of LTV and DSTI limits on household credit in Korea from 2000Q1 to 2012Q4. In particular, he uses a qualitative VAR method to estimate impulse response functions for macroprudential shocks. He found an unexpected tightening in LTV and DSTI limits had a significant effect on household credit growth in Korea.
Using a dynamic stochastic general equilibrium (DSGE) model with housing and household debt, Alpanda and Zubairy (2017) consider the effectiveness of monetary policy, LTV limits and housing-related tax policies on reducing household indebtedness. They find that reductions in mortgage interest payment deductions and regulatory LTV ratios are the most effective tools to limit household credit, as these measures are the most targeted. Rubio and Yao (2017) show in a DSGE model that a macroprudential authority can act as a complementary macro-financial stabiliser for both real and financial cycles when the steady-state interest rate is low and monetary policy hits the zero lower bound.
While all these papers study changes in regulatory caps on LTV ratios, Bachmann and Rueth (2017) analyse the effects of changes in average LTV ratios on output and credit in the United States. They use a structural VAR framework to identify exogenous variation in LTV ratios and find that a 25 basis points tightening in the LTV ratio reduces GDP by approximately 0.1%. At the same time, they find that the Federal Reserve responded to a tightening in LTV ratios with lower policy rates. As a result, mortgage rates fall and residential investment increases after a tightening in LTV ratios.
The narrative approach focuses on evidence derived from the historical record (Romer and Romer (1989)). More specifically, researchers conduct narrative analysis by systemically using qualitative information from contemporary primary sources to construct numerical measures often with the aim of addressing issues of causation. Narrative analysis has been used not only in the context of monetary policy (Friedman and Schwartz (1963) and Romer and Romer (1989)) but also for fiscal policy (Romer and Romer (2007) and Gillitzer (2017)) and financial distress episodes (Romer and Romer (2017)). By contrast, the narrative approach has not been used in the context of macroprudential policy. Budnik and Kleibl (2018) describe a database on policy actions of a macroprudential nature taken by the European Union member countries to affect the banking sector in 1995-2014. The information in the database is based on responses from a survey with the aim of using narrative information for the impact assessment of macroprudential policies.
3. Data and empirical strategy
We base our analysis on quarterly data for 56 economies, including both AEs and EMEs, from 1990Q1 to 2012Q2.The data used in this paper rely on various sources such as the BIS Databank (national sources) and the database on housing market policy actions from Shim et al. (2013). As our dependent variables we use output (real GDP) and the level of the consumer price index from the BIS Databank. For credit variables, we use data from the BIS Databank on bank credit to the private non-financial sector, bank credit to households and housing credit. Explanatory variables include policy variables, other macroeconomic variables, asset prices and structural variables. For policy rates, we use actual policy rates, backdated with one-month or three-month market interest rates obtained from the BIS Databank. For the US policy rate, we use the federal funds rate obtained from Bloomberg.
A major contribution lies in the construction of additional data on macroprudential policy actions: we extend the database for policy actions on housing markets constructed by Shim et al. (2013). While the literature has so far analysed macroprudential measures using dummy variables, we collected data to quantify policy actions. Our main focus is on changes in maximum LTV ratios. This instrument is comparable across countries and the size of a change can be identified in most instances. Because we want to measure the effects of macroprudential policies on output and inflation, we focus on changes in LTV limits that target the financial cycle without being driven by concerns about growth or inflation.
3.1 Narrative identification of macroprudential policy shocks
The greatest challenge in measuring the causal effects of macroprudential policy consists in constructing a measure of macroprudential policy shocks. The following three criteria must be fulfilled: (i) policy actions are exogenous with respect to the current and lagged real variables; (ii) actions are uncorrelated with other shocks; and (iii) they are unexpected. To address the exogeneity condition, we rely in this paper on a novel hand-collected dataset documenting the stated objectives of policy-makers when they change LTV limits.
For the narrative identification, we proceeded in the following way. We first listed all 92 LTV policy actions documented in the database of Shim et al. (2013) from 1990Q1 to 2012Q2 for the 56 economies. Such actions consist of the introduction, tightening, loosening or abolition of the maximum LTV ratio and the prohibition of certain types of loan (that is, applying a zero LTV ratio). We then consulted official documents for each of these policy actions such as press releases announcing these actions, annual reports describing the background of specific policy actions taken and regulatory documents such as circulars to understand the reasoning behind those actions and identify objectives for the implemented LTV actions. We then classified the motivations broadly into real and financial objectives. More specifically, we classified real objectives into the following three categories: GDP, inflation and other real objectives. We also classified financial objectives into the following seven categories: house price, total credit, housing and household credit, bank buffer, risk taking, FX borrowing (including borrowing from non-residents), and other financial objectives. Table 1 provides a summary of the stated objectives of the 92 LTV actions taken among the 56 economies from 1990Q1 to 2012Q2. It should be noted that one LTV action can have more than one stated objective, hence the total number of stated objectives is greater than the total number of LTV actions.
