• No results found

Measuring the Credit Gap: a Structural Approach for the Netherlands

N/A
N/A
Protected

Academic year: 2023

Share "Measuring the Credit Gap: a Structural Approach for the Netherlands"

Copied!
79
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

Measuring the Credit Gap: a Structural Approach for the Netherlands

By: Jorrit Marc Thunnissen

13274481

Thesis for the Completion of the MSc Economics at the University of Amsterdam

First Reader and Supervisor: Aerdt Houben Second Reader: Christian Stoltenberg

Supervisor CPB: Bert Kramer

Date: 15/08/2021

Statement of Originality

This document is written by Student Jorrit Marc Thunnissen who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it. UvA Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(2)

Table of Contents

TABLE OF CONTENTS 2

I. INTRODUCTION 4

II. EMPIRICAL FINDINGS AND THEORIES ON THE DRIVERS OF THE CREDIT CYCLE 6

Why do we have credit cycles, and what are their risks? 6

Financial Frictions 6

Behavioural Theories 7

Empirical Findings of the Credit Cycle, and its relation to Asset Prices and the Financial Cycle 10

III. LITERATURE REVIEW ON MEASURING THE CREDIT GAP 13

The Basel Credit Gap 13

Main Limitations of the BCG 14

Literature Review on Measuring the Credit Gap 15

Statistical Approaches 15

Structural Methods 17

IV. METHODOLOGY 19

Model Set-up 19

Other specifications 21

Constraints 22

Comparison Criteria 23

Robustness Checks 24

AUC Sample 24

HP-filter smoothing parameter and sample size 25

Different Countries 26

Data 26

Graphing the Data 26

V. RESULTS ANALYSIS FOR THE NETHERLANDS 28

Coefficient Analysis of the Main Specifications 29

Comparison Credit Gap Estimates of the Main VECM Specifications and the BCG 33

1980s and 1990s 34

2000s and Onwards 37

Preliminary Conclusions 39

Statistical Comparisons 39

(3)

Other Specifications 40

Statistical Comparisons 41

Bank Credit 42

Credit to Households or Companies 44

Excluding the Real Interest Rate 46

Changing Variables into Nominal Terms 48

Altering the Trend Specification 48

VI. ROBUSTNESS CHECKS 49

Altering the AUC Sample 49

Altering the Smoothing Parameter and Starting Points of the HP filter 50

VEC Model with Asset Prices for Multiple Countries 54

VII. CONCLUSION 56

APPENDIX I 59

APPENDIX II 67

APPENDIX III 69

BIBLIOGRAPHY 71

(4)

I. Introduction

To mitigate the impact of financial crises, policymakers need to be able to identify ahead of time the build-up of macro-financial risks in order to be able to act counter-cyclically. Minsky (1978, 1986) already noted that the build-up of these risks often occurs when the financial system seems to be at its most stable. The build-up also occurs when the cost of raising bank capital is comparatively low, and financial conditions are loose. It is in these periods that capital buffers should be raised, in order for them to be released when banks need capital to absorb their losses, and financial stability is threatened due to fire sales and contagion. What is more, after the deflation of a boom, macroprudential policymakers need to consider whether more adjustment of the credit level is needed or whether the downward correction has gone further than its desired level. To make banks’ capital buffers behave in such a cyclical manner, at the guidance of macroprudential authorities, the Countercyclical Capital Buffer (CCyB) was introduced in 2010 as part of the new Basel III framework.

There is no consensus yet on a preferred measure for excessive credit. This reflects in part the lack of a ‘canonical’ macroeconomic model of financial excesses (Baba et al., 2020).

Nevertheless, since 2010, the Basel Committee on Banking Supervision (BCBS) recommends using the difference between a country’s actual credit-to-GDP and its long-run Hodrick Prescott (HP)-filtered trend as a common reference point for the calibration of the CCyB. This is also called the Basel Credit Gap (BCG). As its use has expanded, however, its limitations have increasingly come to light (Edge & Meisenzahl, 2011; Hamilton 2018; ECB, 2019). For instance, in the Netherlands the BCG measure indicated a historically high credit gap at the height of the Sovereign Debt Crisis (SDC). Contrastingly, the BCG measure was large and negative before the Covid crisis, which implied credit was far below trend. This was also true for many other European countries (Baba et al., 2020). However, with interest rates historically low, output gaps positive, and supportive financial conditions, macroprudential policymakers saw risks building up. As such, they were more inclined to tighten rather than loosen macroprudential measures (Baba et al., 2020). This begs the question whether the BCG is the most appropriate measure for the Dutch credit cycle.

Figure 1. Basel Credit Gap Estimates for the Netherlands. The grey-shaded areas are macroprudential relevant stress events as identified by Lo Duca et al. (2017). Source: BIS (2021), Lo Duca et al. (2017), and author’s own calculations.

-5 -4 -3 -2 -1 0 1 2

1970q1 1975q1 1980q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 2020q1 Financial Crises Basel Credit Gap

(5)

As Greenwood et al. (2020) recently showed, looking at credit with the combination of movements in equity and house prices provides much clearer signals when financial vulnerabilities are building up. What is more, Galan and Mencia (2021) exploit the endogenous feedback effects that credit, GDP, house prices, and the real interest rate have on one another through a structural Vector Error Correction Model, and estimate in real-time the credit gaps. They find that this model significantly outperforms the BCG in its predictive power for financial crises. Combining the insights of both studies, might prove to be enlightening. I therefore pose the following research question: “Does a structural Vector Error Correction Model with asset prices improve on the Basel Credit Gap in terms of the credit gap estimation of the Netherlands?”

To stay within the tradition of much of the literature (e.g., Drehmann et al., 2012; Lang

& Welz, 2018; Drehmann & Yetman, 2018, 2020; Galan & Mencia, 2021), I test the crisis early warning properties of the measure and find a significant improvement when compared to the BCG. These results, however, cannot be taken too far considering the small sample size and the number of macroprudential-relevant stress events. What is more, predicting a financial crisis from happening is different from measuring the credit gap over the entire length of the credit cycle (Baba et al., 2020). I therefore also qualitatively analyse the credit gap estimates to show that, like the BCG, the model yields credit gaps that turn positive ahead of macroprudential relevant stress events for the Netherlands, but, unlike the BCG, do not remain negative for an extended period following the burst of a large and prolonged credit boom. Thus, the findings indicate that the model is a good measure for the entire credit cycle of the Netherlands. The results are less robust when running several model specifications also for France, Germany, Sweden, the United Kingdom, and the United States. This underlines difficulty of finding an internationally homogeneous model for financial vulnerabilities, the differences that exist between domestic financial cycles and thus the importance of a deep country-focused approach, and the importance of national discretion in macroprudential policymaking.

