BIS Working Papers
Measuring financial cycle time
by Andrew Filardo, Marco Lombardi and Marek Raczko
Monetary and Economic Department
JEL classification: demography, ageing, inflation, monetary policy
Keywords: E31, E52, J11
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)
Measuring financial cycle time
by Andrew Filardo, Marco Lombardi and Marek Raczko
Motivated by the traditional business cycle approach of Burns and Mitchell (1946), we explore cyclical similarities in financial conditions over time in order to improve our understanding of financial cycles. Looking back at 120 years of data, we find that financial cycles exhibit behaviour characterised by recurrent, endogenous swings in financial conditions, which result in costly booms and busts. Yet the recurrent nature of such swings may not appear so obvious when looking at conventionally plotted time-series data (that is, observed in calendar time). Using the pioneering framework developed by Stock (1987), we offer a new statistical characterisation of the financial cycle using a continuous-time autoregressive model subject to time deformation, and test for systematic differences between calendar and a new notion of financial cycle time. We find the time deformation to be statistically significant, and associated with levels of long-term real interest rates, inflation volatility and the perceived riskiness of the macro-financial environment. Implications for statistical modelling, endogenous risk-taking economic behaviour and policy are highlighted.
Key words: Financial cycles, continuous-time autoregressive models, time deformation, behavioural economics, endogenous risk-taking behaviour, central banking
JEL codes: E32, G01, F32, F34, E58, E71, D80
The Great Financial Crisis (GFC) reinvigorated interest in studying financial cycles. One strand of the recent research has focused on understanding the domestic origins of wide swings in financial conditions, namely credit and housing prices. Risk-taking behaviour, in particular, has been emphasised as a key channel that transforms supportive financial conditions into boom-bust cycles (see eg Borio (2013)). Another strand has looked into understanding the global dimension of financial cycles. Waves of global liquidity have been found to interact with domestic financial cycles in ways that generate excessive procyclicality in financial conditions (see eg Claessens et al (2011), Bruno and Shin (2015)).
Despite considerable efforts over the past decade, researchers and policymakers have yet to forge a workhorse model of financial cycles. Modelling efforts have emphasised the behaviour of banks and non-banks as well as the role of financial frictions (see, eg, Linde (2018)). These frictions are seen as amplifying, amongst other things, exogenous economic and financial shocks. These shocks, under certain circumstances, can result in financial fragility and destabilising financial headwinds which, if large enough, can turn into financial crises. Many proposed models, especially those built on the bedrock of the dynamic stochastic general equilibrium approach, offer interesting insights about particular aspects of boom- bust developments. But no model has broken out of the pack to provide a comprehensive understanding of financial cycles. One reason for this failure is that there remains considerable disagreement about the key mechanisms behind the recent crisis. Some analysts argue, for example, that the GFC was the result of a unique set of developments that exposed weaknesses in the financial regulatory system. Others point to broad similarities with other boom-bust experiences and argue that these similarities suggest boom-bust cycles are an inherent feature of capitalist economies.1
Our paper takes a different, but in many ways complementary, approach. We apply the ideas of the classical business cycle literature of Burns and Mitchell (1946) to the financial cycle. Namely, we treat financial booms and busts as endogenous, recurrent phenomena with common features shared across time rather than as unrelated exogenous one-off unique events. Burns and Mitchell ask the fundamental question, "what happens during business cycles?"; in a similar vein, we ask "what happens during financial cycles?". This question presupposes that financial cycles share certain commonalities across time that make them, in many respects, all alike.
Having traced the roots of our approach to that of Burns and Mitchell (1946), we find it useful to start by framing the analysis with Burns and Mitchell's famous description of the business cycle:
"Business cycles are a type of fluctuation found in the aggregate economic activityof nations that organize their work mainly in business enterprises: a cycle consistsof expansions occurring at about the same time in many economic activities, followed by similarly general recessions, contractions and revivals which merge into the expansion phase of the next cycle; this sequence of changes is recurrent but not periodic; in duration business cycles vary from more than one year to ten or twelve years; they are not divisible into shorter cycles of similar character with amplitudes approximating their own."
While the description applies to business cycles, it could be considered apt also for financial cycles. Paraphrasing the quote, financial cycles are a type of fluctuation found in aggregate financial conditions of nations that organise their intermediation in financial markets and financial intermediaries such as banks: a cycle consists of expansions occurring at about the same time in many measures of financial activity, followed by similarly general recessions, contractions (which sometimes become crises) and revivals which merge into the expansion phase of the next cycle; this sequence of changes is recurrent but not periodic; in duration financial cycles vary from more than eight years to 20 or so years; they are not divisible into shorter cycles of similar character with amplitudes approximating their own.
It is important to highlight key differences between financial cycles and business cycles. Graph 1 illustrates the empirical nature of the differences between the two cycles. Statistical features of the financial cycle obviously differ in various respects. First, financial cycles comprise expansions and contractions of financial conditions rather than of the real side of the economy. In this graph, the proxy for the financial cycle is an aggregate of credit and housing prices. Second, the duration of financial cycles varies and is generally longer than the duration of business cycles. Third, the amplitude of the financial cycle varies more than the business cycle. Fourth, the synchronicity of financial and business cycles appears fairly low. The business cycle does not appear to drive the financial cycle. However, there is a relationship between financial bust phases and serious macro-financial disruptions.