In the next step, we dropped all policy actions that were primarily motivated by real objectives. Yet since a policy action can have more than one stated objective, we went through all actions motivated by financial objectives and verified that at the same time the authorities did not voice concerns over real imbalances in the economy. Among the 92 LTV actions, three actions had a stated real objective, hence we excluded them from the sample. The resulting sample consists of 89 LTV actions accompanied by stated financial objectives only.
3.2 Intensity adjustment
So far, papers in the literature on measuring the effects of macroprudential policy in cross-country data use dummy variables taking the value 1 for tightening actions and zero otherwise, or those taking value 1 for tightening actions, —1 for loosening actions and zero for no change. Such variables do not capture the intensity of policy actions of the same type. By definition, in such a research design a decrease in the maximum LTV ratio by 10 percentage points and a decrease in the ratio by 20 percentage points are treated equally. The coefficients on these dummy variables show the impact of a "typical" policy action in a certain type. That is because these coefficients only show the average impact of all policy actions with different magnitudes of actual changes.
To measure the economic magnitude of the impact of a certain type of macroprudential policy on target variables such as output and inflation, we need to construct a variable measuring the size of policy changes. This is especially important for policy-makers when they try to calibrate the size of the change in regulatory ratios to achieve a certain amount of slowdown in the growth rate of output, credit or asset prices. It should be noted that when we construct such a variable, we need to consider both the size of the change in the relevant ratio and the scope of such policy actions being applied, that is, the range of loans to which the change in the maximum LTV ratio applies.
Specifically, the following criteria are applied to construct the intensity-adjusted LTV action variable. We denote this variable by AMaPP;,t.
• When the maximum LTV ratio is lowered by 10 percentage points, the LTV variable takes a value of 10. When the maximum LTV ratio is raised by 10 percentage points, the LTV variable takes a value of -10. This is in line with the aforementioned
convention of assigning 1 to tightening actions and —1 to loosening actions for dummy variables.
Among the 89 LTV actions without stated real objectives, two LTV actions were taken in one quarter by the same country in the case of four country-quarter observations (i.e., 2010Q2, 2010Q3 and 2011Q1 for China, and 1996Q2 for Singapore). When we use the intensity-adjusted quantifiable LTV change variable, we sum up all LTV actions taken in a quarter. As a result, we have 85 distinct LTV actions in the sample. Among the 85 LTV
actions, 32 actions involve either introductions or abolitions of the maximum LTV ratio or loan prohibition for which we do not have information about the actual size of the change in the maximum LTV ratio. For our baseline regressions, we do not consider these 32 actions with insufficient information, but only consider the remaining 53 quantified actions. Table 2 provides the list of the 53 actions that were not taken with reference to the real cycle and for which we have accurate information on the size of the change.
For robustness check, we also construct a quantified LTV action variable for which we do not adjust for the scope of loans to which the changes in LTV limits apply. Table 3 shows the summary statistics of the 53 LTV variables with and without scope adjustment. The average size of the scope-adjusted LTV change variable is 7.1 percentage points for the tightening actions and -7.9 percentage points for the loosening actions. When we do not adjust for the scope of LTV changes applied, the average size of the change in the maximum LTV ratio is 22.5 percentage points for the tightening actions and -16.0 percentage points for the loosening actions. When we compare the scope-adjusted and scope-unadjusted LTV change variables, scope adjustment reduces the average size of the LTV change variable into one third to one half of the scope-unadjusted LTV change variable. The large difference between the two types of the quantifiable LTV variable is also due to loan prohibition actions which tend to take very large values of the scope-unadjusted numerical LTV change variable (such as +90 or -80).
As a robustness check based on the intensity-adjusted measure AMaPPi>t, we also define an additional measure. AMaPP-ndex is an index variable based on our intensity adjusted policy action variable:
Similarly, A Tlght MaPPi,t and ALooseMaPPi,t are dummy variables for tightenings and loosenings, respectively:
3.3 Are LTV changes exogenous to the real cycle?
In the previous section, we presented the new narratively identified data on the stated objectives and size of LTV changes. In our empirical analysis we combine both to estimate the responses of real variables to a one percentage point change in LTV limits. As described before, focusing on actions referencing only the financial objectives allows a causal interpretation of the results.
A relevant concern in this setting is that policy-makers may target real objectives, without stating them explicitly when implementing macroprudential actions. In this section, we test this important prerequisite for a causal interpretation of our results.
We first determine cyclical deviations of the real variables. To do so, we calculate the deviations of real GDP and prices from trend. To detrend the data we follow Hamilton (2017). The procedure is based on the idea that the trend component of a variable at time t + h is the value we could have predicted based on historical data. The cyclical component will be the difference between the realised value and this trend.