Because I focus on one country and one alternative econometric method, I am able to deepen my analysis in three ways as compared to the existing literature on financial vulnerabilities indicators. First, I can analyse my results qualitatively. This allows for the analysis of the performance of the econometric methods across the entire length of the credit cycle, rather than only right before a crisis (Baba et al., 2020). Although I also compare the AUC-values for the different methods, that only tells part of the story. Second, I am able to expand more on the choice of the included variables. Due to the large variety of macroprudential measures and their focus on banks, an indicator that solely focuses on total credit from all sectors to the nonfinancial private sector does not provide enough information. Altering the source of credit (from banks) and disaggregating it by the type of borrower (households and corporates) could provide policymakers with more tailored information. And third, I am able to scrutinise some of the underlying assumptions and the dynamics of the HP filter used in estimating the BCG. As Galan (2019) showed for Spain, altering several features of the HP filter improved the BCG’s performance significantly. Jokipii et al. (2020) recently showed for Switzerland that changing the starting point of the data sample of the HP filter altered its results as well. In the majority of the studies that propose new indicators, these alterations to the HP filter are not taken into account. Doing so would increase the robustness of the findings. These three expansions allow for a more thorough answering of the research question that would otherwise have been the case. A last point in where I expand on the existing literature, is that I also estimate the structural model in real-

(6)

time. This is of particular relevance for policymakers as they make their decisions in real-time as well. Furthermore, this allows for a truly proper comparison with the BCG, as it too is estimated in real-time.

The structure of the rest of the paper is as follows. Section II discusses the most recent empirical findings on the credit cycle, and theories on what drives this credit cycle. Section III sets out a literature review on measuring credit gaps. Section IV lays out the methodology of the research. Section V discusses the results for the Netherlands. Section VI sets out three robustness checks of the results from section V. Section VII concludes the paper.

II. Empirical Findings and Theories on the Drivers of the Credit Cycle

Before discussing how credit cycles are measured, it is useful to first analyse the existing ideas on what causes them and what the current empirical findings on credit cycles are. What will become evident is that only looking at the quantity of credit does not capture all the dynamics associated with the credit cycle. As such, the BCG, which solely focuses on quantity variables, most likely falls short in its characterisation of the credit cycle. A measure which also includes asset prices might come closer to the characterisation of the credit cycle and its associated dynamics. The first half of this section will cover theories on why credit cycles exist. The second half of the section will focus on the empirical findings concerning credit cycles, and why a measure including more than only credit and GDP might be better at characterising the credit cycle.

Why do we have credit cycles, and what are their risks?

There is a multitude of theories that explain how credit booms can lead to economic downturns or financial crises. They can largely be grouped into two groups: financial frictions theories and theories that focus on investor behaviour.

Financial Frictions

The macroeconomic discipline has long integrated financial frictions into models to study aggregate fluctuations. An early example is Fisher (1933) who discusses the role of debt- deflation dynamics at the time of the Great Depression. Modern examples are Bernanke and Gertler (1989), Kiyotaki and Moore (1997), and Geanakoplos (2009). They all share some core elements: firstly, all agents are assumed to be rational and to have rational expectations.

Secondly, debt contracts are the primary source of external finance, due to a multitude of agency issues. And thirdly, frictions in the debt market are prevalent in the forms of constraints on the ability to borrow. This is either endogenous through constraints on collateral value or borrower net worth, or an exogenous debt limit (Stein, 2021).

All of these ingredients together create propagation and amplification effects: At the time of a negative shock to the economy, the net worth of firms and households is reduced as they have borrowed to finance past spending. Due to debt market frictions, they are forced to cut back on their borrowing, and reduce future consumption and investments. The accompanying reduced aggregate demand then leads to a further decline in economic output.

(7)

This then again leads to a further reduction in collateral values and borrower net worth, and the downturn deepens, accentuating the financial cycle.

Some recent studies take this approach further and are especially relevant in light of the recent Global Financial Crisis (GFC). Brunnermeier and Sannikov (2014) find that the amplifying effects of financial frictions might be substantially nonlinear. This means that the response of an economy to a small external shock is much smaller than its response to a larger shock. Hall (2011) and Eggertsson and Krugman (2012), in turn, find that the downturn which results is prolonged when the Zero Lower Bound (ZLB) on nominal interest rates impedes the equilibrating process of the economy (Stein, 2021).

However, as the agents in these macro-models are entirely rational, and the rapid and large-scale accumulation of debt makes the economy fragile, why would agents do so in the first place? The common answer that is provided in the literature is that the leverage choice contains externalities: individuals do not fully take into account the imposition of vulnerabilities on the economy as a result of their (individual) borrowing decisions. This leads them to over-borrow from the social planner’s perspective (Stein, 2021). These externalities can be imbedded in aggregate demand spill-overs when accompanied by the ZLB (Farhi &

Werning, 2016; Korinek & Simsek, 2016), or in the presence of fire-sale effects (Lorenzoni, 2008; Stein, 2012).

In short, these financial frictions models can explain why economies with significantly leveraged households, firms, or investors are potentially susceptible to the effects of exogenous shocks. They also explain why individual decisions of these agents can then result in ex ante high leverage, despite the potential costs. What is more, they affirm the empirical studies that utilise balance-sheet measures of leverage to forecast potential economic recessions. This is because they emphasise that leverage captures the fragility of the economy.

Nevertheless, these financial-frictions models are theories of propagation and amplifications, and always depend on a shock from outside the system to put the gears into motion. Therefore, these theories have usually very little (or nothing at all) to say concerning how and when a credit-driven downturn is triggered. In relation to this, the duration of the credit cycle is also often disregarded. To illustrate, consider that the arrival of large negative shocks only happens infrequently. Then, policymakers who observe that an economy is in a highly leveraged fragile state might have to (on average) wait quite some time before they see the forecasted downturn, because they have no extra information concerning the probability of an exogenous shock occurring. Perhaps this is the reason that Borio (2014) commented that macroeconomists have largely included the financial cycle only very conservatively in their models.

Behavioural Theories

A different way to study credit cycles is found in the narratives of Kindleberger (1978, 2015) and Minsky (1977, 1986, 1995), and works within the behavioural finance discipline that focuses on the imperfectly rational beliefs of investors. Two important recent studies in this strand are Bordalo et al. (2018) and Greenwood et al. (2020). These studies attempt to uncover three credit cycle aspects: First, why investors can become overoptimistic, which drives credit spreads to excessively low levels. Second, what leads this overoptimism to

(8)

endogenously reverse, causing a consequent worsening of credit conditions. And third, the accompanying macroeconomic dynamics of this.

In their ground-breaking work, Kahneman and Tversky (1972) laid out their idea of the representativeness heuristic. According to them, when an attribute is diagnostic for a population, which means that the attribute occurs more in that population compared to a relevant reference population, then that attribute will be judged to be common excessively so in that population. In a financial context, this means that investors overweigh future states whose likelihood has increased the most with the new information they are presented with, relative to what they already know. Therefore, when investors receive good news, they overestimate the probability of a good future state. This is referred to as ‘diagnostic expectations’, because investors overweigh diagnostic information.