These differences should not be interpreted as suggesting that financial cycles have little to do with business cycles. Indeed, cyclical credit conditions played a role in traditional business cycle analysis. Burns and Mitchell (1946), Moore (1961), Eckstein and Sinai (1986) and Zarnowitz (1992) emphasised the dynamics of credit conditions and asset prices when describing the evolution of the stages of the business cycle.2 At business cycle frequencies, cyclical credit conditions and business cycles were consistently found to be interrelated. The more recent literature on the macro-financial linkages during business cycles has highlighted the role of the so-called financial accelerator (Kyotaki and Moore (1997), Bernanke et al(1999)) and financial frictions (Hall (2013), Duval et al (2017)). However, as pointed out by DeLong and Summers (1986) and Zarnowitz (1992), the historical relationship between financial panics and business cycle contractions is not as close as sometimes thought.
Our study focuses on the longer swings associated with financial cycles, a phenomenon plainly visible in Graph 1. Various features of the financial cycles are notable. The financial cycle does not coincide with the business cycle. So we are not focusing on the amplification mechanism of the conventional financial accelerator but capturing phenomenon of the type described by Borio et al (2012) and Schularick and Taylor (2012). Financial cycles are generally of a longer duration (or, in other words, they are a lower frequency phenomenon) than business cycles, and the duration (and amplitude) changes over time.
In addition, when looking at measures of the financial cycle over an even longer time span, the financial cycles do not appear to be solely regime-dependent, at least in some respects. Since the late 1880s, for example, various exchange rate, monetary, fiscal and regulatory regimes have been in place. This is not to say that these regimes, amongst other things, did not influence the shape of the financial cycle. Certainly, periods of financial repression, for example, tended to influence the shape of the financial cycle (see, eg, Burnside et al (2016)). But through it all, recurrent long swings in financial forces are evident.
One challenge in exploring the financial cycle's recurrent nature is the ostensible differences of each cycle over time, as is evident in Graph 1. Conventional time-series plots of financial cycles, on the face of it, suggest that the financial cycles are not at all alike. However, as Burns and Mitchell (1946) emphasised when researching business cycles regularities, conventional time-series plots (in calendar time) may obscure some of important regularities that all cycles share. Put differently, calendar time might not be the best way to compare cycles over time.3
Indeed, Burns and Mitchell advocated a phase-centric approach to characterising the regularities of business cycles. They did this by defining units of cyclical time as the sequence of phase turning points. So, for example, the expansion phase of a cycle (from trough to peak) would represent one unit of cyclical time, regardless of the elapsed calendar time in that phase. They then summarised changes in economic variables in each of the phase-based time units and assessed the co-movements. Essentially they were stretching and compressing the time series in calendar time so as to transform the data into business cycle time. This made it possible to establish clearer stylised facts about the regularity of business cycles. Much of the research that followed from this conjecture emphasised a unique cyclical, rather than calendar, time in understanding business cycles.
Using analogous methods, the aim of our paper is to shed light on the nature of the factors associated with, and the forces behind, the financial cycle. In addition to using the graphical analysis of Burns and Mitchell, we apply the pioneering statistical approach of Stock (1987) in order to assess the extent of the differences between financial cycle time and calendar time, ie technically, we call this time deformation in the financial cycle. The statistical approach allows us to characterise the regularities of financial cycles that may arise from differences between calendar time and financial cycle time.4 Technically, we demonstrate how to jointly estimate a continuous-time autoregressive process and the parameters of the non-linear function that describe the relationship between financial cycle time and observable macrofinancial variables.5 The model allows statistical testing for the presence of time deformation and the statistical significance of the related macro-financial variables driving it. We find evidence that low real interest rates, inflation volatility and macro-financial risk environments are associated with financial cycles of longer duration.
Various potential benefits arise from being able to characterise financial conditions in financial cycle time rather than calendar time. From a modelling perspective, if fluctuations in financial conditions exhibit more stable statistical behaviour in financial time than in calendar time, forecasters can take advantage of this timing information and, in principle, produce more accurate predictions of both financial conditions and the conditional probability of a boom ending in a bust. From a more economic point of view, knowing that financial cycles are all alike in financial cycle time, analysts may be able to make more reliable comparisons with past cycles and hence make better inferences about current and future financial conditions. Policywise, if it is the case that financial cycles are recurrent and driven consistently by endogenous forces, central banks may be able to design policy frameworks which systematically respond to financial developments in a way that moderate financial cycles and thereby improve economic performance.6
The rest of the paper is structured as follows. The next section introduces our working definition of the financial cycle. Section 3 outlines the modelling approach of time deformation. Section 4 presents the empirical results. Section 5 offers some interpretations of the type of behaviour leading to time deformation. The final section draws conclusions.
We define financial cycles as long swings found in aggregate measures of financial conditions. These swings are plainly visible in Graph 1, where we plot our working definition of the financial cycle for the United States based on Borio et al (2012). Various features of the financial cycle are notable. First of all, the financial cycle does not coincide with NBER-dated business cycles. Second, swings in the financial cycle are generally longer in duration (or, in other words, of a lower frequency) than those in business cycles; and their amplitude tends to display more variability over time.
Graph 2 points to striking visual evidence of time deformation in the financial cycle going back into the 19th century.7 One simple way to illustrate this point is to perform a (nonparametric) Burns and Mitchell-type assessment of time deformation. We first identify the peaks and troughs in calendar time with the Bry-Boschan (1971) turning point dating algorithm. Then, we map these turning points of the cycle into the equally-spaced financial cycle time-scale intervals. This mapping is depicted in the lower panel. Note that the differing lengths of the phases in calendar time represent a differing number of observations being mapped into each financial time interval.
