چگونه بحران مالی همبستگی میزان بازده آمریکا را تغییر می دهد؟

We investigate the pairwise correlations of eleven U.S. fixed income yield spreads over a sample that includes the Great Financial Crisis of 2007–09. Using cross-sectional methods and nonparametric bootstrap breakpoint tests, we characterize the crisis as a period in which pairwise correlations between yield spreads were systematically and significantly altered in the sense that spreads comoved with one another much more than in normal times. We find evidence that, for almost half of the fifty-five pairs under investigation, the crisis has left spreads much more correlated than they were previously. This evidence is particularly strong for liquidity- and default-risk-related spreads, long-term spreads, and the spreads that were most likely directly affected by policy interventions.

A persistent state of turmoil engulfed the international financial markets – particularly U.S. equity, debt, credit, and derivatives markets – between the summer of 2007 and the late spring of 2009. A number of papers (e.g., Caprio et al., 2010 and Galí, 2010) have labeled such a state, characterized by unsettled and dysfunctional markets, as the “Great Financial Crisis.” There is now little doubt that the Great Financial Crisis ravaged U.S. fixed income (debt and credit) markets in unprecedented ways (see Dwyer and Tkac, 2009). Using data from the epicenter of the crisis, a range of U.S. fixed income markets, we pose two questions. First, can the Great Financial Crisis be truly seen as an approximately 2-year crisis episode that progressively abated to leave markets in a “normal” state similar to the one that had prevailed before 2007? Or second, to the contrary, was the Great Financial Crisis so pervasive that it left the relationships among different fixed income segments permanently altered? More generally, some economic literature has investigated the fabric that turns a state of turmoil in the fixed income market into a persistent regime so severe to merit being dubbed a crisis, or even a “great” one nonetheless. A number of papers have focused on anomalies in the univariate dynamics of the first (the level) and second (the volatility) moments of yield spreads, constructed as the difference between the yield to maturity of a risky asset and the yield of a comparatively less risky (or riskless) asset (see, e.g., Guidolin and Tam, 2010 and Muir, 2013), sometimes also using event studies (see, e.g., Nippani and Smith, 2010). However, a multivariate approach focused on the comovement patterns across fixed income markets should also contribute to a useful economic characterization of the Great Financial Crisis. For instance, in a rare example of multivariate analysis, Dungey et al. (2010) develop a method based on structural identification via heteroskedasticity to separate market contagion from hypersensitivity during crises. They exploit the ability of multivariate generalized autoregressive conditional heteroskedasticity (GARCH) models to forecast the dynamics of correlations. However, their application concerns only the 1997–98 Asian crisis. Another strand of the literature has intensely debated the exact dating of the Great Financial Crisis. The differing conclusions have often reflected the priors of the different researchers as well as their specific methodological approaches. With a few exceptions (e.g., Campbell et al., 2011 and Frank and Hesse, 2009), most papers have agreed on early August 2007 as a potential starting date of the Great Financial Crisis, even though only a few have traced this claim back to the actual behavior of financial data. Moreover, only a handful of papers have ventured into establishing an end date for the Great Financial Crisis (see, e.g., Aït-Sahalia et al., 2009, Campbell et al., 2011 and Guidolin and Tam, 2010). Usually such attempts have consisted of generic and informal claims about the possibility that the effects of the crisis were reabsorbed around mid-2009. In this paper, we develop a characterization of the Great Financial Crisis based on the multivariate behavior of a large set of fixed income yield spreads that offers a novel perspective on the end of the crisis. Yield spreads measure various dimensions of risk. While studying the level of yields per se has value in certain contexts, many papers and articles in the literature favor the analysis of yield spreads because they offer a clearer picture of the underlying trade-offs for investors. We focus on fixed income yield spreads for several reasons. First, filtering a financial crisis through the lenses of spread data is an implicit way to relate financial events to the business cycle (see, e.g., Gilchrist et al., 2009). In general, yield spreads are likely to be informative of the channels through which financial prices affect the real side of the economy. In particular, fixed income spreads tend to widen shortly before the onset of recessions and to narrow again before recoveries. Analyzing the behavior and the common dynamics of yield spreads based on interest rates derived from the core of the financial crisis (i.e., fixed income markets) sheds light on some important aspects of the turmoil and on the relationship between economic activity and the evolution of fixed income markets. Second, economists are generally interested in understanding the mechanisms that link variables in a given market. The extent of comovement between fixed income yield spreads may have consequences for the cost of borrowing, the portfolio diversification strategy of investors of various types, and the modeling and forecasting of interest rates in the market under investigation. Moreover, a good understanding of the dynamics of credit and liquidity risk premia incorporated in the prices of fixed income securities potentially has a number of practical implications for portfolio managers and policymakers. On the one hand, through such understanding, financial market regulators may be able to improve capital allocation functions and the information aggregation process in fixed income markets. Furthermore, they may be able to evaluate the robustness of such markets to shocks to the financial system. On the other hand, investors may want to look at the dynamics of credit and liquidity risk premia