Let h denote the horizon for which we build such a prediction. Then the cyclical (detrended) component is the difference between the realised value at time t + h and the expectation about this value formed at time t. To build this expectation, Hamilton (2017) proposes a regression of the value y at time t + h on the four most recent values of y at time t, i.e. yt, yt-1, yt-2 and yt-3. Formally, this regression can be written as
The choice of h depends on the horizon we attribute to the cyclical component. As suggested, we choose a horizon of eight quarters, so the residual is the deviation of the realised value yt+8 from the expectation formed at time t based on information on yt, yt-1, yt-2 and yt-3. We normalise this variable by its country-specific standard deviation.
Figure 1 shows scatterplots for the size of our LTV changes and detrended real GDP as well as the detrended price level. The data show no clear pattern that could be interpreted as an indication that LTV changes are a reaction to the output gap or price level gap. In particular, changes in the maximum LTV ratio implemented when the output gap (positive and negative) was larger than two standard deviations, do not stand out as large LTV changes. We obtain similar results when we use standard H-P filters and when we use longer lags of detrended real variables.
In Table 4 and Table 5 we turn to formal procedures to test the relationship between the treatment (that is, the implementation of macroprudential actions involving LTV limits) and real economic variables. Note that in an ideal randomised allocation of treatment and control, there would be no difference between treatment and control sub-populations.
In Table 4, we differentiate between two treatments, a tightening and a loosening, and the non-treated control group of observations. We then compare the means of those sub-populations and test for their equality. In the upper panel we compare the group of tightening observations to the control group. We compare real GDP and the price level in treated and non-treated observations based on two measures. First, we compute the smoothed growth rate of these variables over the previous year, which is over four lags, and demean this measure at the country level to account for the fact that fast growing EMEs have historically been more active in using macroprudential policies. The results show no significant difference between the two sub-populations. We also compare the lag of the
output gap and the detrended price level between the two sub-populations, but do not find a significant difference.
In the bottom panel of Table 4, we present the results of tests for the equality of means between loosening actions and the control group. Again, we do not find a statistically significant difference between the two sub-populations.
In Table 5 we test whether we can predict either AMaPPij or AMaPPfndex based on the one-period lagged detrended GDP and the price level. Column (1) shows the results of a regression of the LTV change on the lagged output gap, the detrended price level and a constant. In column (2) we additionally include country-fixed effects. Columns (3) and (4) use AMaPPiI,tndex as the dependent variable instead. Reassuringly, the coefficients are across all specifications insignificant and the variables have little explanatory power. Based on the results in this section, we conclude that LTV changes are not predicted by real economic variables and that their implementation can be seen as orthogonal to the real cycle.
Another concern may be that macroprudential actions are anticipated by market participants. We therefore conducted additional tests whether bank equity indices display abnormal returns before the announcement of an LTV action. When we analyse monthly and quarterly returns, we find no evidence that actions were anticipated by the market. Looking at a higher frequency of daily returns, there seems to be some information leakage about the policy change around 10 days before the announcement. For our exercise, however, it matters that actions were not anticipated in the previous quarter.
4. The output and price effects of changes in LTV limits
For the main part of the empirical analysis, this paper uses local projection methods. Jorda (2005) introduces local projections as a way to compute impulse responses without specification and estimation of the underlying multivariate dynamic system. Local projections are estimated at each period of interest rather than extrapolating into increasingly distant horizons from a given model as it is done with VAR models. He discusses the advantages of local projections such as being more robust to mis-specification, being simple in joint or point-wise analytic inference, and being able to easily accommodate experimentation with highly nonlinear and flexible specifications.
Jorda et al. (2013) use local projection methods to condition on a broad set of macroeconomic controls when studying how past credit accumulation affects key macroeconomic variables. Here we study the path of output and prices conditional on a change in LTV limits and macroeconomic controls. We denote the dependent variables, real output and the price level of country i at time t, by yi,t. Ahy;,t = yi,t+h — yi,t denotes our response variable of interest, the change in real output or in the price level between base quarter t and quarter t + h over varying prediction horizons h = 1,2,..., H, where H is 16 in our specifications. We are interested in the response of this variable to a perturbation in our measure of macroprudential policy AMaPPi/t. Specifically, we estimate
for h = 1,..., 16. AMaPPiit denotes changes in macroprudential policy, here the regulatory LTV limit, implemented in country i and quarter t.
In various robustness tests we vary this treatment variable AMaPPit in a number of ways. We first present the results based on the intensity-adjusted variable described in the data section (AMaPPi>t) to assess the impact of a one percentage point change in LTV limits. In the following specification we replace AMaPP/j with an index to connect to the existing literature, assigning a value of i to a tightening action in LTV limits, a value of —1 to a loosening action and zero if there is no action. We denote this variable by AMaPPlndex. We also show results for tightenings (positive values of AMaPP/,t) and loosenings (negative values) separately. We include a rich set of covariates in each specification. These include country dummies to control for country-specific growth rates ah as well as time-fixed effects 7^ to control for global trends. X/,t is a vector that contains the GDP growth, inflation and policy rate changes.