Time dilation and compression are evident in the second half of the sample. The post- WWII period witnessed a very long cycle in calendar time, which appears much shorter in financial cycle time. By way of contrast, recent financial cycles appear relatively short in calendar time relative to financial cycle time. Taken together, these findings suggest that swings in financial conditions, at least when crudely dated using the peaks and troughs of the financial cycle, exhibit some form of time deformation.
The nonparametric Burns and Mitchell approach is subject to the criticism that it is merely highlighting interesting, but possibly spurious, patterns in the data. To put our analysis of time deformation on a more sound statistical footing, we model the financial cycle data using the pioneering approach of Stock (1987, 1988). This allows us to formally estimate and test for time deformation.
We assume financial conditions can be modelled as an n-dimensional data vector generated by a continuous-time A;th-order process, %(s), with time invariant parameters,satisfying the stochastic differential equation:
where At are matrices of AR coefficients,9 D is a differentiation operator, and ((s) is a n- dimensional Gaussian white noise process with mean 0 and covariance matrix D. Note, the time scale s - the so-called operational time (in our application, financial cycle time) - differs from the calendar time scale (which will be denoted by t). As well, the parameters of this process are time invariant with respect to financial cycle time, ie constant with respect to s.
We restrict our attention to a first-order continuous-time auto-regressive (CAR) process that has the following closed-form solution:
Given that ^(s) is not directly observable, we must infer it from financial data, Yt, which is observed in calendar time. In this paper, we assume there is a function g which maps calendar time into financial time, ie s = g(t). This yields the equation linking financial cycle time to calendar time observations:
A few properties of the g function are worth highlighting. First, a reasonable restriction on the g function is the requirement that financial cycle time runs forward and does not run backward. This implies that the g function be monotonically increasing in t. Second, the g function may depend on a set of m observable variables, z, which effectively dilate and compress calendar time relative to financial cycle time.
As in Stock (1988), a convenient form for g is an exponential function:
where c is a m-dimensional vector of parameters. When c = 0, the increments in the g function from t - 1 to t are unitary irrespective of the values of zt-1. In other words, there is no time deformation when the time scales s and t are proportional. The exponential representation also ensures that calendar time and financial cycle time run forward. As for the z variables, we require them to be pre-determined at time t to simplify the estimation.10
Given the g function and the normalisation of the A matrix described in Appendix B, the CAR(1) representation in (discrete) calendar time can be written as
Equation (5) is the state equation governing the evolution of the (unobserved) financial cycle in calendar time. Plugging (5) into (3) and adding a random error term, the measurement equation takes the following form:
where G is a coefficient matrix capturing the financial cycle's influence on Yt. The system defined by the state and measurement equations is a standard linear state space, and its para- meters-as well as the coefficients c in the g function-can be estimated using a Kalman filter.
Table 1 reports the baseline results from the time deformation model. We find evidence of time deformation in the financial cycle, with estimates of the model being intuitively plausible and statistically significant at conventional levels of confidence.
The first two lines in the table report the CAR(1) parameters and values of t-student statistics. The continuous-time autoregressive parameter, A, is fairly stable across models;11 the R2 is roughly 0.85 (not reported in table). The likelihood ratio test indicates that the null hypothesis of no time deformation is rejected at the conventional critical values.12
The positive coefficient estimates on the zt_1 variables are statistically significant, indicating that a decline in the long-term real interest rate, inflation volatility, the corporate spread and NVIX dilates the financial cycle in calendar time. In other words, observed financial cycles tend to get longer.
Testing the statistical significance one variable at a time, the corporate spread appears to have the largest influence on time deformation, with the impact of the interest rate and the NVIX being somewhat less significant and the inflation volatility even less so. When we include all the variables in the g function, the corporate spread still plays a dominant role, with the others smaller but statistically significant.
Graph 3 (left-hand panels) shows the observed financial cycle in calendar time (top) and in financial time (bottom) based on our CAR(1) time deformation model in column 6 of Table 1. A few features stand out. First, the early swings in the data in financial cycle time are compressed in calendar time. Second, the first post-World War II swing appears to be considerably dilated; the 50 years of calendar-based data account for only a few periods in financial cycle time. Third, the last two cycles in the panels are roughly equally spaced in calendar and financial time. Finally, comparing the results in this graph to the B-M approach results in Graph 2 displays some interesting differences. The CAR(1) model finds much more evidence of time deformation than the B-M approach; note the arrows matching the peaks in the series differ. This may reflect the limitation of the B-M nonparametric approach which only uses the peaks and the troughs in the observed series to map calendar time into financial cycle time. In Graph 3, all the data are used in the CAR(1) model.
Also note that the estimates of the time deformation function, g, are plotted in the bottom-right panel. We see a clear association between the value of the g function and the pace of calendar time. The lower (higher) is the value of g, the slower (faster) calendar time proceeds relative to financial cycle time. For example, the first post-World War II cycle is associated with a persistently low value of g. During this period often characterised by financial repression, corporate spreads were below average as were long-term interest rates and inflation volatility (top-right panel).
Recent developments in g may shed light on real-time financial market conditions when trying to distinguish developments associated with boom-bust behaviour and with cyclical fluctuations. For example, the estimates of g at the end of the sample period show an upward trend. One narrow interpretation is that the financial cycle accelerating in calendar time and this information could be useful in forecasting. A somewhat broader, more speculative, interpretation would emphasise that very easy financing conditions may be much less related to broad swings in the financial cycle that lead to costly boom-busts and might be more related to phenomenon associated with the financial accelerator (ie the mechanism resulting in more procyclical business cycles).
