BIS Quarterly Review, June 2000

Mico Loretan William B English

(+1 202) 452 2219 (+1 202) 736 5645
mico.loretan@frb.gov william.b.english@frb.gov

III. Special feature: Evaluating changes in correlations

during periods of high market volatility*

In computing measures of the market risk of a portfolio, such as Value at Risk, portfolio managers
typically rely on estimates of correlations between returns on the financial instruments in the portfolio
and on the volatility of those returns. This task is relatively simple if the correlations and volatilities do
not change over time, and if there are sufficient data to allow them to be estimated fairly precisely.
The task is vastly more difficult if the correlations change abruptly as a result of structural breaks in
the mechanisms that determine asset returns – perhaps owing to the impact of contagion on the links
between markets, changes in the sources of shocks, or new market structures or practices.12 However,
changes in correlation patterns may be no more than the natural and predictable effects of fluctuations
in asset return volatility. In such cases, the problem facing risk managers should be less difficult, as
the empirical challenge then consists of modelling the time-varying nature of asset return volatilities.
In periods of heightened market volatility, correlations between returns on financial assets tend to
increase relative to correlations estimated during periods of normal volatility. For example, the
average correlation between yield spreads for selected fixed income securities rose to 0.37 following
the Russian crisis in August 1998 from 0.11 in the first half of 1998 (Committee on the Global
Financial System (1999), Table A18). The increased correlation of returns during periods of high
volatility is often explained as resulting from changes in the underlying relationships that determine
returns.13 Yet, probability theory shows that correlations between asset returns depend on market
volatility even if the underlying relationships between returns have not changed; variations in
correlations measured over different periods of time may merely be the consequence of variations in
realised volatility.
This article explores the link between volatility and correlation, which has until recently largely been
overlooked in the economics and finance literature.14 The next subsection provides two numerical
examples that demonstrate the dependence of correlations on volatilities, and also states a theorem that
links variances and correlations. An empirical application is presented next, focusing on the behaviour

*
This article is based on a longer BIS conference research paper (Loretan and English (2000)). The authors are members
of the research staff of the Board of Governors of the Federal Reserve System. The analysis and conclusions in this
article are those of the authors and do not indicate concurrence by the Board of Governors, by the Federal Reserve
System or by the BIS. We thank Jim Clouse, Mike Gibson, Michael Gordy, Brian Madigan, Henri Pagès, Matt Pritsker,
Vince Reinhart and participants at the 1999 Central Bank Economists’ Autumn Meeting at the Bank for International
Settlements for helpful comments and discussions.
12
Recent discussions of possible routes for contagion include Drazen (1998), Eichengreen et al (1996) and Gerlach and
Smets (1995). Kodres and Pritsker (1999) present a structural model of the contagion-like transmission of shocks.
13
For example, the increase in market volatility of US and other government securities in 1994 was accompanied by an
increase in sampling correlations. In its 1995 annual report, Bankers Trust (1995) stated that movements in interest rates
in 1994 were “unusual in the degree to which interest rates across international markets moved together” (p 23). The bank
went on to note that “this phenomenon of increased correlation among interest rates reduced the risk management
benefits derived from diversification across interest-sensitive instruments” (p 23). The bank responded to this situation by
withdrawing from substantial market positions (p 24).
14
For previous economic and finance studies of the link between volatility and correlations, see Ronn (1995), Boyer et al
(1999) and Forbes and Rigobon (1999).

29
BIS Quarterly Review, June 2000

of equity returns in the United Kingdom and Germany during the past decade. We find that quarters in
which the volatility of equity returns was high also tended to be quarters with above average
correlations, in a manner that is consistent with a constant unconditional data generating process for
equity returns. The final subsection discusses the implications of the link between volatility and
correlation for risk management and for financial supervision.