Diagnostic expectations have important implications. An upward path of improving news will lead investors to neglect bad future outcomes and focus on the good ones, causing excessive optimism. On the other hand, a downward path of deteriorating news will lead investors to neglect positive future outcomes and only focus on the bad ones, causing excessive pessimism. There is an important component of truth here: investors revise their expectations following news, but do this excessively so. This means that when change slows, and the continual stream of good news dries up, there is a sentiment reversal. Thus, crises can even occur in good economic times without the need for the actual deterioration of fundamentals.

Related to the idea of diagnostic expectations is the notion of heterogeneity of market uncertainty (Knight, 1921; Bewley, 1986, 1989; Rigotti & Shannon, 2004; Slovik, 2011).

According to the efficient market hypothesis (Fama, 1970), asset prices incorporate all known information and changes in these prices reflect the emergence of new information. Thus, by default, these prices cannot reflect unknown information. These types of information form two distinct information sets, and are affected by two opposing processes. First, as time passes, new information becomes available, so the set of known information grows, and the set of unknown information shrinks. Second, due to the ever-changing nature of the socio- economic system, the available information becomes outdated, thereby shrinking the known information set and expanding the unknown information set. Significant changes to the available knowledge and information can occur gradually over time, or very suddenly. An example for the latter is a break-through innovation within a sector. When a sudden change to the known information set occurs, a substantial price change is to be expected, as prices reflect the set of known information. In short, market prices are built on available knowledge, while market uncertainty signifies the uncertainty of the validity of that knowledge.

As Slovik (2011) states, however, the market uncertainty hypothesis, implies that not all efficient market prices are created equal. The uncertainty of the validity of market prices is especially high if available information is limited. Again, this happens most often when a substantial change happens to the known information set, such as a technological innovation in a sector. When this happens, market instability will increase, due to the increase of uncertainty that is not embedded in the prices. Then, even with the economic agents being fully rational in utilising known information, the financial system is in a more fragile state. The financial system would then also be more leveraged than would be optimal from a social planner’s point of view.

Related to the idea of diagnostic expectations and the uncertainty hypothesis, is the notion of ‘herd behaviour’. Already mentioned by Keynes (1936), and formalised by Scharfstein and Stein (1990), herd behaviour is the tendency of investors to mimic the actions

(9)

of other investors after observing and interacting with the practices of each other. What is more, herd behaviour leads to investors following the behaviour of other investors and disregarding their own private signals or prevailing market fundamentals, as they are fearful of the consequences for their reputation as sensible investors that their contrarian behaviour might have (Scharfstein & Stein 1990; Erdenetsogt & Kallinterakis 2016). This behaviour then feeds off the increasing number of investors partaking in certain behaviour, causing excessive upward and downward reactions of credit markets.

In the models described above, investors’ extrapolative beliefs of the future or the effects of their actions on their reputation lead to time-varying credit market sentiment.

Alternatively, Myerson (2012) and Stein (2013) argue that mistaken beliefs, whilst important, are not the entire story. Their theories relate to the significant funds of others which financial intermediaries have at their disposal. This leads to bankers holding substantial financial power, and they might be tempted to abuse this power for their own gain. An example of this would be that bankers would seek larger short-term gains at the expense of long-term returns, because that is more beneficial for their bonus pay-outs. This is called moral hazard.

Becker and Stigltiz (1974) and Shapiro and Stiglitz (1984) advocated for solving this issue by promising larger late-career rewards for individual investors or bankers who keep strong performance records throughout their career. Myerson (2012) argues that this requires that bankers must anticipate at least some form of long-term relationship with investors. Thus, even when the vast majority of physical investments are in nature short-term, investors can be compelled to agree to limits on the liquidity of their investments. The ability of investors to trust their bankers must depend on expectations concerning future profits in banking in the long-term, as the motivation of bankers to identify good investments is driven by the promise of large late-career rewards. This means that net value of a mid-career banker’s position depends, at any point in time, is dependent on recent economic activity. As such, the value of a mid-career banker’s position becomes a state variable which is able to affect the amount of current investment. Aggregate current investment must then decline when trusted bankers become fewer in number. Therefore, dynamic forces which drive fluctuations in aggregate economic activity can be created through long-term solutions for financial moral hazard.

Stein (2013) argues that this agency problem between financial intermediaries and their stakeholders increases when interest rates are low. Investors demand from their bankers some form of monetary return on their investments. In a low interest rate environment, safe investments, such as government bonds of Western countries, yield a much lower return than in a high interest rate environment. This then makes intermediaries more likely to ‘search for yield’, meaning that they are willing to accept lower risk premia for bearing credit and duration risk (CPB, 2020). For more work on this, see for instance Rajan (2005) Borio and Zhu (2012), Hanson and Stein (2015) and Gertler and Karadi (2015).

Whatever the cause for this procyclicality, the idea that financial markets are procyclical is what underpins the entirety of macroprudential policy. Traditional micro- prudential policies were judged to be too cyclical in light of the events of the GFC. Most notably, risk-based capital requirements are cyclical, because low risk indicators imply low capital requirements. This is an essential consideration that led to the CCyB, which would require higher risk-weighted capital requirements when credit market sentiment is elevated, so as to account for this procyclicality.

(10)

Of course, while these models are logically distinct, it does not seem as if they are mutually exclusive. Actually, they might be very good complements. A way to see how this might be so is to understand that models based on financial frictions are good at explaining how and why an economy might be in a vulnerable highly levered state. And yet, they usually require a shock from outside the system to trigger a downturn. Essentially, they are models of vulnerabilities, but not triggers (Stein, 2021). On the other hand, sentiment-models, which focus on overoptimistic beliefs endogenously unwinding, get more close to theorising triggers. Indeed, Minsky (1986, 1995) placed this interplay between mispricing and leverage central in his seminal works.

Empirically, some recent work combining these two theories are Krishnamurthy and Muir (2020) and Kirti (2020). These use an interactive regression specification and indeed find evidence for the interrelation of these two theories. However, much work is still to be done.

Another step would be to use (semi-)structural models and integrate sentiment proxy- variables into the specification. Which variables these might be, will be discussed in the second half of section II, after laying out two stylised groups of facts concerning credit cycles.

Empirical Findings of the Credit Cycle, and its relation to Asset Prices and the Financial Cycle

To ease the overview, I emphasise two stylised groups of facts: quantity-oriented- and credit- market sentiment-based evidence. Concerning the first, when looking at quantity data that capture growth in aggregated credit, credit growth foreshadows at fairly low frequencies adverse economic outcomes. These can either be financial crises or other types of economic slowdowns. However, credit-quantity variables do not capture by themselves all the negative information concerning future credit growth. This is where the second group concerning credit market sentiment comes in.

‘Sentiment’ can in this sense be seen as the time-variation in expected returns. In this way, when credit-market sentiment is quite high, that would be equal to stating that bearing credit risk generates low expected returns. And it turns out that higher sentiment in this regard negatively indicates future economic activity. A way of interpreting this channel is that investors are over-optimistic and are at a higher risk of being disappointed, when sentiment is elevated. When these investors are disappointed, credit conditions tend to reverse quite sharply. This corresponds to credit supply shifting inward, which leads to a contraction in economic activity. Again, the overarching idea of both types of evidence is that credit booms, and especially those that are accompanied by elevated sentiment (such as the aggressive pricing of credit risk and asset prices), usually do not end well.