4.1 Main results
We start with our baseline specification, and include the intensity-adjusted LTV change variable AMaPP/t, which refers to the percentage point change in regulatory maximum LTV ratios between t — 1 and t constructed as described earlier in the data section. The results of estimating Equation 1 using the numerical measure of AMaPP/j are visualised in Figure 2. The two panels display the cumulative responses of output and of the price level to a one percentage point change in the maximum LTV ratio over the following 16 quarters. Note that we define AMaPP/,t such that a positive value refers to a tightening in macroprudential policies. A value of AMaPP/,t = 10 refers to a 10 percentage point decrease in the regulatory maximum LTV ratio, for example a tightening from 80% to 70%.
Figure 2 shows that the response to a one percentage point change in LTV limits is a 0.05% lower real GDP after two years, which increases to a 0.11% loss after four years. These results are rather imprecisely estimated: the coefficient is only significant for very short horizons immediately after the LTV action is taken. As we will see later this effect immediately after the implementation of a policy can be attributed to loosening actions only. Consider again the tightening in the maximum LTV ratio by 10 percentage points from 80% to 70%: our estimates correspond to 0.5% lower real GDP after 24 months and 1.1% lower real GDP after 48 months for this scenario.
The price response appears to be slightly positive, but insignificant at most horizons and we caution against a structural interpretation of these rather imprecise estimates. The "price puzzle" disappears once we drop LTV actions taken in Hong Kong around the Asian crisis and the following deflationary period (see Figure A.3). The responses of GDP and prices at different horizons are reported in Table 6. To get a sense for the magnitude of the effects, we will compare our results to existing estimates for the effects of monetary policy shocks below.
LTV actions are sometimes taken in combination with other macroprudential policies. Such correlated treatments could bias our results. To address this concern, we control for regulatory changes affecting debt service-to-income ratios, risk weights and reserve requirements. We furthermore add four lags of changes in the credit-to-GDP ratio, the current-account-to-GDP ratio and foreign currency lending as a proxy of cross-border credit. We do not include these control variables in our baseline specification as the number of observations is reduced by 15%. The results shown in Figure 3 are in line with our previous findings. The output response is a bit stronger than in the baseline specification (-0.17 instead of -0.11 after 16 quarters). The response of the price level is similar to the baseline specification and close to zero after 16 quarters.
Another concern could the potential interaction between macroprudential and monetary policy: if monetary policy reacts to the positive price effects of LTV changes, the contractionary effects we find might be due to monetary policy actions. In the regressions we control for changes in policy rates, but we also analysed the responses of policy rates to a change in LTV limits and did not find evidence for a systematic response.
Figure A.6 in the appendix shows that the response of real consumption is similar to the response of GDP. The coefficient is also estimated rather imprecisely, fluctuating around -0.1, in line with our estimate for the path of real GDP. In the following subsections we will vary the treatment variable in the baseline specification in a number of ways.
4.2 Index variable specification
We now assess the average effects of changes in LTV limits on real economic output and the price level. To connect to the existing literature on macroprudential policies that has focused on changes in macroprudential policy expressed as binary or index variables, we estimate our baseline specification, including the index variable and the set of control variables described earlier. This boils down to the following expression
for h = 1,..., 16. The results are displayed in Figure 4 and show that this typical LTV action has insignificant contractionary effects on real GDP, while there is almost no effect on the price level. The coefficients for various horizons are displayed in Table 7. Increasing our index variable by 1, which corresponds to varying it from no action to a tightening action, lowers real GDP by 1.52% after 16 quarters. This response is again rather imprecisely estimated and not significant as indicated by the shaded area in light grey referring to 1.96 standard deviations. Comparing this result to the coefficient using the numerical value shows that the index specification overestimates the strength of the output effects: the average size of the change in the quantified LTV variable in this sample is around 7.5 percentage points and hence the effect in the index specification is almost twice as high as in the baseline results. The response of the price level is closer to zero over all horizons h.