Finally, it is important to note the implications for linear model estimation of the financial cycle. Statistically significant time deformation implies a non-stationary time series representation of the financial cycle when observed in calendar time. Technically, a CAR(1) has an approximate time-varying parameter AR(1) representation. If the financial cycle is estimated with a linear model, however, the AR coefficient estimators could be seriously biased. To get a sense of the size of the bias, Graph 4 displays the AR(1) coefficient implied by the CAR(1) baseline estimates. Taken at face value, the variation indicates a risk that linear time series models of the financial cycle are subject to considerable misspecification biases.
In this subsection, we look at alternative measures of the financial cycle and consider additional variables that might improve the fit of the g function.
Alternative measures of the financial cycle Consistent with the approach of Borio et al (2012), we look at a multivariate definition of the financial cycle based on the co-movements in different types of financial cycle indicators, namely credit, house prices and equity prices. To this end, we extract the common component from these indicators via the method of principal components, and treat the first principal component as a measure of the financial cycle. Table 2 reports estimates using this principal component approach.
The estimates in Table 2 are similar to those found in Table 1, with some differences in statistical significance. All variables related to time deformation have a positive sign when evaluated one variable at a time (columns 2-5); this implies that increases in these variables (which might be interpreted as evidence of a more risky macro-financial environment) correspond to a faster pace of calendar time relative to financial cycle time. The joint estimation of all the variables in the g function (column 6) underscores the statistical significance of the corporate spread and the real interest rate at conventional test levels.
Additional candidate zvariables in the p function In Table 3, we extend our baseline model to consider additional variables that may help to capture the nature of time deformation. The first two - slope of the yield curve (ie the long-term nominal yield less the short-term interest rate) and GDP volatility - are alternative risk measures of the macro-financial environment. The coefficient on the slope of the yield curve is negative and statistically significant at the 10% confidence level. The negative sign suggests that as the slope of the yield curve flattens or inverts, calendar time tends to dilate. This evidence is consistent with a reduced risk premia associated with the risk-taking channel during financial booms. Note that inclusion of the yield curve slope reduces the statistical significance of the real interest rate, inflation volatility and the NVIX. These findings support the view that the yield curve is correlated with various developments in the macro-financial environment.
The coefficient on GDP volatility is positive and statistically significant at conventional levels of confidence. The positive sign indicates that great macroeconomic volatility is associated with the compression of calendar time. This is generally consistent with the baseline results that macro-financial risks drive the time deformation.15
We also considered business cycle phase interactions with the financial cycle. We find a strong correlation between business cycle recessions and calendar-time deformation. Recessions tend to speed up financial cycles in calendar time. However, some caution is needed in interpreting this result. The ex-post dating of business cycles implies the use of future information, which could influence inferences. Nonetheless, these correlations do point to possible interactions between business cycles and financial cycles that deserve further exploration.
The coefficient on the financial cycle downturn dummy is positive and statistically significant. Taken at face value, this suggests that time deformation may depend systematically on the phase of the financial cycle. One might conjecture that financial busts naturally lead to an acceleration in calendar time due to the rapid changes in the macrofinancial environment. An alternative implication is purely statistical in the sense that financial downturns may entail nonlinear dynamics associated with doom loops and other amplification mechanisms that are not captured in our g function. Such possibilities may also call for a more elaborate statistical models, such as a CAR(2).
As a final robustness check, we restrict the z variables to be dummy variables which correspond to historical regimes, starting with the Gold Standard period of 1890-1912. The regimes correspond to conventional designations over the past 120 years: Gold Standard, world wars, Roaring Twenties, Great Depression, Financial Repression, Financial Liberalisation and the Great Recession. The coefficient on each dummy variable reflects the average level of time deformation in each era. This test has the benefit of avoiding spurious correlations that may arise when using financial market covariates. Table 4 shows that the average time deformation systematically differs across economic eras. The estimates of Ag during each economic era indicate that calendar time ran slower relative to financial cycle time when below 1 and faster when above 1. Graph 5 displays the implied AR coefficients and variances in calendar time. For example, the Gold Standard and Financial Repression eras exhibit high AR coefficients in periods of low error variances. But the Roaring Twenties, Great Depression and Great Recession eras by contrast were characterised by lower ARs and higher variances.
Our findings raise questions about the possible mechanisms that could be responsible for endogenous, recurrent financial cycles subject to time deformation. Various economic mechanisms could be at work. We group these into two broad categories: behavioural and informational.
Behavioural economics has recently made concrete strides in building models of agents' cognitive processing based on experimental and experiential data (see eg Gabaix (2016, 2017) and the references therein). This interest in cognitive processing, however, has a long history in macroeconomics and monetary policy.18 Allais (1966, 1972) introduced the idea of psychological time into the discussion of money demand. He argued that a psychological time scale, reflecting the need to convert past price experiences into expectations about the future, naturally led one to consider stages of the cycle rather than a year as the appropriate unit of time.19 This strand of the literature had a particular impact on the thinking of Friedman and Schwartz (1982) when reconciling the Gibson paradox with the logic behind the Fisher equation.