Arguably, two of the most influential works regarding quantity-oriented evidence have been that of the aptly named “Credit Booms Gone Bust” by Schularick and Taylor (2012), and “when Credit Bites Back” by Jorda et al. (2013). In the paper by Schularick and Taylor (2012) fourteen developed economies are studied during the period 1870-2008. 1 They follow Bordo et al. (2018) and Reinhart and Rogoff (2009) and define crises as public interventions in the banking systems and/or bank runs. Their key finding is that the probability of a financial crisis in one period is significantly increased when there is a growth of bank loans relative to

1 These countries include: Australia, Canada, Denmark, France, Germany, Italy, Japan, the Netherlands, and the USA.

(11)

GDP five years prior to this period. What is more, this occurs at relatively low frequencies. In other words, it will usually take some time before a boom turns into a bust. Thus, if an econometrician analyses a country with rapid growth of (total or bank) credit, he could come to the conclusion that the economy is vulnerable but not per se forecast that by next year there is an imminent reversal in credit market sentiment.

With regards to credit market sentiment, López-Salido et al. (2017) study the impact of sentiment for the United States (US) for the period of 1929 to 2015. To start with, they build on the previous research of Greenwood and Hanson (2013), who found that when the share of high-yield bond issuance of total corporate bonds issuance is high, and when credit spreads are narrow, the expected returns of taking on credit risk are relatively low. To put it differently, narrow credit spreads and a relatively large share of high-yield bonds indicate elevated sentiment in credit markets. In turn, a reversal of sentiment in credit markets is likely to lead to a decline in economic activity, as discussed earlier in this section. López-Salido et al. (2017) show that significantly elevated credit-market sentiment in one year leads to lower economic activity in the two and three years after this elevation in sentiment. According to them, what underlies this result is the prevalence of predictable mean reversion in credit market conditions when sentiment is elevated. They say that spreads widen after credit risk is aggressively priced. The moment the spreads widen is then strongly related to a dampening of economic activity. What is more, López-Salido et al. (2017) find that elevated sentiment in credit markets in one period, leads to a net debt issuance fall two periods after. This corresponds a reversal of conditions in credit markets which lead to a contraction of the supply of credit (Stein, 2021).

More recently, Krishnamurthy and Muir (2020) expand on these findings by using a long panel spanning nineteen countries and going back 150 years. They assess the behaviour output and credit across a financial crisis where they use credit growth and credit spreads information. What they find is that an economy transitions into a crisis when sentiment, proxied by credit market spreads, drops substantially. Furthermore, they find that credit spreads narrow, indicating an increase in sentiment, just before a crisis. They conclude that the interrelation between prices and credit quantities indicate that expansions of the credit supply are a prelude to crises. Differently from López-Salido et al. (2017), however, they do not find a significant effect of a credit supply expansion on output. Nevertheless, what emerges once more is that is important to measure both increases in investor sentiment as well as the quantity of credit to capture the elements of credit booms that appear to be associated with subsequent financial crises. Greenwood et al., (2020) reports similar findings, but equate ‘sentiment’ not to credit market spreads but to asset prices more generally. They find that the likelihood of a financial crisis happening significantly increases when asset price inflation and rapid increases in credit move in tandem.

These results suggest that a model such as the BCG, which only looks at quantity variables such as credit and GDP, disregards an important component of the credit cycle:

sentiment. Including a proxy for sentiment might therefore significantly improve the identification of the credit cycle and increasing (and decreasing) systemic risks. One way of doing so, is to include asset prices in the model (Greenwood et al., 2020). For instance, Brown and Cliff (2005) state that asset prices are strongly affected by investor investment in the long-run. Namely, investors that are overly optimistic drive asset prices above fundamental values of these assets, and that these errors in pricing tend to revert over a horizon of multiple years. Different asset prices, reflect sentiment in different markets. A rapid upward shift in equity prices indicates an increase in sentiment of investors in companies. A rapid upward

(12)

shift in real estate prices indicates an increase in sentiment of households or firms regarding the value of property. This suggests that including both house- and equity prices would be the better at incorporating sentiment in general, rather than separately.

Furthermore, even though credit might be an essential component of financial crises, and cycles more generally, it is not as clear whether it is enough. Gorton and Ordonez (2016), for instance find that not all credit booms lead to a financial crisis. In this sense, credit’s role in capturing financial cycles could be complemented by asset prices when arguing through the ‘balance sheet’ channel. As Bernanke and Gertler (1999) put it, credit market frictions imply, amongst other things, that an agent’s balance sheet is an important determinant for his ability to lend and borrow. From this perspective, because assets are used as collateral and determine the level of an agent’s leverage, changes in asset prices adjust an agent’s net worth and consequently affecting the scale of lending and borrowing. Many studies on leverage cycles have been centred on this interaction (e.g., Gromb & Vayanos, 2002;

Geanakopolos & Fostel, 2008; Adrian & Shin, 2010; Schüler et al, 2017).

Therefore, I supplement credit in my credit gap estimator with a composite asset price indicator which includes house- and equity prices. The relevance of house prices has for instance been stated by Iacoviello (2005), who integrated constraints on collateral which are dictated by house prices into a monetary business cycle model because previous findings suggest that a significant part of borrowing is secured by properties (another example is Justiano, et al., 2015). What is more, Jordà, Schularick, and Taylor (2015a, 2016) also state that financial stability risks are closely linked to developments in mortgage lending. The importance of equity prices has, for instance, been noted by Jordà et al. (2015b), who provide empirical evidence that the majority of build-ups of systemic risk after World War II included both house- and equity prices. Furthermore, Claessens et al. (2011, 2012) find empirical evidence that recessions where equity prices, together with house price go bust are longer and deeper.

These findings suggest that a joint analysis of credit and other asset prices might be necessary to aptly characterise the credit gap, and that the BCG therefore falls short in doing so (as it only includes credit and GDP). As noted by Jorda et al. (2015b), this more comprehensive set of asset prices (as opposed to only house prices) might be crucial in identifying risks, also on the global scale (Breitung & Eickmeier 2014; Rey 2015). While the inclusion of asset prices in the characterisation of credit gaps has been done before, it has not been done through estimating a real-time structural Vector Error Correction (VEC) model that includes both equity and house prices. As will be discussed in more detail later, this type of model allows for the leveraging of the endogenous feedback effects that credit and asset prices have on one another to estimate a credit gap. In light of the discussion in section II, this might result in more robust credit gap estimates.

What is important for policymakers, however, is not only the theory explaining what credit cycles are and why they exist. What is also important is accurately measuring this credit cycle, as this is needed to tailor counter-cyclical policy. Section III discuss more in depth what the current state of the art is regarding these measures. First, I will briefly discuss the Basel Credit Gap, the most popular and wide-spread credit gap measure within the policy sphere, and its limitations. In the second half I will discuss the academic literature on other measures for measuring the credit gap.