The link between volatility and correlation

According to probability theory, when the movements of random variables are more volatile, sampling
correlations between those variables should be elevated even if the underlying process generating the
variables remains unchanged. Boyer et al (1999) provide a formal proof of this link (see the box on the
next page).
To demonstrate the intuition that underlies this theoretical result, consider a pair of random variables,
x and y, and suppose that the possible outcomes for these variables are distributed jointly normally
with means equal to zero, variances equal to one and contemporaneous correlation equal to 0.5. A
sample of 1,000 independent draws of such pairs is shown in the graph below. The thick and thin
ellipses denote the areas that contain 50% and 95% of the total mass of the distribution respectively.
Now suppose we split this sample into two subsamples based on the outcome of the variable x. One
subsample would be “low volatility” and would include all x and y pairs for which the absolute value
of x is less than 1.96.15 The other subsample would be “high volatility” and would include all pairs for
which the absolute value of x is greater than or equal to 1.96. Intuitively, the effect of excluding
observations with large values of x should be to reduce the sample correlation between x and y. By
contrast, the correlation for the high volatility subsample should be enhanced because one portion of
that subsample picks up the large positive values of both variables while the other portion picks up the
large negative values. As is noted in the graph below, the difference between the correlations in the
two subsamples is large: the correlation for the high volatility sample is 0.81, while that for the low
volatility sample is 0.45. Note that the correlation in the latter subsample is close to the population
value of 0.5; this result may not be surprising since the low volatility subsample includes 95% of the
data.
Bivariate normal random numbers, ρ = 0.5

-2

-4
-4 -2 0 2 4
ρ ( (x,y) | |x|<1.96)=0.45, ρ ( (x,y) | |x| >= 1.96) = 0.81

15
The distributions of x and y are standard normal by assumption. Hence, the absolute value of x is less than 1.96 with a
probability of 95%.

30
 BIS Quarterly Review, June 2000

A formal result
The intuitive link between volatility and correlation can be derived formally. Boyer et al (1999) provide the
following theorem.
Theorem. Consider a pair of i.i.d. bivariate normal random variables x and y with standard deviations σ x
and σ y , respectively, and covariance σ xy . Let ρ ( = σ xy /(σ xσ y ) ) denote the unconditional correlation
between x and y . The correlation between x and y conditional on an event x ∈ A , for any A ⊂ ¶ with
0 < Prob( A) < 1 , is given by

−1 / 2
 σ x2 
ρ A = ρ  ρ 2 + (1 − ρ 2 )  (1)
Var( x | x ∈ A) 
 
Proof. ➀ Let u and v be two independent standard normal random variables. Now construct two bivariate
normal random variables x and y with means µ x and µ y , respectively, standard deviations σ x and σ y ,
respectively, and correlation coefficient ρ :

x = µ x + σ xu (2)

y = µ y + ρσ y u + 1 − ρ 2 σ y v (3)

Consider an event x ∈ A , for any A ⊂ ¶ with 0 < Prob( A) < 1 . By definition, the conditional correlation

coefficient between x and y , ρ A , is given by

Cov( x, y | x ∈ A)
ρA = (4)
Var( x | x ∈ A) Var( y | x ∈ A)

By substituting for u in (3) using equation (2), then substituting the resulting expression for y into (4), and
using the fact that x and v are independent by construction, one can rewrite this as
( ρσ y / σ x ) Var( x | x ∈ A)
ρA = (5)
Var( x | x ∈ A) ( ρ 2σ 2y / σ x2 ) Var( x | x ∈ A) + (1 − ρ 2 )σ 2y

which can, in turn, be simplified to yield the expression in (1).

Thus, the conditional correlation between x and y is larger (smaller) than ρ in absolute value if the
conditional variance of x given x ∈ A is larger (smaller) than the unconditional variance of x .
_____________________________
➀
This proof is based on the property of bivariate normal random variables that each component can be expressed as the
weighted average of the other and of an independent variable that is also normally distributed. See, for example,
Goldberger (1991, p 75).