4.3 The effects of tightening and loosening actions
The negative coefficient for AMaPPi,t suggests that there are small negative spillovers to the real economy from tightening macroprudential policies. Importantly, this applies to policies specifically targeting the financial cycle. The result could also be interpreted such that loose macroprudential policies may be used to stimulate output. To distinguish between the two domains of policy-making, we analyse whether there is a systematic difference between the responses to the quantified LTV tightening and loosening actions. To do so, we include the quantified LTV change in tightening and loosening actions separately. Let us define two variables as follows:
Hence, A>0MaPPi/t measures the size of a tightening action only. A<0MaPPi/t is the loosening analogue:
We can add these two variables to the baseline specification:
for h = 1,..., 16. In Equation 2, ^h denotes the coefficient on changes in maximum LTV ratios in the positive domain and Kh in the negative domain. Remember that our MaPP variable is defined such that tightenings in policy are associated with a positive value of AMaPPi,t. Hence, the fih coefficient refers to the effect of a one percentage point lower maximum LTV ratio when a tightening action is implemented, while kh refers to the effect of a one percentage point higher maximum LTV ratio when a loosening action is implemented. Hence, the baseline estimates are a weighted average of these two estimates.
The results are displayed separately in Figure 5 and Figure 6. The responses of output displayed on the left-hand panels in these figures show a difference between a one percentage point tightening and loosening LTV actions. In particular, the average slightly negative response displayed in the first seven quarters is mainly driven by loosening actions and the response after eight quarters is mainly driven by tightening actions. The price level increases slightly as a reaction to a tightening action, while it shows almost no response to a loosening action. These findings from the visual inspection are confirmed in Table 8.
In this section we check the robustness of our results in a number of subsamples. In particular, we investigate whether the results differ across subsamples of our data. As described earlier, macroprudential policies increasingly received attention after the global financial crisis and this is also true for policies implementing LTV limits. To rule out that results are driven by some characteristics associated with the post-2007 period, we test whether our results also hold in a pre-2007 subsample. We therefore run our baseline local projection using the quantified LTV changes only for policies implemented until 2006. As shown in Figure 7, in this subsample the responses of GDP and prices to a change in the maximum LTV ratio look very similar to the full sample results. The negative response of output is slightly stronger at longer horizons in this sample than in the full sample.
We also ensure that our results are not driven by a single country. We therefore report in the appendix specifications where we drop, one by one, China, Hong Kong SAR and Iceland from our sample. In our sample, these are the countries that have used LTV policies most frequently in EMEs and AEs, respectively. Figure A.3, Figure A.4 and Figure A.5 show that we obtain consistent results when we exclude these economies one by one.
Morover, we test if the efficacy of macro-prudential policies differs between countries at various development stages. We address this possibility in our next test and run our baseline specification for subsamples of EMEs and AEs separately. Our sample contains 23 AEs and 33 EMEs, using the BIS classification of economies.
Figure 8 and Figure 9 show that the responses of output and prices indeed depend on the development stage. The negative output response is entirely driven by EMEs (see Figure 9). In these countries the response is negative at all horizons and this result is statistically significant in the first two years after the policy is implemented. In the sample of AEs (see Figure 8) the response of GDP is very close to zero at all horizons. The price level also displays heterogeneous responses between the two groups. While the response of the price level is small and mostly insignificant in both cases, the pattern differs: the price level first increases slightly in EMEs and returns to zero, while it decreases in AEs first and becomes slightly positive after two years. These results indicate that the costs of using LTV limits are smaller for AEs. One possible channel for these differences could be that policies are better calibrated to desired targets in AEs and hence do not trigger a misallocation of credit.
In addition, we report here the results of using the scope-unadjusted measure for changes in LTV limits. As explained in the data section, we adjusted the size of LTV limit changes for the scope of policies implemented. In this test we assume all changes in LTV limits affect all types of credit and hence no adjustment for scope is necessary. Figure 10
shows the results for our baseline specification using this approach: the coefficients are even smaller, although more precisely estimated.
The appendix shows the results of additional tests: responses over a 30-quarter horizon (Figure A. 7) and responses in boom (GDP above trend) and slump (GDP below trend) subsamples. We show in Figure A.8 and Figure A.9 that the immediate negative response of output is driven by slump periods, while the negative output response in the long run is driven almost entirely by actions taken in a boom. We also analyse separately the set of actions that are intended to moderate the financial cycle and are not intended to foster resilience. We classify actions as targeted at moderating the financial cycle if the stated objectives refer to the growth of house prices, total credit, housing/household credit and FX borrowing. Figure A.10 shows that the output effects of such actions are initially a bit stronger than in the baseline, but remain stable after eight quarters.
In all our specifications we use the implementation dates of changed policies. Usually, the announcement of a policy and its implementation fall into the same quarter. We did, however, check whether our results change if we employ announcement dates rather than implementation dates in the analysis. In total, six actions were affected. In two cases the announcement was in the last week of a quarter. Therefore we assume there was no time for economic actors to adjust their behaviour in this quarter. We then use the announcement
quarter rather than the implementation quarter for the other four changes. The results for the baseline specification are also shown in the appendix (Figure A.11), and there is no visible change compared to the results based on the implementation quarter.