Other mechanisms have been emphasised in the behavioural literature. A recent paper by Malmendier and Nagel (2016) argues that agents may overweight personal experiences when forming subjective expectations and making decisions over time.20 In the case of inflation expectations, this learning-from-experience mechanism implies that agents rely more heavily on inflation experienced during their lifetimes than on other historical data when forming expectations. The authors conclude that this perspective helps to explain why persistent disagreements about inflation expectations are associated with age cohorts.21 Looking at labour market dynamics, Storesletten et al (2004) find it useful to use cyclical- phase-dependent cross-sectional moments as far back to 1930 as a means to understand labour income risk perceptions in their panel from 1968-93. Studies of financial markets have also emphasised the role of phase-dependent behaviour. For example, Malmendier and Nagel (2016) look at whether age cohorts assess financial risks differently because of the macroeconomic experiences that they lived through. They find systematic age-cohort effects on portfolio allocations based on having lived through long periods of either low or high returns. They note that this type of return behaviour has a strong influence on the behaviour of the young, which may help to explain why financial boom-bust cycles are associated with periods of long financial upturns in calendar time; in other words, having never experienced a financial downturn young investors may be particularly prone to endogenous risk-taking behaviour. Bordalo et al (2017a, b) develop a theory of diagnostic expectations in macrofinancial data, arguing that there is a 'kernel of truth' property where belief updating results in bigger revisions in expectations than constant-gaining learning would suggest. And, Koudijs and Voth (2014) find corroborating historical evidence looking back over the centuries that personal investment experiences, especially large losses, play a disproportionate role in influencing risk tolerance and hence leverage cycles. Finally, the literature on confirmation bias would naturally be related to the risk-taking channel and stretching of calendar time.22
Informational frictions also provide a mechanism through which financial cycles may be subject to time deformation. Such frictions influence the timing of updates by agents when forming subjective expectations. Stock (1987) mentions a number of earlier papers that justify a distinction between calendar time and business cycle time, such as Barro (1970); Jorda (1999) offers a more recent example of agents' (S, s) decision rules and the nature of the time deformation that this can induce in time series data. In the case of financial booms and busts, Zeira (1999) and Burnside et al (2016) highlight informational setups that lead to nonhomogenous expectation formation. These models focus on informational overshooting in the sense that agents may put too much weight on optimistic outcomes, only to be disappointed. At a more fundamental level of information processing, the sticky information model of Mankiw and Reis (2002) assumes that information disseminates slowly across an economy and hence results in delayed responses of agents and therefore induces persistence in macroeconomic outcomes. Sims (2010) models rational inattention using Shannon entropy to reflect the finite information processing capacity of economic agents. In his model, the size of shocks and their persistence influences the pace at which agents update their expectations and make decisions. This type of behaviour provides a foundation for time deformation that may help to explain important features of financial booms, such as underestimating adverse tail risks during long periods of favourable macro-financial performance.
It should be also noted that the approach taken in this paper is very much subject to the criticism that it is measurement without theory, similar to that which was initially aimed at Burns and Mitchell.23 However, our statistical approach does take us down the road from the ad hoc statistical approach by offering a statistical structure that, in principle, could be embedded in a theory of financial cycles. Recent advances in the field of behavioural economics may also present opportunities to connect our results to underlying endogenous risk-taking economic behaviour of consumers, firms, financial institutions and investors.24 In general, we view the findings in this paper as representing a set of empirical observations-so called, stylised facts about financial cycles-for which a more fulsome model of the financial cycle would need to account.
Finally, our empirical results are consistent with models featuring endogenous risk-taking channels. During risk-on periods when imbalances are growing and the financial cycle is in a boom, real interest rates, corporate spreads and real GDP volatility (along with the VIX) tend to be low, which according to our results correspond to periods when calendar time becomes dilated relative to financial cycle time. Conversely, during bust periods, some or all of these risk measures can sharply rise. Empirically, this type of time deformation would result in long financial booms and shorter financial busts of the type that we have seen in the historical data across many countries.
Our paper demonstrates that the statistical methods initially designed for measuring business cycle time are useful for measuring financial cycle time. We showed that it is possible to jointly estimate the time series behaviour of financial conditions and the nature of the time deformation. As a by-product of this analysis, we produced a new set of stylised facts about financial cycles, which may help motivate theoretical modelling. However, to understand the wider implications of these findings, we need new theoretical models that can capture the mechanisms that generate this type of behaviour. In particular, a model that purports to capture financial cycle dynamics should be able to match the stylised facts about time deformation. Our identification of the variables associated with the extent of the time deformation provides useful clues about what such a model might look like.
With our approach to measure financial cycle time, we have found statistical evidence of significant time deformation in the US financial cycle. The extent of the time deformation appears to be associated with a set of variables measuring subjective risk perceptions. These variables include the level of the long-term interest rate, which has been identified as an important indicator of the strength of risk-taking channel. The other key variables are the volatility of inflation, corporate spreads and the NVIX. These are more direct measures of the riskiness of the macro-financial environment.
The recurrent nature of the financial cycle in financial cycle time also highlights the strength of the underlying forces influencing swings in financial conditions. Over the past 120 years covered by our dataset, there has been a wide range of economic and policy environments that influenced the financial cycle. The United States went from being an emerging market to an advanced economy with a highly open financial market. It has gone through the Gold Standard period when inflation was low as well as the 1970s when inflation was high and volatile. Over this long historical period, the price stability credentials of central banks have waxed and waned. As well, fiscal and regulatory policies have varied considerably. Through all this, the financial cycle dynamics remain a constant feature of the economy.
This leads us to speculate that financial cycles are intrinsic to all financially liberalised, market-oriented economies. In this respect, we find attractive those theories that point to the endogenous risk-taking behaviour of agents in the way they process information when forming expectations about the future and especially about risk assessments. This is not a new observation. For example, Minsky (1982) and Aliber and Kindleberger (2015) put forth similar perspectives using historical narratives. The main difference with our study is that we are led to this conclusion by looking at historical data via a formal statistical model emphasising the lens of financial cycle time.