The theoretical link between volatility and correlation holds in a time series context as well. Consider
subdividing a long time series of two variables, x and y, which are observed daily, into quarterly
subsamples. For each subsample, calculate the variance of x and the correlation between x and y.
Finally, order the subsamples by the variance of x. The table on the next page shows the results of
such an exercise under the assumption that x and y are independent and normally distributed, with unit
variances, and a constant correlation coefficient equal to 0.5 (as in the graph on the previous page).
The first column of the table shows ranges for the ratio of the quarterly sampling variance in x to its
population value (which is 1). The other three columns show the distribution of quarterly correlation
values for the samples in those ranges. For quarters with in-sample variance of x close to its population
value (0.9 to 1.1), the median sampling correlation is 0.50. However, the distribution of sampling

31
BIS Quarterly Review, June 2000

In-sample correlations when conditioning on volatility

Range of variances of x Conditional correlation of x and y

relative to its population variance Bottom 5% Median Top 5%

0.3–0.5 0.17 0.36 0.53

0.5–0.7 0.24 0.43 0.58
0.7–0.9 0.29 0.47 0.61
0.9–1.1 0.34 0.50 0.64
1.1–1.3 0.38 0.54 0.67
1.3–1.5 0.41 0.57 0.69
1.5–1.7 0.45 0.59 0.71
1.7–1.9 0.48 0.61 0.72

Note: The random variables x and y are i.i.d. ELYDULDWH QRUPDO ZLWK D SRSXODWLRQ FRUUHODWLRQ FRHIILFLHQW  RI  5HSRUWHG YDOXHV IRU
Corr(x,y) are based on 2.5 million random draws of “quarters” (consisting of 60 “daily” data pairs). There were too few observations with
a variance of x less than 0.3 or greater than 1.9 times its population value for values of Corr(x,y) to be reported with confidence.

correlations is fairly wide, with a 90% confidence interval running from 0.34 to 0.64. In contrast, for
quarters with in-sample variance of x between 1.7 and 1.9 times its population value, the median
correlation is 0.61, with the 90% confidence interval running from 0.48 to 0.72. In other words, in this
time series example, periods of increased sampling volatility are also periods of relatively high
measured correlations, even when the population correlation remains constant.

An empirical application
In order to assess the real-world applicability of this theoretical link between volatility and correlation,
we need to consider whether it can explain the historical relationship between pairs of asset returns.
Are contemporaneous changes in sampling variances and sampling correlations empirically consistent
with an unchanged underlying distribution of asset returns, and, in particular, with a constant
population correlation?
We consider stock prices as measured by the FTSE and Dax stock price indices.16 These data series
represent large and liquid markets and reflect market conditions at roughly the same time, and so we
do not have to be concerned about the implications of non-synchronous data collection.17 Our data are
daily observations from the beginning of 1991 to the middle of 1999. The returns are calculated as
daily percentage changes in the respective price indices.
The graph on the next page shows time series plots of the within-quarter variances (left-hand panel)
and correlations (right-hand panel) of the daily stock market returns. It is clear that autumn 1998 was a
period of high volatility and, just as the theoretical results would suggest, one of elevated correlation.
To evaluate the importance of the theoretical link between volatility and correlation more generally,
we show in the graph on page 34 a scatterplot of the quarterly in-sample correlations against the in-

16
In Loretan and English (2000) we present results for returns on government bonds and foreign exchange as well. See also
Forbes and Rigobon (1999) for a detailed examination of the link between volatility and correlation in equity prices. The
FTSE and Dax data are from Bloomberg, and reflect closing quotes.
17
For a discussion of the problems associated with non-synchronous data collection, see RiskMetrics (1996,
pp 184-196).

32
 BIS Quarterly Review, June 2000

sample volatility of the return on the Dax (the lines in the graph are discussed below).18 The graph
clearly shows a generally increasing relationship between the sample variances and sample
correlations; the observations for the final two quarters of 1998 comprise two of the three observations
at the top right. Although the upward slope in the graph on the next page is consistent with theoretical
expectations, the data also show a considerable dispersion in the sample correlation for a given level of
sample volatility. In order to provide a more compelling test of whether the population correlation is
constant, we need to determine whether the empirical relationship lies mostly within a confidence
band around the expected average relationship between volatility and correlation, where the expected
relationship and the confidence band are based on the assumption of a constant distribution of the asset
returns. One way to construct the theoretical expectations and confidence band is to use a bootstrap,
which is based on repeatedly drawing observations from the actual data. Specifically, we select a
random sample of a quarter’s worth of observations (60 pairs of returns) from the observed data series
and calculate the sample variances of the two returns and the sample correlation between the return
series. We then repeat the process a large number of times (2 million random samples in total), thereby
producing a very large number of correlation-variance pairs. We then use these random observations
to calculate the median value of the correlation as a function of the volatility as well as 90%
confidence intervals around that median.19 The resulting lines are plotted in the graph on the next page.