4.5. Comparing LTV changes with monetary policy
Our data on the size of LTV changes also allow us to compare the effects of quantified LTV changes to those of monetary policy. Table 9 displays estimates of the response of real output to a 100 basis point increase in policy rates found in several studies. In particular, we present estimates based on a narrative identification approach (Romer and Romer (2004)), estimates based on high-frequency identification (Gertler and Karadi (2015)) and estimates that are obtained by exploiting the open-economy trilemma for identification (Jorda et al. (2017)). Furthermore, we report the state-dependent effects analysed in Tenreyro and Thwaites (2016).
We compare our estimates of the effects of LTV limit changes with those of monetary policy. We do this based on the following question: what is the monetary policy adjustment that yields the same output response as a 10 percentage point LTV limit change? In our
baseline regression, the estimated response to this change after two years is -0.5%.
This response would correspond to a change of 12 basis points in the policy rate based on the results reported in Romer and Romer (2004). Based on a high frequency identification procedure, Gertler and Karadi (2015) find smaller effects and our estimate of a 10 percentage point LTV limit tightening would correspond to a change of around 42 basis points in the federal funds rate. In another study, Jorda et al. (2017) estimate the response of real GDP to short-term interest rates and find an effect of -1.9% after two years. This estimate is most likely the best comparison for our purpose, as it is also based on local projections using an international panel dataset. Here, our estimated response of real GDP corresponds to a 26 basis point interest rate change, falling in the middle of the range between the other results.
What do we learn from this exercise? A monetary policy shock of 26 basis points is certainly not negligible, but our estimates are rather imprecise compared to those identified in the literature on monetary policy shocks. Tenreyro and Thwaites (2016) and Jorda et al. (2017) show that the output costs of monetary policy are much higher in the boom than in the slump. We show in the appendix that these effects are very similar for LTV changes. There seems to be little room for policy-makers to exploit state dependence of the effects when choosing between monetary and macroprudential policies. The good news for policy-makers is that the output response to a change in LTV limits is attenuated in the sample of AEs. Here the output costs seem rather low, while the costs of interest rate policies discussed above applied almost exclusively to the United States or a sample of AEs.
5 The effects on financial variables
What are the benefits of macroprudential tightenings? The previous section has shown that tightenings in LTV limits are associated with modest output costs, but the cost- benefit tradeoff of implementing macroprudential policies also depends on the efficacy of macroprudential policies to dampen the financial cycle. In this section we therefore analyse the responses of real household and mortgage credit and of asset prices to a tightening in the maximum LTV ratio. To answer the question, we cannot rely on the same identification strategy we used to assess the output costs. As we previously have argued, the objectives of LTV limit changes are mostly related to the financial cycle and policies are often implemented to affect credit variables. We therefore need to approach this question using a different identification strategy.
To estimate the response to a tightening action in LTV limits, which is the treatment, we employ an inverse probability weighted regression-adjusted (IPWRA) estimator. An inverse probability weighted (IPW) estimator gives more weight to those treatments that are difficult to predict based on observables and less weight to those instances that are endogenous due to the other factors. In combination with the local projections approach we employ in this paper, controlling for Xi,t, we obtain the IPWRA estimator. Jorda et al. (2016) use an IPWRA estimator to account for the probability of observing a financial crisis driven recession rather than a normal recession. Other applications study the response to sovereign defaults (Kuvshinov and Zimmermann (2016)) and austerity (Jorda and Taylor (2016)).
We use the IPWRA estimator to analyse the responses of financial variables to a tightening in LTV limits. Let d,t be the tightening dummy that takes value one if AMaPPi ,t > 0 and zero otherwise. The estimation proceeds in two stages. In a first step, we specify a logit model to estimate the probability that LTV limits are tightened
Zi,t-i is a vector of macroeconomic controls at time t — 1, where we include smoothed four-quarter changes in real and financial variables. In addition, we include country-fixed effects to account for country-specific usage of LTV policies. We refer to the probability of a tightening as the propensity score and its estimate from Equation 3 is denoted by pi,t. In the second stage, we estimate local projections using regression weights given by the inverse of pi,t. Weighting by the inverse of the propensity score puts more weight on those observations that were difficult to predict and thereby re-randomises the treatment. In our application, this implies putting more weight on regulatory LTV tightenings that were taken as a surprise based on observables, and putting less weight on those tightenings that could be predicted.
Ahyi,t now denotes the change in financial variables between time t and t + h. To implement the second stage, we run the following specification using weighted least squares (WLS)
The weights are defined by wi,t = di,t/piit + (1 — di,t)/ (1 — pi,t), where we truncate wi't at 10. In this specification, we include the baseline set of controls from our previous exercises and lagged changes in the response variable. Further details on the methodology can be found in Jorda and Taylor (2016) and Jorda et al. (2016).