Our findings also shed light on the type of economic mechanisms that might be relevant. Namely, the estimated time deformation is consistent with the information processing and behavioural mechanisms highlighted in the introduction. One facet of these mechanisms that stands out is that they are associated with basic human behaviour and are not regime dependent. In other words, these mechanisms could be seen as reflecting the fundamentals of human behaviour when agents are free to choose from the opportunities that confront them. As such, it may be reasonable to look over long histories to better understand current- day challenges, even though the 19th and early 20th centuries, for example, would appear to be so fundamentally different, say, economically and culturally than today.
With respect to policy implications, our paper provides a new way to analyse the state of financial conditions. For example, easy financial conditions during a period of strong calendar time dilation (relative to financial cycle time) appear to be consistent with rising concerns about the behaviour associated with costly booms and busts. In contrast, easy financial conditions during a period of calendar time compression is more closely associated with models of the financial accelerator, which have less to do with booms and busts and more to do with the amplification role financial conditions can play in the business cycle. Therefore, the information about the financial cycle viewed from the time deformation perspective may help to clarify when policy should 'lean against the wind' and when it simply needs to be more countercyclical.
While our study focuses on past financial cycles, it would be useful in the future to explore the nature of time deformation spillovers from the financial cycle to the business cycle - especially during balance sheet recessions - and to financial cycles in other countries. From a modelling perspective, more attention deserves to be spent on models capturing endogenous financial cycles that are subject to time deformation and the implications for the design of policy frameworks.
Measure of financial conditions: US credit to the private non-financial sector (scaled by potential GDP). Annual nominal credit data for 1952-2017 (year-end values) from the BIS; for 1880-1952, we proxy credit data using growth rates of total bank loans growth from Jorda et al (2017). Nominal GDP data for 1880-2013 from Jorda et al. (2017); for 2014-2017, from the US Bureau of Economic Analysis. Potential GDP estimated as the trend from a Hodrick- Prescott filter (smoothing parameter of 100). Finally, we use a band-pass filter for frequencies 8-32 years for smoothing the credit ratio.
Real house price index data for 1950-2017 is taken from Schiller (2015); for the 1880-1945 period, due to poor quality of house price measures, we proxy the real house price index with real equity index (also based on Schiller (2015)); for 1945-1950 we proxy house price index by a weighted average of real house price index and real equity price index. In order to smooth the credit ratio, we use a band-pass filter for frequencies of 8 to 32 years.
Long-term real interest rate: nominal long-term rate less the 10-year centred average rate of annual inflation. Annual long-term interest rate for 1880-2013, from Jorda et al. (2017). For 2014-2017, annual average yield on US Treasury securities (10-year constant maturity) from the Federal Reserve. Annual inflation data for 1880-1913 calculated as year-on-year changes in the General Price Index from the NBER macro-history data. For 1914-2017, year-on-year change in annual CPI from FRED.
Inflation volatility: Centred 5-year moving standard deviation of the annual inflation data.
Corporate spread: Long-term corporate yield less the long-term US treasury yield. For 18801918, annual corporate yield from Gordon (1986); from 1919-2017, Moody's annual (seasoned) Baa corporate bond yield from FRED.
Volatility index: stock market implied volatility. For 1986-2017, annual average of the CBOE VXO from FRED; for 1890-1986, the NVIX volatility index from Manela and Moreira (2017).
These candidate variables for the g function: Data lagged one year and smoothed using a onesided Hodrick-Prescott filter.
A continuous-time AR processes satisfying equation (1) can be estimated using the Kalman filter (Harvey and Stock, 1985) with the modification proposed by Stock (1988) to allow for time deformation. This section gives further details on the methodology.
We start by re-writing (1) in stacked form, so that the state equation for a general CAR(^) process can be expressed as:
Note that the integral in equation (B.6) depends on the shape of the g function and the assumption that z is pre-determined at time t. Furthermore, under the simplifying assumption that g can be well approximated by a piecewise linear function, that is
and the covariance matrix of the prediction error updates according to the following equation:
which can be maximised using standard algorithms to obtain parameter estimates and (filtered and smoothed) estimates of the latent process, f.
In Appendix B, we justify the use of the Kalman filter methodology assuming that zt is predetermined with respect to Yt. In this appendix, we discuss the exogeneity conditions that expand the set of admissible zt variables for the g function. Consider Yt and zt as being represented by a general linear model of the economy, which can be express as
where ut is a white noise error term uncorrelated with the past values of Yt and has a generic covariance matrix S.
Based on Engle et al (1983), zt is pre-determined in equation (C.1) if
in this case, equation (C.1) has a strong structural form.
In general, such conditions cannot be guaranteed. However, economic theory may impose a set of cross-equation restrictions with implications for the exogeneity conditions. The constraints on A0 and B0 are particularly relevant as they relate to the contemporaneous correlation structure of Yt and zt; one such useful constraint can be expressed as:
where O(L) are polynomials in the lag operator and the covariance matrix of the errors is block-diagonal. Under the contemporaneous correlation structure in equation (C.2), zt is weakly exogenous if and only if
Condition (C.6) implies a contemporaneous correlation structure that is consistent with the cross-equation restrictions with shocks wY t and wzt being orthogonal. Condition (C.7) implies an additional requirement that z should not Granger-cause Y.