Within-quarter variances and correlations, Dax and FTSE indices

Within-quarter variances Within-quarter correlations (in %)

4 100

Dax
FTSE
3 75

2 50

1 25

0 0
91 92 93 94 95 96 97 98 99 91 92 93 94 95 96 97 98 99

18
The correlations could be plotted relative to the volatility of either return; use of the FTSE index yields similar results.
Note that the within-quarter variance of the return on the Dax has been expressed relative to its full-sample value.
19
Note that the median correlation and the confidence intervals are based on the actual distributions of the data series rather
than on an assumed distribution, such as the bivariate normal. Our earlier study, Loretan and English (2000), shows both
the bootstrap results and those based on a bivariate normal distribution. The median values are similar, but the confidence
contours are wider under the bootstrap; this appears to be due to the fact that the actual returns have more outlier
observations than would be implied by a normal distribution. The bootstrap procedure preserves the unconditionally
heavy-tailed nature of the distributions as well as the contemporaneous correlation structure of the data. However, it does
not take account of serial dependence features such as GARCH, which, as discussed in Loretan and English (2000),
appear to be present in the data.

33
BIS Quarterly Review, June 2000

Quarterly variances vs quarterly correlation, bootstrapped confidence bands

In percentage points¹

100

80

60

40

20

0
0 1 2 3 4
¹ The vertical axis shows the quarterly correlation of the Dax vs the FTSE. The horizontal axis shows the relative quarterly variances of
Dax returns.

The equity data fit the pattern implied by the simple theory surprisingly well.20 The observations are
scattered fairly evenly around the median line, and only a few of the 34 observations lie outside the
90% confidence contours. While a more comprehensive test is beyond the scope of this article, our
results suggest that one should not be too quick to conclude that fluctuations in correlations during
periods of market volatility, including those observed in the second half of 1998, represent true
changes in the distribution of asset returns. Rather, they may be nothing more than the predictable
consequences of observing certain (low probability) draws from an unchanged distribution. This
conclusion need not imply that “contagion” does not occur: rather, it suggests that if one defines
contagion to mean elevated sample correlations between asset returns, then contagion can be a natural
by-product of high sampling volatilities.

Implications
The statistical link between sampling volatilities and correlations of asset returns has important
implications for the evaluation of portfolio risk by market participants and investors as well as for the
supervision of financial firms’ risk management practices.
Risk managers sometimes use data from a relatively short interval when calculating correlations and
volatilities for use in risk management models. Some estimation methods are based on longer intervals
of data, but they apply geometrically declining weights, thereby reducing the effective number of
observations employed. The theoretical and empirical results presented here suggest that the use of
relatively short intervals of data for estimating correlations and volatilities may be dangerous. If the
interval happens to be atypically stable, then not only may the estimated volatilities be too low, but,
perhaps more important, the estimated correlations between returns will be lower than average. As a
result, assessments of market risk may overstate the amount of diversification in a portfolio, leading
the investing firm to take on excessive risk. Conversely, if the interval of data employed is a relatively

20
Our results are based on the volatility of asset returns with no distinction made between increases and decreases in asset
prices. In a related study, Longin and Solnik (1998) find that measured correlations between equity returns in different
countries behave as the theory would suggest when there are large positive stock market returns but are higher than the
theory would suggest when there are large negative returns. We leave an examination of this issue for future research.