5.1. Credit responses
We first apply this procedure to estimate the response of credit variables targeted by LTV limits. These targeted variables are normally household credit and mortgage credit. Table 10 presents the results of our first stage. We run logit classification models for the tightening dummy ATight MaPPit as we want to account for increases in financial variables presumably targeted by tightening actions. Hence, we include in this regression the smoothed growth rates over the previous four quarters of detrended GDP and the detrended price level. Furthermore, we include the growth rate of real credit variables and the real stock price index over the previous four quarters in the regression as well as country-fixed effects. Credit variables, and here particularly the growth rate of total real private credit, emerge as the best predictors of a tightening action in LTV limits. The number of observations is reduced to 437, which is due to country fixed effects, because not all countries in our sample introduced an LTV tightening during our sample period. Furthermore, housing and mortgage credit data are available only for a subset of our observations. We report the AUC statistics which stands for area under the receiver operating curve. The statistic measures the ability of a model to correctly sort observations into the "tightening" and "no tightening" bins as combinations of true positive and false positive rates that result from changing the threshold variable for classification. In other words, it yields a summary measure of predictive ability that is independent of individual cut-off values chosen. The AUC takes on the value of 1 for perfect classification ability and 0.5 for an uninformed classifier or the results of a 'coin toss'. Here the AUC is 0.77 which is a significant improvement over the coin toss.
Figure A.1 in the appendix plots the estimated probabilities of treatment based on the first stage, differentiating between treated units (red) and control units (blue). Our procedure in the second-stage regression now assigns a higher weight to the treated observations that were less likely to be treated based on this analysis, i.e. those observations with very low probabilities. Figure 11 presents the results of these inverse propensity weighted local projections. We find that a tightening action in maximum LTV ratios decreases credit variables as intended. Real household credit is reduced by almost 6% after two years and mortgage credit by more than 5%. Both coefficients are statistically significant. The coefficients remain stable for longer time horizons, while confidence intervals widen
and, as a result, the effects are no longer significant after four years. These results are in line with a number of recent studies that find negative effects of macroprudential tightenings on credit. Remember that the number of observations is reduced in this set-up, as we use a large set of fixed effects and covariates in the first stage logit regression to predict the treatment probabilities. Figure A. 12 shows that conditional estimates, which do not employ an IPWRA estimator, display a stronger decline in credit. We also analysed the response of total domestic bank credit (deflated with the CPI). Similar to the findings in Akinci and Olmstead-Rumsey (2018), we do not find a significant response of total bank credit.
5.2 Asset prices
In addition to the effect of LTV tightenings on credit variables, we assess the effect of these measures on asset prices; here we focus on stock prices and house prices. The first stage is again a logit classification model for the tightening dummy. We use a similar set of predictors as in the previous set-up, but replace household and mortgage credit by real house price growth, which is also only available for a subset of observations. Figure A.2 in the appendix plots the estimated probabilities of treatment between treated units (red) and control units (blue). Table 12 presents the results of this approach. In line with the first stage for credit variables, we find that real growth of private credit is a driver of the probability of an LTV limit tightening being enacted.
Table 13, Figure 12 and Figure A. 13 present the results of this exercise. We see that both, real stock and house prices, initially fall after a tightening action has been implemented. There is strong heterogeneity in the response of stock prices as displayed in the left-hand panel. While the coefficient is negative and large, the confidence intervals are wide and, as a result, this negative effect is insignificant. The response of real house prices is quite different. Here, the confidence intervals are rather small and the negative effect becomes significant after two years. The coefficient further declines the longer the horizon, reaching a highly significant 8% decline in real house prices after four years. This result is in line with many findings in the literature as well as the targeted nature of LTV limits. As we have seen from analysing the response of credit variables, LTV actions seem to be effective in reducing mortgage credit and they also dampen house prices.
This paper uses a narrative identification approach to determine the causal effect of changes in maximum LTV ratios, an important element of the macroprudential toolkit, on output and inflation. Our main objective is to understand to what extent, if at all, macroprudential policy interferes with the main objectives of monetary policy to stabilise output and inflation. For this purpose, we introduced a dataset that codes exogenous and intensity-adjusted changes in maximum LTV ratios for a sample of advanced and emerging market economies.
Our main result is that changes in maximum LTV ratios appear to have relatively modest effects on output and inflation. The output effects tend to be imprecisely estimated and small in advanced economies. After employing a number of different specifications and a battery of robustness tests, we show that our results point to more sizeable effects of LTV tightening than loosening, as well as to consistently larger effects in emerging market economies. The effects of LTV changes on inflation tend to be negligible. As a rule of thumb, over a two-year horizon the mean output effect of a 10 percentage point change in maximum LTV ratios corresponds roughly to that of a 25 basis point change in policy rates, but the standard errors are large. Using inverse propensity weighting methods to mimic random allocation of the LTV treatment, we also provide evidence that LTV changes have substantial effects on credit growth and house prices.