Under these assumptions, the Kalman filter recursion can be rewritten with contemporaneous values of zt appearing in the g function. Equation (B.4) can be written as
This derivation highlights the role of the exogeneity assumption; namely, innovations to the Qt matrix (and hence to the state vector) are orthogonal to those of the observed equation. Furthermore, under the strict exogeneity assumption, the shocks to zt would influence Yt only through the changes in the integration limits of equation (C.10).
According to equation (C.13), the prediction error variance of the latent variable is the same as in the case with pre-determined zt (B.13). As soon as zt becomes available, however, one would need to update equation (C.13) with the new estimate Qt. This would then be used to update the estimate of the state variable as in (B.16).
This appendix has outlined the conditions under which we can extend the Kalman filter approach in the paper to allow for contemporaneous values of zt in the g function. Weak exogeneity can be secured by imposing theory-based constraints on the contemporaneous correlation structure. However, the additional assumption ensuring strong exogeneity is that of Granger non-causality from Yt to zt, which is necessary to guarantee that zt does not directly influence Yt, but only does so through the time deformation function, g. In practical terms, this assumption can be quite restrictive as it implies that all the potential drivers of the financial cycle only affect it through time deformation.
Adrian, T, and F Duarte (2018): "Financial vulnerability and monetary policy", CEPR Discussion Papers, no DP12680, February.
Adrian, T, N Boyarchenko and D Giannone (2017): "Vulnerable growth", Federal Reserve Bank of New York Staff Reports, no 794.
Adrian, T, F Grinberg, N Liang and S Malik (2018): "The term structure of growth-at-risk", IMF Working Papers, no 18/180, August.
Albuquerque, R, M Eichenbaum, D Papanikolaou and S Rebelo (2015): "Long-run bulls and bears", Journal of Monetary Economics, 76(S), pp S21-S36.
Aliber, R, and C Kindleberger (2015): Manias, Panics, and Crashes: A History of Financial Crises, seventh edition, Palgrave.
Allais, M (1966): "A restatement of the quantity theory of money", American Economic Review, vol 56(5), pp 1123-57.
Allais, M (1972): "Forgetfulness and interest", Journal of Money, Credit and Banking, vol 4(1), part 1, pp 40-73.
Bank for International Settlements (BIS) (2016): 86th Annual Report, chapter IV.
Barro, R (1970): "Inflation, the payments period, and the demand for money", Journal of Political Economy, vol 78(6), pp 1228-63.
Barthalon, E (2014): Uncertainty, Expectations and Financial Instability: Reviving Allais's Lost Theory of Psychological Time, Columbia University Press.
Bernanke, B, M Gertler and S Gilchrist (1999): "The financial accelerator in a quantitative business cycle framework", in (J Taylor and M Woodford, eds) Handbook of Macroeconomics, Elsevier, vol 1, chapter 21, pp 1341-93.
Bordalo, P, N Gennaioli, Y Ma and A Shleifer (2017a): "Overreaction in macroeconomic expectations", unpublished Harvard University manuscript, December.
Bordalo, P, A Gennaioli and A Shliefer (2017b): "Diagnostic expectations and credit cycles", The Journal of Finance, 73(1), pp 199-227.
Borio, C (2013): "On time, stocks and flows: understanding the global macroeconomic challenges", National Institute Economic Review, vol 225(1), pp R3-R13.
Borio, C, P Disyatat, M Drehmann and M Juselius (2017): "Monetary policy, the financial cycle, and ultra-low interest rates," International Journal of Central Banking, vol 13(3), pp 55-89.
Borio, C, M Drehmann and K Tsatsaronis (2012): "Characterising the financial cycle: don't lose sight of the medium term!", BIS Working Papers, no 380, June.
Brunnermeier, M, and Y Sannikov (2014): "A macroeconomic model with a financial sector", American Economic Review, 104(2), pp 379-421.
Bruno, V, and H-S Shin (2015): "Cross-border banking and global liquidity", Review of Economic Studies, vol 82(2), pp 1-30.
Bry, G, and C Boschan (1971): Cyclical Analysis of Time Series: Procedures and Computer Programs, National Bureau of Economic Research.
Burns, A, and W Mitchell (1946): Measuring Business Cycles, National Bureau of Economic Research.
Burnside, C, M Eichenbaum and S Rebelo (2016): "Understanding booms and busts in housing markets", Journal of Political Economy, vol 124(4), pp 1088-147.
Cagan, P (1956): "The monetary dynamics of hyperinflation", in (M Friedman, ed) Studies in the Quantity Theory of Money, University of Chicago Press.
Chevillon, G, and S Mavroeidis (2017): "Learning can generate long memory", Journal of Econometrics, vol 198(1), pp 1-9.
Claessens, S, M Ayhan Kose and M Terrones (2011): "Financial cycles: what? how? when?", NBER International Seminar on Macroeconomics, vol 7(1), pp 303-34.
De Grauwe, P, and M Grimaldi (2006): The Exchange Rate in a Behavioral Finance Framework, Princeton University Press.
DeLong, B, and L Summers (1986): "Cyclical variability of economic activity in the United States", in (R Gordon, ed) The American Business Cycle: Continuity and Change, University of Chicago Press.
Duval, R, GH Hong and Y Timmer (2017): "Financial frictions and the great productivity slowdown", IMF Working Papers, no 17/129, May.
Eckstein, O, and A Sinai (1986): "The mechanisms of the business cycle in the postwar era", in (R Gordon, ed) The American Business Cycle: Continuity and Change, University of Chicago Press.
Engle, R, D Hendry, and J Richard (1983): "Exogeneity", Econometrica, vol 51(2), pp 277-304.