(13)

III. Literature Review on Measuring the Credit Gap

In this section, I will first discuss what the BCG is, and what its most important limitations are.

In the second half of this section, I will discuss the literature on credit gap measures more generally. In particular, I split the literature up into two groups: purely statistical measures, and structural measures. Structural measures have as a distinct competitive advantage compared to statistical measures that they allow for the estimation of the credit gap as a function of economic fundamentals, such as asset prices. Also, there is no need for preliminary assumptions on the length of the credit cycles, as is the case with many other statistical measures.

The Basel Credit Gap

The BCG rose to prominence after the introduction of the Basel III reforms. These Basel III reforms were introduced as the new global standard on bank capital adequacy and liquidity by the Basel Committee on Banking Supervision (BCBS) in 2010. Part of these reforms was a new macroprudential toolkit, which included measures such as the CCyB. These macroprudential measures were designed to help lean against the credit cycle. To do this, these macroprudential measures should be activated and increased during the boom phase of the credit cycle, and released during downswings to allow banks to absorb losses without cutting back on lending (Baba et al., 2020).

To guide the actual implementation of these new measures, in particular the CCyB, the BCBS recommended as a common reference point to use the BCG, which is defined as the difference between the actual credit-to-GDP ratio and the long-term Hodrick Prescott (HP)- filtered trend. This specific indicator was chosen due to it performing well in predicting banking crises early on (BCBS 2010; Baba et al., 2020). According to the guideline, the CCyB should be activated at the point the BCG is above a certain threshold, and continue to rise until a maximum value of the BCG is reached. The CCyB should be released again once it reaches a minimum value. Thus, the BCG should be able to estimate the credit cycle correctly both in its boom and bust phase. However, the early warning properties only reflect the ability of the BCG to predict relatively rare financial crises events. The BCBS (2010) also acknowledge that the BCG does not function properly in all countries, or in every time period, and calls for expert judgement to be a major component of the measures (Baba et al., 2020).

The calculation of the BCG is the difference between a country’s actual credit-to-GDP ratio and its long-run HP-filtered trend. Credit is defined as total credit to the non-financial private sector. This includes credit from foreign entities and non-banks. The estimation of the long-run trend is done recursively, and it is as such called a ‘one-sided’ HP filter. The smoothing parameter (lambda) of the HP filter is set to 400,000 and is applied to quarterly data. This means that the underlying assumption of the credit cycle’s length relative to the business cycle is four times as long (Drehmann et al., 2012; Jokipii et al., 2020; Baba et al., 2020).2

Several of the BCG’s limitations have also come to light over time (for instance, see Hamilton 2018, ECB 2019, Jokipii et al., 2020, Baba et al., 2020, or Drehmann and Yetman 2020). I will set out the five most important limitations. The first two issues are related to the BCG in particular, and the three others are related to HP-filters in general.

2 This is explained in section IV.

(14)

Main Limitations of the BCG

The first issue with the BCG in particular is the normalisation of credit by GDP (Repullo &

Saurina, 2011; Drehmann & Tsatsaronis, 2014; Baba et al., 2020; Jokipii et al., 2020). As credit is a stock variable, and GDP is a flow variable, the former adjusts more slowly than the latter.

This leads to a negative correlation between GDP growth and the BCG. In other words, ironically, the BCG itself is procyclical. As a result, the credit gap can not only turn positive when credit grows relative to its long-term trend, it can also turn positive in an economic downswing. The BCG might therefore signal to macroprudential policymakers to activate macroprudential measures which would aggravate the economic downturn. This is the opposite of the goal of these macroprudential policies. This particular issue is also relevant for the Netherlands, as figure 1 shows. During the Sovereign Debt Crisis, when GDP growth was relatively low, the Dutch credit gap reached its highest level of the entire data sample.

The second issue is with the assumption on the relative length of the credit cycle to the business cycle (Galan, 2019; Baba et al., 2020). With the smoothing parameter set to 400,000, the assumption implicit in the BCG is that the credit cycle’s length is four times as long as the business cycle. However, empirical evidence seems to suggest that the actual relative length is not only shorter than that, but it also varies over time (Galan, 2019; Baba et al., 2020). Drehmann et al. (2012) state that the reason this large smoothing parameter was chosen, is that it improved on average the early warning properties of the BCG. However, as Baba et al. (2020) show, in times of a long boom phase of the credit cycle, the BCG seems to follow the actual data with several lagged years. As such, the BCG might fail to signal building imbalances after a prolonged boom phase. This is possibly also prevalent for the Dutch case (see figure 1), where after the (wrongfully) indicated boom phase during the SDC, the credit gap was historically low. This meant that in the time when both equity and house prices rose considerably, both signalled that vulnerabilities were building up in the financial system.

Interestingly, Galan (2019) finds that for the case of Spain, adjusting the smoothing parameter downwards to 25,000 improves the predictive properties of the BCG. As such, I will also test this assumption later on in my robustness checks.

The other three main issues concern HP-filters in general. The first of these is the so- called ‘downward bias’ (Lang & Weltz, 2017; Lang et al., 2019; Baba et al., 2020; Jokipii et al., 2020). It is related to the second BCG-specific limitation. Due to its mathematical properties, a boom phase of the credit-to-GDP ratio is integrated into the trend estimation. This then elevates the trend estimation, perhaps more than it should. This means that after a prolonged boom, the trend estimation during the deflation and normal period would be higher than it would have been without this preceding boom. As such, the credit gap is then biased downward.

Second is the start-point problem (ECB, 2019; Baba et al., 2020; Jokipii et al., 2020).

What this means is that the start of the time series has significant implications for the size of the BCG. This is also the reason why the BCBS recommend to start using the BCG only when there is a minimum length of the data sample of ten years. Empirical evidence seems to suggest, however, that this is longer. Jokipii et al. (2020) show for instance that for the Swiss case it takes around 30 years for the different start points to converge. I evaluate this assumption for the Netherlands also later on in this study and find that it takes around twenty-five to thirty years for the BCG estimations to converge after a different starting point.

(15)

Third is the end-point issue (Edge & Meisenzahl, 2011; Baba et al., 2020; Jokipii et al., 2020). This means that the end-point of the sample has a strong influence on the estimate of the underlying trend. It must be noted, however, that these issues are prevalent in all statistical measures. Nevertheless, the ECB (2018) find that while the BCG estimates the phase of the financial cycle relatively correctly, it underestimates the boom-bust sizes.

As a response to these weaknesses, many authors have proposed additional measures to measure the credit gap. Those measures most influential and relevant to this study will be discussed in the next section.

Literature Review on Measuring the Credit Gap

After the Great Financial Crisis (GFC) and the creation of novel macroprudential regulation with a focus on using cyclical instruments, the literature on measuring the credit cycle and indicators for early warnings of systemic crises has elevated in its importance and size. The literature can largely be divided into two groups. First are statistical approaches, where mostly filters are used to obtain a cyclical and a trend component from credit data itself (univariate) or with the inclusion of other variables (multivariate). Second are structural approaches, where a model based on certain underlying credit fundamentals is estimated.