to derive better information about the return and diversification properties of major asset classes. Overall, a careful assessment of the main characteristics of such premia may be associated with better investment and policy decisions over the business cycle. Given these premises and objectives, we systematically investigate the empirical behavior of pairwise correlations between eleven U.S. fixed income spreads over a sample of weekly data between 2002 and 2011. These spreads are distinct in terms of the securities or markets to which they refer, the maturity of the underlying securities, and whether they were affected by specific policy measures by the Federal Reserve and policymakers more generally (e.g., the Treasury and the Federal Deposit Insurance Corporation) in reaction to the Great Financial Crisis. Our series measure yield spreads for a variety of instruments and markets, namely, 3-month London interbank offered rate (LIBOR) unsecured deposits, 3-month unsecured financial and asset-backed commercial paper (ABCP), 5-year swaps, 5-year Resolution Funding Corporation (REFCorp) strips, 5-year Aaa private-label commercial mortgage-backed securities (CMBS), 10-year off-the-run Treasury securities, 20-year Moody's Baa-rated and Aaa-rated corporate bonds, 20-year Moody's Bbb-rated and Aa corporate bonds, and 30-year conventional fixed-rate mortgage-backed securities (MBS). This list also includes two typical mortgage-related risk premia because the U.S. mortgage market is identified as the catalyst of the financial crisis (see Frank and Hesse, 2009).3 We use a mixture of cross-section econometric methods to test the existence of nonzero correlations for groups of spreads and test for breaks in the correlations between spreads. We generally reject both the null hypothesis of no cross-section correlation between spreads in all subperiods we consider and the null hypothesis of constant pairwise correlations over time. We provide a characterization of the Great Financial Crisis as a period during which pairwise correlations between yield spreads were systematically and significantly altered, with spreads comoving with one another much more strongly than in “normal” times. Our work is consistent with the intuition that the Great Financial Crisis was a period of structural and systematic alteration of correlations between spreads, possibly (but not exclusively) induced by a soaring exposure of the underlying securities to common crisis factors (such as declining risk appetite, liquidity shortages, and funding problems for intermediaries; see Brunnermeier and Pedersen, 2009). Implicitly, we consider the possibility that unconditional average correlations may be unstable over time. Although rich but stationary models may be applied to describe the time variation of conditional correlations, we identify and interpret medium-term movements in average correlations across pre- and post-crises regimes.4 Our results are considerably more intriguing than just a novel characterization of the Great Financial Crisis as a shock wave that has affected spread correlations in addition to their means and volatilities. Using nonparametric bootstrap methods, we find evidence that the Great Financial Crisis has left the spreads much more correlated than before the crisis. This evidence appears to be particularly strong for three (occasionally overlapping, but clearly defined) subsets of spread pairs, defined according to spread features — that is, characterized by liquidity problems, measuring default risk, or directly influenced by policy interventions. We also discuss which factors might have driven the correlations during and after the Great Financial Crisis. First, we find that the correlations between a majority of liquidity-related spreads increased so substantially during the Great Financial Crisis that they have not reverted to normal levels in the aftermath of the crisis. This finding suggests that, for most spreads, their exposure to a liquidity factor has been substantially altered by the Great Financial Crisis. Second, almost two-thirds of the default risk spread correlations have remained altered even after the Great Financial Crisis, consistent with permanently altered exposures of default-risk-related spreads to a common default risk factor. Third, about half of the correlations between spreads affected by policy interventions have returned to levels that exceed the pre-crisis norm. Pairs of spreads that typically capture default risk (e.g., the Baa–Aaa corporate or the corporate junk spreads) were all simultaneously affected by policy interventions. These spreads may have inherited patterns of behavior in the post-crisis period that reflect the possibility of future interventions. However, the higher correlation levels in the aftermath of the Great Financial Crisis can also be considered as indication that the vast array of policy measures deployed to counter the effects of the crisis may have affected the set of investment opportunities in structural and possibly undesirable ways.5 Our findings concerning the failure of many pairwise spread correlations to revert to their normal pre-Great Financial Crisis levels cast doubts on some of the recent literature that has concluded – perhaps too quickly and dismissively – that the crisis was over by mid-2009. Even though the means and volatilities of many spreads have indeed returned to their pre-crisis norm, permanently altered (higher) correlations between spreads may also have produced severe long-run effects. In this sense, and possibly in the light of the sovereign debt crisis that has affected the international fixed income markets since 2010, the Great Financial Crisis may have been over much later than commonly believed and possibly beyond the end of our sample. Using a different methodology and a smaller but more diverse set of underlying assets, a related paper by Dungey et al. (2012) also highlights the possible existence of the structural shifts we discuss in our analysis. Their paper uses a parametric smooth transition structural GARCH model to endogenously detect simultaneous structural shifts in the relationships (dynamic correlations) among U.S. stocks, real estate, and Treasury securities during different stages of the Great Financial Crisis that are