34
 BIS Quarterly Review, June 2000

volatile one, then the resulting estimates of correlations will be atypically high and could lead the firm
to take positions that are excessively risk-averse.
This does not necessarily imply that the use of longer time series produces more reliable calculations.
Indeed, short intervals have some desirable features. Since financial markets can change over time,
one may not want to depend on data from the distant past.21 Moreover, the emphasis on recent data
allows account to be taken of time-varying volatility, which appears to be a feature of actual returns.
However, our results suggest that when determining the appropriate time interval to use, risk managers
should not exclude periods of relatively high or low volatility. Such periods contain important
information about the underlying relationship between asset returns.
Another way in which the link between in-sample volatility and correlation could cause problems for
risk managers is in the calculation of worst case scenarios and in stress testing. Put simply, risk
managers should not consider the possible effects of high return volatilities without also taking into
account the higher correlations between asset returns that would generally accompany the elevated
volatility (see Ronn (1995) for a related discussion). One way to do so would be to employ
information from historical periods of high volatility in order to form estimates of correlations
conditional on being in a period of heightened volatility.22 These conditional correlations could then be
used to evaluate the distribution of returns under a high volatility scenario. Put differently, the method
used for stress testing a portfolio must not (inadvertently) exclude the empirical feature that periods of
high volatility are also likely to be periods of elevated correlation.
Supervisors of financial institutions also need to be aware of the link between volatilities and
correlations when assessing firms’ risk management practices. For example, in evaluating such firms’
internal models, supervisors need to keep in mind the difficulties noted earlier with relying on a
relatively short interval of data for information on correlations and the need to form appropriate
conditional correlations for stress tests.

References
Bankers Trust (1995): Annual Report 1994.
Boyer, Brian H, Michael S Gibson and Mico Loretan (1999): “Pitfalls in Tests for Changes in
Correlations”, International Finance Discussion Paper No. 597R, Federal Reserve Board, March.
Chase Manhattan Corp (1999): Annual Report 1998.
Committee on the Global Financial System (1999): A Review of Financial Market Events in Autumn
1998, Bank for International Settlements, October.
Drazen, Allan (1998): “Political Contagion in Currency Crises”, mimeo, University of Maryland.
Eichengreen, Barry, Andrew Rose and C Wyplosz (1996): “Contagious Currency Crises”, NBER
Working Paper No. 5681, July.
Forbes, Kristin and Roberto Rigobon (1999): “No Contagion, Only Interdependence: Measuring stock
market co-movements”, NBER Working Paper No. 7267, July.
Gerlach, Stefan and Frank Smets (1995): “Contagious Speculative Attacks”, European Journal of
Political Economy.

21
Similarly, if the assets under consideration are firm-specific (rather than indices), the behaviour of firms can change over
time as managers or business strategies are changed, making older information less useful.
22
Alternatively, firms might want to use actual data from earlier periods of high volatility to stress test their portfolios. For
example, Chase Manhattan uses asset price movements during three historical episodes - the bond market sell-off in
1994, the 1994 Mexican peso crisis and the 1997 Asian markets crisis - as well as internally developed scenarios, when
assessing the risk of its portfolio (Chase Manhattan (1999, p 37)).

35
BIS Quarterly Review, June 2000

Goldberger, Arthur S (1991): A Course in Econometrics, Harvard University Press.

Kodres, Laura and Matthew G Pritsker (1999): “A Rational Expectations Model of Financial
Contagion”, mimeo, Federal Reserve Board, May.
Longin, François and Bruno Solnik (1998): “Correlation Structure of International Equity Markets
During Extremely Volatile Periods”, mimeo, Group HEC, April.
Loretan, Mico and William B English (2000): “Evaluating ‘Correlation Breakdowns’ during Periods
of Market Volatility”, BIS Conference Papers, Vol 8, Bank for International Settlements, March.
RiskMetrics (1996): RiskMetrics - Technical Document, 4th Ed, JP Morgan/Reuters.
Ronn, Ehud I (1995): “The Impact of Large Changes in Asset Prices on Intra-Market Correlations in
the Stock and Bond Markets”, mimeo, Department of Finance, University of Texas at Austin, July.

36