Statistical Approaches

Much of the statistical literature revolves around the use of univariate filters (Drehmann et al., 2012; Detken et al., 2014; Richter et al. 2017; Drehmann & Yetman 2020, 2018; Jokipii et al., 2020). The most widely used univariate filter is the BCG, as this was the initially proposed method by the BCBS (2010) for measuring the credit gap and the deployment of macroprudential measures such as the CCyB. Other filtering approaches that are in use as well are the Baxter and King (BK) (1999) and Christiano and Fitzgerald (CF) (2003) filters (Aikman et al. 2015). These two filters allow for a range of frequencies and are therefore more flexible than the HP filter. But these two filters also supply a substantial range of credit cycle estimations depending on the assumption of the relative length of the credit cycle.

In this sense, the first studies which supported the Basel Credit Gap (BCG) for measuring the early warning for systemic crises (BCBS, 2010; Drehmann et al., 2010), were very quickly followed by studies that found many shortcomings (Edge & Meisenzahl, 2011;

Repullo & Saurina Salas, 2011; Salas, 2011; Castro et al., 2016; Hamilton, 2018). Most of its limitations were found to be tied to the statistical nature of the BCG. Its most notable critic is Hamilton (2018). The main fault that he finds is that the HP-filter (and indirectly other filters also) produce spurious dynamics. And therefore, he argues, it should never be used. Rather, he proposes a linear projection approach where the trend value of variable y at t + h is defined as the value that the value of y would have been estimated to be at time t based on the past x values. This would result much less spurious dynamics and it greatly diminishes the end- point issue of the HP-filter.

However, the Hamilton approach estimates are very volatile and have frequent sign changes. This makes it very difficult for macroprudential-policymakers to implement.

Drehmann and Yetman (2018, 2020) criticise the approach also for the fact that it has much weaker predictive power for financial crises than the BCG. They acknowledge the weaknesses of the BCG, but still find that “..in the absence of clear theoretical foundations, any proposed gap measure is nothing more than an indicator, including when derived with more

(16)

sophisticated empirical methods.” (Drehmann and Yetman, 2020, p. 3). They then proceed to test a whole range of univariate methods on their predictive ability for banking crises, and find that overall, the BCG is still the best univariate indicator available. Jokipii et al. (2020) also come to the same conclusion for the case of Switzerland, although they stress that more indicators should be used to capture the entire credit cycle.

This is, however, not as easy as it appears. As noted by Alessi and Detken (2011), no indicator performs clearly better than the rest across many countries and over a longer time span. What is more, it is often the case that well-performing indicators provide conflicting signals in certain situations. Babecky et al. (2012), then, propose to combine indicators to improve their early warning properties. They base their findings on a large set of indicators of a sample of 40 countries. Regardless, these signals are only reliable at short time horizons.

This might not be sufficient time for macroprudential policymakers to act effectively. Lang et al. (2019) therefore propose aggregating a large set of individual variables into one large composite indicator. They find that this improves the early signalling properties of the BCG and other individual indicators. However, a major weakness of this study is that this indicator’s performance is not assessed in real-time. Nor do the differences between the individual indicators, the BCG, and the composite indicators appear to be statistically significant. This puts into question the findings’ reliability. These are weaknesses that are prevalent in the majority of the previous studies on financial vulnerability indicators. To my knowledge, only Galan and Mencia (2021) assess all of them fully.

Some have also proposed the idea of designing early warning systems based on the information of a set of indicators. Coudert and Idier (2016) suggest taking the average of a set of logit models which include a multitude of macro-financial indicators and the combination with their transformations. Their findings suggest that it increases the indicators’

performance in terms of the early warning properties. Alessi and Detken (2018) suggest using a non-parametric method. Specifically, they advocate a system of decision-trees that allows for the identification of relevant thresholds of a large group of macro-financial indicators which indicate cyclical risk building up.

Interestingly, in most of these recent articles, the choice of the smoothing parameter is not put into question, even though this would significantly decrease the downwards bias prevalent in the BCG. The choice of the smoothing parameter determines what the author believes the relative length of the credit cycle to the business cycle to be. A lower smoothing parameter would mean a relatively shorter length of the credit cycle. The majority of studies assess their new model’s predictive properties versus the standard Basel Credit Gap with a smoothing parameter of 400,000. A recent exception is Galán (2019), who assesses the predictive properties of a variety of filters and other methods for determining the credit cycle for Spain, and finds that the HP-filter with a lower lambda of 25,000 actually improves on the predictive properties compared to the lambda of 400,000 while also having a significantly lower downwards bias (by design, of course). He then concludes that the BCG, adjusted for a lower lambda, is still the best indicator for Spain. Considering these findings, a proper evaluation of a new model’s usefulness should be done whilst similarly assessing whether the assumption underlying the choice of the smoothing parameter of the HP-filter is empirically sound.

All these methods claim to improve on the BCG. However, they also are all purely statistical methods that do not provide economic interpretations of credit deviations from its long-run trend. And this is crucial for a policymaker to be able to distinguish between benign- (financial deepening) and bad deviations (excessive credit growth) from its trend.

(17)

Furthermore, policymakers are also faced with Goodhart’s Law (1975): if policymakers base their policy decisions on a certain variable – such as the BCG – then the relationship with this variable and the end objective (flattening the financial cycle) is likely to become unstable. For instance, banks are likely to redefine their credit provision (such as through disintermediation) in order to reduce the penalty they are levied for credit growth. Models based on structural relationships originating from economic theory, can mitigate these two issues, as these structural relationships between variables can be stronger than purely statistical relationships. These models will be discussed below.

Structural Methods

A different way to signal systemic financial vulnerabilities is by estimating credit’s trend by regressing credit on its fundamental determinants, and using the regression’s residuals as the identification of the credit gap. Compared to the purely statistical approaches mentioned above, a model-based method can explicitly capture components that alter credit’s trend across countries and over time. These fundamental can include, for instance, changes in the real interest rate, income levels, financial reforms, and changes in asset prices. This can improve the economic robustness of the credit gap estimations. A gap in this literature is that few of these structural models have been estimated real-time, so as to evaluate their relevance for policymakers and properly compare them to the BCG (which is also estimated in real-time). Due to the lack of this type of studies, I lay out structural methods for estimating the financial cycle more generally first, and zoom in on studies that estimate models in real- time later on.

The most prominent group of (semi-)structural methods is the implementation of parametric filters based on the use of uni- and multivariate Unidentified Components Models (UCMs) (Galan & Mencia, 2021). These have the advantage of avoiding making any assumptions on the (relative) length of the financial- and business-cycle. Although the use of the Kalman filters has been widespread in the business cycle literature, its application to financial variables is much more recent. Lucas and Koopman (2005) pioneered this part through disentangling credit and business cycles through a UCM. Chen et al. (2012) expand on this method to model cyclical patterns between output, interest rates, asset prices, and credit. They find that all these variables are highly correlated in the long run. Galati et al.