consistent with ours (see Section 4.1). A stark outcome of their empirical efforts is that financial conditions in 2009–10 were not back to where they were earlier in the decade and, in particular, that the contemporaneous linkages between bond and stock markets did not return to pre-crisis conditions. Similar to our conclusions, their findings suggest that the Great Financial Crisis left persistent effects at least through 2010. As our main focus is on a larger but more homogeneous set of yields spreads at the epicenter of the Great Financial Crisis, we see our results and theirs as complements. Both Dungey et al. (2012) and we address the problem raised by Dungey and Zhumabekova (2001) that testing for changes in correlations in small samples may seriously affect the power of the test. However, while Dungey et al. (2012) work within a tight parametric framework that allows them to isolate the effects of the correlation variations and rule out potential biases derived from contemporaneous instability in the volatilities, we adopt a nonparametric framework. In particular, we implement a nonparametric bootstrap methodology to test for breaks in signed pairwise correlations with the objective of characterizing the evolution of correlations without imposing any specific – and potentially misspecified – parametric structure. We also correct for the small-sample biases discussed in Dungey and Zhumabekova (2001). Our paper is related to a strand of the literature that proposes increasingly sophisticated models for the dynamics in yield spreads, such as Davies (2008), who analyzes the determinants of U.S. credit spreads over an 85-year period that covers several business cycles. His analysis demonstrates that econometric models are capable of explaining up to one-fifth of the movement in the spreads. Davies also reports that maximum explanatory power is achieved using nonlinear econometric frameworks of the regime-switching type. One can interpret our modeling, in which structural change in second moments is allowed, as an additional case for such a regime-switching behavior. The rest of the paper is organized as follows. In Section 2 we sketch the methodological aspects of our empirical investigation. Section 3 is devoted to the description of the dataset and the summary statistics of the eleven spreads over three subsequent subsamples of the 2002–11 period. We present our empirical results in Section 4, including details on the dating of the Great Financial Crisis. We conclude in Section 5.

In this paper, we have systematically investigated the empirical behavior of the correlations of eleven U.S. fixed income yield spreads over a 2002–11 period surrounding the so-called Great Financial Crisis of 2007–09. We have used a combination of cross-section methods to determine the existence of nonzero correlations in groups of spreads and break tests concerning pairwise correlations to provide a novel characterization of the Great Financial Crisis. Our results indicate that during the crisis most pairwise correlations between yield spreads were systematically and significantly altered in the sense of spreads comoving with one another much more strongly than in normal times. These results should be contrasted with the more traditional studies that have examined either the effects of the crisis on spreads only at a univariate level or those focused only on its effects on the (conditional) mean and variance of the spreads. Our work confirms the heuristic idea that the Great Financial Crisis was also a period of structural and systematic alteration of correlations possibly (but not necessarily) induced by the common and soaring exposures of the securities underlying the spreads to common factors (such as disappearing risk appetites, liquidity shortages, and funding problems for intermediaries). The adoption of a nonparametric bootstrap approach provides evidence that for almost half of the fifty-five pairs of spreads investigated, the Great Financial Crisis has left fixed income spreads more highly correlated than before the crisis. This evidence appears particularly strong for three (occasionally overlapping, but clearly defined) subsets of spread pairs. From a financial point of view, significantly altered correlations might affect investment decisions and the composition of portfolios and their characteristics in terms of diversification. We have also discussed which factors might have driven the correlations during and after the Great Financial Crisis. First, we found evidence that the increase of the correlations of the majority of the liquidity-related spreads during the Great Financial Crisis was so substantial that such correlations have failed to revert to normal levels after the crisis. The exposure of most spreads to a liquidity factor appears to have been substantially increased by the Great Financial Crisis. Second, almost two-thirds of the default risk spread correlations remained altered even after the Great Financial Crisis. This finding is consistent with permanently altered exposures of default-risk-related spreads to a common default risk factor. Third, about half of the correlations affected by policy interventions eventually reached levels exceeding the pre-crisis standards. This result may be deemed an indication of the possibility that the broad array of policy measures deployed to counter the effects of the crisis in fixed income markets structurally affected the set of investment opportunities. While it would be interesting to develop and estimate a factor model to theoretically frame our results, we leave this extension to future work. Finally, our findings regarding the failure of many pairwise spread correlations to revert to their pre-Great Financial Crisis levels cast doubts on the results of some of the recent literature, which may have too quickly and dismissively concluded that the crisis was over as early as mid-2009. Even though means and variances of many spreads have returned to their pre-crisis norms, permanently altered (higher) correlations between spreads may have produced undesirable long-run effects. An important implication of our results is that the Great Financial Crisis may have come to an end much later than commonly believed and possibly beyond the last observation in our sample.