This paper contains several potentially important implications for policy makers. First, our results suggest that central banks in advanced economies could be in a position to use macroprudential instruments to manage financial booms without interfering with the monetary policy objectives in a major way. Second, the use of a scope-adjusted quantified LTV change variable in this paper makes first inroads towards calibrating macroprudential tools. Finally, the evidence in this paper demonstrates that changes in maximum LTV ratios introduced under financial objectives tend to have rather substantial effects on activity in credit markets and house prices as intended.
Aikman, David, Bush, Oliver, and Taylor, Alan M. 2016. Monetary versus Macroprudential Policies: Causal Impacts of Interest Rates and Credit Controls in the Era of the UK Radcliffe Report. Bank of England Staff Working Paper No. 610.
Akinci, Ozge, and Olmstead-Rumsey, Jane. 2018. How Effective are Macroprudential Policies? An Empirical Investigation. Journal of Financial Intermediation, 33, 33-57.
Alpanda, Sami, and Zubairy, Sarah. 2017. Addressing Household Indebtedness: Monetary, Fiscal or Macroprudential Policy? European Economic Review, 92, 47 - 73.
Bachmann, Rudiger, and Rueth, Sebastian. 2017. Systematic Monetary Policy and the Macroeconomic Effects of Shifts in Loan-to-Value Ratios. CEPR Discussion Paper 12024.
Bernanke, Ben S., Boivin, Jean, and Eliasz, Piotr. 2005. Measuring the Effects of Monetary Policy: A Factor-Augmented Vector Autoregressive (FAVAR) Approach. Quarterly Journal of Economics, 120(1), 387-422.
Boar, Codruta, Gambacorta, Leonardo, Lombardo, Giovanni, and da Silva, Luiz Pereira. 2017. What are the Effects of Macroprudential Policies on Macroeconomic Performance? BIS Quarterly Review.
Bruno, Valentina, Shim, Ilhyock, and Shin, Hyun S. 2017. Comparative Assessment of Macroprudential Policies. Journal of Financial Stability, 28,183 - 202.
Budnik, Katarzyna, and Kleibl, Johannes. 2018. Macroprudential Regulation in the European Union in 1995-2014: Introducing a New Data Det on Policy Actions of a Macroprudential Nature. ECB Working Paper, 2123.
Cerutti, Eugenio, Claessens, Stijn, and Laeven, Luc. 2015. The Use and Effectiveness of Macroprudential Policies: New Evidence. IMF Working Paper 15/61.
Claessens, Stijn, Ghosh, Swati R., and Mihet, Roxana. 2013. Macro-prudential Policies to Mitigate Financial System Vulnerabilities. Journal ofInternational Money and Finance, 39, 153 - 185.
Dell'Ariccia, Giovanni, Igan, Deniz, Laeven, Luc, Tong, Hui, Bakker, Bas, and Vandenbussche, Jerome. 2012. Policies for Macrofinancial Stability: How to Deal with Credit Booms. IMF Staff Discussion Note 12/06.
Friedman, Milton, and Schwartz, Anna J. 1963. A Monetary History of the United States 1867-1960. Princeton: Princeton University.
Gertler, Mark, and Karadi, Peter. 2015. Monetary Policy Surprises, Credit Costs, and Economic Activity. American Economic Journal: Macroeconomics, 7(1), 44-76.
Gillitzer, Christian. 2017. Do Output Contractions Cause Investment in Fiscal Capacity? American Economic Journal: Economic Policy, 9(2), 189-227.
Glocker, Christian, and Towbin, Pascal. 2015. Reserve requirements as a macroprudential instrument Empirical evidence from Brazil. Journal of Macroeconomics, 44(Supplement C), 158 - 176.
Hamilton, James D. 2017. Why You Should Never Use the Hodrick-Prescott Filter. Review of Economics and Statistics. Forthcoming.
Jorda, (Oscar. 2005. Estimation and Inference of Impulse Responses by Local Projections. American Economic Review, 95(1), 161-182.
Jorda, (Oscar, and Taylor, Alan M. 2016. The Time for Austerity: Estimating the Average Treatment Effect of Fiscal Policy. The Economic Journal, 126(590), 219-255.
Jorda, (Oscar, Schularick, Moritz, and Taylor, Alan M. 2013. When Credit Bites Back. Journal of Money, Credit and Banking, 45(2).
Jorda, (Oscar, Schularick, Moritz, and Taylor, Alan M. 2015. Betting the House. Journal of International Economics, 96(Supplement 1), 2-18.
Jorda, (Oscar, Schularick, Moritz, and Taylor, Alan M. 2016. The Great Mortgaging: Housing Finance, Crises, and Business Cycles. Economic Policy, 31(85), 107-152.
Jorda, (Oscar, Schularick, Moritz, and Taylor, Alan M. 2017. Large and state-dependent effects of quasi-random monetary experiments. CEPR Discussion Paper 11801.