(2016) use this approach for several euro area countries and the US and find that credit and real estate prices have very similar cycles, but that the duration and amplitude of these cycles vary substantially among countries. Rünstler and Vlekke (2017), again using a similar method, similarly find long and large cycles for both credit and real estate prices, which seem to be highly correlated to GDP. Most definitely, credit and real estate prices have been found to have cyclical similarities (Claessens et al., 2012; Schüler et al., 2015; Jorda et al., 2015). The studies also find substantial heterogeneity between the amplitude and duration of cycles across economies. This is mostly related to the different specificities of the national housing markets (Galan & Mencia, 2021).

Another structural approach to identify the relation between financial and macroeconomic variables that has recently come to the fore is the use of Vector Autoregressive (VAR) models. The particular advantage of this type of models is that it leverages endogenous relations between the variables that are included. This makes it especially relevant in light of the earlier discussion on the theories and empirical facts of the credit cycle. There, I laid out why credit and asset prices most likely strengthen each other’s

(18)

movements, creating larger financial vulnerabilities in the process. And much of the VEC literature does find strong relations between credit, GDP, and other macro-financial variables.

Goodhart and Hoffman (2008) take a VAR model to characterise the role of house prices in explaining output and credit, and find that after a significant house price bubble inflation credit is also increased. Hristov et al. (2012) and Duchi and Elbourne (2016) look more closely at the influence of credit shocks on GDP. They find that there is a strong long-run relation between business and financial variables. Juselius et al. (2018) put forth the idea of characterising the financial cycle through a Vector Error Correction (VEC) model which is a VAR model with a long-run component in its specification. They suggest that this might account better for long-run relations between credit, output, asset prices and interest rates.

They indeed find that the long-run relation between debt service and leverage drives the length and size of financial cycles. They do not take this model further, however, to estimate the financial cycle, nor do they do this real-time.

Clearly, UCM and VEC models have been used to characterise the relationship between credit, other financial variables, and financial crises. However, they have rarely been used as a real-time early warning indicator for systemic crises, let alone assessing whether these real-time indicators are also good measures for the entire length of the credit cycle. To the best of my knowledge, the only authors that used either a UCM or a VAR/VEC model to predict systemic crises are the recent articles of Lang and Welz (2018), and Galan and Mencia (2021). Baba et al. (2020) do not assess their measure’s predictive properties, but do assess whether their credit gap measures are a good approximation of the entire credit cycle. Lang and Welz (2019) focus on household credit and propose to model its trend based on several fundamentals which are institutional quality, population, long-term real interest rates, and GDP. Their findings suggest that the extracted cycle is less persistent than the BCG. Yet, because credit to the entire nonfinancial private sector is excluded from their model, periods of excess credit growth which are not correlated enough to household credit can be missed.

This makes the measure less relevant for the CCyB. What is more, they do not include house prices in their analysis. This is troubling for two reasons. First, mortgages are a very important component of household credit. And second, a strong relationship between house price cycles and credit has been identified in the literature (Drehmann et al., 2012; Claessens et al., 2012; Borio, 2014; Schüler et al., 2015; Galati et al., 2016).

Baba et al. (2020) and Galan and Mencia (2021) each propose both a UCM and a VECM for estimating the credit gap. Baba et al. (2020) first suggest using a semi-structural method based on several key economic theories, such as the Taylor Rule and the Phillips Curve, estimate the model on a per-country basis, and graph the resulting Kalman Filter estimates of the credit gap. Their second suggestion is estimating a VEC on an unbalanced panel of 40 advanced economies ranging from 1970 to 2018, where the residuals for every country would indicate its respective credit gap. The underlying fundamentals for the credit gap are the old- age dependency ratio, the real interest rate and GDP. While they do not mention it, the second model implicitly assumes the presence (and dominance) of an international financial cycle, as discussed in Rey (2020). That would be quite a step away from the majority of financial and credit cycle measurement studies undertaken up to this point, as most focus on domestic financial cycles. They conclude that for both their specifications, the downwards bias is less than that of the BCG, and that they would be of added value to the BCG for indicating the entire credit cycle. One weakness of this study is that they do not explain at length what the implications are for the estimations of the credit cycle for each country, even though their results differ quite significantly per country. Nor do they estimate their models

(19)

in real-time. They only estimate their models around 2008, to see what the models would have indicated the credit gaps position to be at that time. This diminishes the added value of their findings. Lastly, asset prices are not taken into account in their models. Considering the earlier theoretical and empirical discussion on the empirics of the credit cycles, this could mean that there is significant omitted variable bias in their estimations.

Galan and Mencia (2021) also propose a UCM and a VECM. For both models, house prices and real interest rates are the major variables. They then estimate these models on a per-country basis for a panel of seven European countries from the period of 1970 to 2016.

Using the AUC, they then rank the best performing models. They find that the performance of the models differs per country, but that they consistently outperform the Basel Credit Gap.

For the Netherlands, the VECM was the best model and the best specification included house prices. However, the study has several components that can be expanded on. First, they only include house prices, but do not include equity prices into their specifications. Given the recent findings of Greenwood et al. (2020), and the earlier discussion on the importance of sentiment in the credit cycle, also adding equity prices can prove to be of added value for the credit gap estimation. Second, they do not evaluate the model for bank credit only, which would allow for a more precise localisation where signalled imbalances might be coming from. And third, they do not analyse their results more deeply for each country. They primarily use the AUC-value to base their conclusions on. As discussed before, only basing the conclusion on how well a vulnerabilities indicator performs on the predictive properties of relevant stress events, leaves out the consideration whether the indicator will actually signal vulnerabilities correctly over the entire credit cycle. These are all considerations that will be taken on in this study.

VEC models allow for delving deeper into endogenous relations between a set of variables and their lags. Thus, they allow for analysing relations between variables across time. Considering the earlier discussion in section II on the relation between credit and asset price cycles, this seems particularly interesting. It is therefore surprising that the literature leveraging these models to produce real-time estimates for the credit cycle is so thin.

Nevertheless, the studies that do produce interesting results. As such, it would be of added value to the literature to evaluate a VEC model with credit, asset prices and several other fundamentals to see whether this model improves on the credit cycle estimation compared to the BCG, and is of potential added policy value. In section IV, I will discuss how I do this.

IV. Methodology

Model Set-up

As stated before, a structural model allows for the explicit modelling of the connections between credit fluctuations and other macro(-financial) fundamentals. Furthermore, it offers a direct economic interpretation of the model’s estimates, which might aid in the decision- making of policymakers. It also does not require any prior assumptions on the duration of the financial cycle, which has been shown to vary significantly across time and countries (see, for instance, Drehmann et al., 2012, and Schüler et al., 2015).

In particular, I follow Galan and Mencia (2021), and propose to estimate a VEC model, where several of the variables are cointegrated. This model allows for the estimation of a common long-run stochastic trend of the variables. The main specification is as follows:

(20)

! ΔC ΔY ΔAP

ΔR (

!