Filardo, A, and D Guinigundo (2008): "Transparency and communication in monetary policy: a survey of Asian central banks", paper presented at the BIS-BSP Research Conference, April.
Filardo, A, and P Rungcharoenkitkul (2016): "A quantitative case for leaning against the wind", BIS Working Papers, no 594, December.
Friedman, M, and A Schwartz (1982): Monetary Trends in the United States and United Kingdom: Their Relation to Income, Prices, and Interest Rates, 1867-1975, National Bureau of Economic Research.
Gabaix, X (2016): "A behavioral New Keynesian model", NBER Working Papers, no 22954, December.
Gabaix, X (2017): "Behavioral inattention", NBER Working Paper, no 24096, December.
Gilchrist, S, and E Zakrajsek (2012): "Credit spreads and business cycle fluctuations", American Economic Review, vol 102(4), pp 1692-720.
Gordon, R (1986): The American Business Cycle: Continuity and Change, University of Chicago Press.
Gourio, F, A Kashyap and J Sim (2017): "The tradeoffs in leaning against the wind", NBER Working Papers, no 23658, August.
Hall, R (2013): "Financial frictions", International Journal of Central Banking, vol 9(2), pp 15563.
Hall, R (2017): "High discounts and high unemployment", American Economic Review, vol 107(2), pp 305-30.
Harvey, A (1989): Forecasting, Structural Time Series Models and the Kalman Filter, Cambridge University Press.
Harvey, A, and J Stock (1985): "The estimation of higher order continuous time autoregressive models", Econometric Theory, vol 1, pp 97-112.
Jorda, O (1999): "Random-time aggregation in partial adjustment models", Journal of Business and Economic Statistics, vol 17(3), pp 382-95.
Jorda, O, M Schularick and A Taylor (2017): "Macrofinancial history and the new business cycle facts", in (M Eichenbaum and J Parker, eds) NBER Macroeconomics Annual2016, vol 31, University of Chicago Press.
Koopmans, T (1947): "Measurement without theory", The Review of Economics and Statistics, vol 29(3), pp 161-72.
Koudijs, P, and H-J Voth (2014): "Leverage and beliefs: personal experience and risk taking in margin lending", NBER Working Papers, no 19957, March.
Linde, J (2018): "DSGE models: still useful in policy analysis?", Oxford Review of Economic Policy, vol 34(1-2), pp 269-86.
Lucas, R (1981): Studies in Business-Cycle Theory, MIT Press.
Malmendier, U, and S Nagel (2011): "Depression babies: do macroeconomic experiences affect risk taking?", The Quarterly Journal of Economics, vol 126(1), pp 373-416.
Malmendier, U, and S Nagel (2016): "Learning from inflation experiences", The Quarterly Journal of Economics, vol 131(1), pp 53-87.
Manela, A, and A Moreira (2017): "News implied volatility and disaster concerns", Journal of Financial Economics, vol 123(1), pp 137-62.
Mankiw, G, and R Reis (2002): "Sticky information versus sticky prices: a proposal to replace the New Keynesian Phillips curve", The Quarterly Journal of Economics, vol 117(4), pp 12951328.
Melino, A (1994): "Estimation of continuous-time models in finance", in (C Sims, ed) Advances in Econometrics: Sixth World Congress, vol II, Cambridge University Press.
Meltzer, A (2003): "Rational and nonrational bubbles", in (W Hunter, G Kaufman and M Pomerleano, eds) Asset Price Bubbles: The Implications for Monetary, Regulatory, and International Policies, MIT Press.
Mian, A, and A Sufi (2018): "Finance and business cycles: the credit-driven household demand channel", NBER Working Papers, no 24322, February.
Minsky, H (1982): Can "It" Happen Again?, ME Sharpe.
Moore, G (1961): Business Cycle Indicators, Princeton University Press.
Reinhart, C, and K Rogoff (2011): This Time Is Different: Eight Centuries of Financial Folly, Princeton University Press.
Runstler, G, and M Vlekke (2016): "Business, housing and credit cycles", ECB Working Papers, no 1915, June.
Schularick, M, and A Taylor (2012): "Credit booms gone bust: monetary policy, leverage cycles, and financial crises, 1870-2008", American Economic Review, vol 102(2), pp 1029-61.
Shiller, R (2015): Irrational Exuberance, third edition, Princeton University Press.
Sims, C (2010): "Rational inattention and monetary economics", in (B Friedman and M Woodford, eds) Handbook of Monetary Economics, vol 3, Elsevier, pp 155-81.
Stock, J (1987): "Measuring business cycle time", Journal of Political Economy, vol 95(6), pp 1240-61.
Stock, J (1988): "Estimating continuous-time processes subject to time deformation: an application to postwar U.S. GNP", Journal of the American Statistical Association, vol 83(401), pp 77-85.
Stock, J, and M Watson (1998): "Median unbiased estimation of coefficient variance in a time- varying parameter model", Journal of the American Statistical Association, vol 93(441), pp 349-58.
Storesletten, K, C Telmer and A Yaron (2004): "Cyclical dynamics in idiosyncratic labor market risk", Journal of Political Economy, vol 112(3), pp 695-717.
Tversky, A, and D Kahneman (1973): "Availability: a heuristic for judging frequency and probability", Cognitive Psychology, vol 5(2), pp 207-32.
Zarnowitz, V (1992): Business Cycles: Theory, History, Indicators, And Forecasting, University of Chicago Press.
Zeira, J (1999): "Informational overshooting, booms, and crashes", Journal of Monetary Economics, vol 43(1), pp 237-57.