= 𝜹 + ∑$%&"'& Γ" ! ΔC ΔY ΔAP

ΔR (

!%"

+ 𝜶𝑢!%& + 𝜀! ; (1)

𝑢!%&= 𝝁 + 𝜷 !

C Y AP

R (

!%&

The log of real credit to the non-financial private sector, is represented by C!. This is the primary variable of interest. I prefer this measure of credit for three reasons. First, the BCG uses this definition, and thus allows for a more proper comparison of the different models under similar conditions (BCBS, 2010; Drehmann et al., 2012; Baba et al., 2020; Jokipii et al., 2020; Galan & Mencia, 2021; BIS, 2021). Second, the goal of the majority of the macroprudential measures is to flatten the financial cycle and mitigate systemic risk caused by excessive credit growth from any source (banks and non-banks) and to any type of borrower (households and non-financial corporations). Thus, taking credit to all sectors into account would most align with the goal of the macroprudential measures. Third, although bank credit to households, and other more narrow credit measures, have been found to be drivers for previous financial crises in several economies, this has not been so in all economies and it is not guaranteed that that future crises will be driven once more by these specific measures. What is more, prudential regulation is oftentimes sharpened concerning the sector that played a significant role in a financial systemic crisis. Due to this increased regulation the relation of the credit sector to systemic risk could be broken, and the value of an indicator focused on that sector could diminish. These issues are alleviated by using a broader measure.

The log of real GDP is represented by Y!. I include this variable for two reasons. First, it proxies the economy’s opportunities and capacity to borrow (and repay) and affects both the demand and supply of credit positively (Duchi and Elbourne, 2016; Baba et al., 2020).

Second, it enhances the comparability of the VEC model estimations with the BCG, as GDP is the BCG’s denominator.

Asset prices, either in logs or levels, is represented by AP!. This can be real house prices, real equity prices, or a composite indicator of both. The composite indicator of both would be a simple addition of both, so as not to complicate the calculations. It also does not shift the indicator’s weight into one direction, to reduce bias in favour of an asset group as much as possible. The reasons for including this, as discussed at more length in section II, is twofold. First, asset prices proxy for sentiment within (credit) markets. Asset prices carry information regarding sentiment of investors. When sentiment is higher, investors regard the future as more favourable and will be more willing to invest. This would eventually drive credit upward. Presumably, this is more so for equity prices than for house prices. As the largest share of financiers of equity are professional investors, these prices reflect mostly what investors expect to be the discounted future stream of profits from their investment.

For house prices, of which the largest consumers are non-professional consumers, the buyer’s expected future stream of income also plays a major role in their investment decision. As such, it dilutes the information component on sentiment within house prices. While credit also carries information on sentiment, asset prices adjust to shifting sentiment more quickly

(21)

than the stock variable credit. Furthermore, financial crises are more likely to happen when both credit growth and sentiment are high, so including an extra variable to proxy sentiment would be beneficial in characterising the financial cycle (Reihnart & Rogoff, 2012; Schüler et al., 2017; Greenwood et al., 2020; Krishnamurthy & Muir, 2020). Second, asset prices influence credit (and indirectly GDP) through the ‘balance-sheet’ channel. That is, assets can be used as collateral for credit, next to it being a source of revenue or service streams.

Changes in asset prices change the borrower’s net worth e.g., through the asset’s use as collateral and its associated influence on leverage, thereby affecting the scale of borrowing and lending (Gromb & Vayanos, 2002; Geanakopolos & Fostel, 2008; Brunnermeier &

Pedersen, 2009; Adrian & Shin, 2010; Geanakopolos, 2010; Schüler et al., 2017; Greenwood et al., 2020).

The long-term real interest rate (from here on out the ‘real interest rate’) is represented by R!. It captures the relationship between debt and its interest payments. In essence, lower interest payments allow firms and households to maintain same stock of debt with lower income in the long run. More generally, the real interest rate positively affects credit demand and negatively affects credit supply. Most studies using a VEC model in relation to the financial cycle integrate this variable (e.g., Borio et al., 2016; Baba et al., 2020; Galan &

Mencia, 2021). However, there is a debate on whether the real interest rate is on a persistent downward path (Carvalho et al., 2016; Borio et al., 2016; Schmelzing, 2021), or that this is simply a temporary phenomenon, and they are actually stable (Blanchard, 2021). As such, I will estimate the majority of my specifications both with and without the real interest rate.

A VEC model would account for the endogeneity between these variables, which would characterise the possible feedback loop between credit, GDP, the real interest rate, and asset prices. The rest of the parameters are the following. 𝜹 is a vector of the intercepts of each equation. 𝜶 represents a matrix of adjustment coefficients of long-run deviations containing one vector for each of the variables of dimension m, where m represents the number of cointegrating relations. 𝝁 is a vector of constant parameters. 𝜷 represents a matrix of parameters composed of four vectors of dimension m. Γ" represents p-1 vectors of parameters associated with the lagged underlying variables, where p is the lag order of the VAR in levels. Δ is the first difference of the variables.

Other specifications

Next to this main specification are four other components of interest to include or exclude:

different sources of credit, disaggregating credit by the type of borrower, including variables in nominal terms, and altering the trend specification of the model. All will be briefly discussed in this section.

Changing the credit variable to include credit from banks only is of additional value to macroprudential policy makers. The reason is that the majority of macroprudential tools is targeted towards banks. Prominent examples are the CCyB or the maximum Loan to Value (LTV) ratio on mortgage credit. While the model with credit from all sectors might indicate that there are financial imbalances building up in the economy, this might come from financial intermediaries other than banks. If so, putting into action the CCyB might not be effective, and perhaps even damaging to banks (as it puts them at an even further competitive disadvantage). If a stable measure can be found also for banks, this can guide policymakers in seeing where the vulnerabilities might be coming from.

Referenties

GERELATEERDE DOCUMENTEN

The household-indebtedness channel argues that due to the relation between credit creation for unproductive purposes and asset prices, households have to take increasing loans to

To that end, future research should examine the effects of school interventions directed at either improving pupils’ attitudes towards epistemic curiosity or their implicit

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence.. Any further distribution of this work must maintain attribution to the author(s)

U-Multirank onderscheidt zich van de bestaande rankings door zijn diversiteit.. Zo reikt de geografische focus verder

Keywords: Corporate Social Responsibility (CSR), Corporate Political Activity (CPA), complementarity, CSR-CPA complementarity, Government Dependence, Corporate Financial

Additionally, just as Voelkl (2012) questioned whether the responses she received from her participants were altered due to the fact that she was a white woman, I question whether

The difficulty the Harlem Renaissance writers and their protagonists felt when they attempted to create an identity for themselves can in part be credited to the racial boundaries

Om echter goed te kunnen begrijpen hoe goedwillende mensen zich voortdurend proberen aan te passen aan alle eisen die aan hen gesteld worden, moet u ook iets weten over hoe