When Swans Are Grey

When swans are grey: VaR as an
early warning signal

Received (in revised form): 7th June, 2010
Daniel Satchkov
is the President of RiXtrema Inc., a risk modelling and consulting firm that focuses on the extreme financial
market events. Prior to founding RiXtrema, he was an Associate Director of Risk Research at FactSet, where he
was responsible for researching and developing software products in the areas of risk measurement and risk
reporting. He has spoken at numerous conferences and published articles dealing with risk management
issues in such magazines as Journal of Asset Management, Investment and Pensions Europe, as well as in a
number of white papers and an e-book. Daniel’s current research is in the area of extreme market events,
credit cycles and behavioural finance. Daniel holds BS and MBA degrees from the University of the Pacific.
RiXtrema Inc., 14–17 150 Street, Whitestone, NY 11357, USA

Tel: þ1 (917) 600 9679; E-mail: d.satchkov@rixtrema.com
Abstract The market events of 2008 will be remembered as much for their
extreme volatility as for a widespread failure of the risk management, which
contributed to the near collapse of many firms thought to be among the leaders
in that field. This paper identifies a key deficiency in the way that the historical
data are currently utilised in the estimation of risk. This deficiency stems from
the conception of the marketplace as an equilibrium-seeking and continuous
system and it led to the financial firms’ unpreparedness for sudden market
reversals. A different framework for risk estimation is proposed based on linking
the risk modelling with the existing literature on financial instability. One
possible application of the proposed method to the estimation of value-at-risk
(VaR) is demonstrated, and empirical tests comparing it with the traditional
methods are performed using S&P 500’s history from 1989 to 2010. The new
measure, called the instability VaR, is shown to dominate all traditional
methods of calculation.
Keywords: market crises, VaR, early warning indicators, market bubbles,

risk mispricing, financial instability hypothesis
INTRODUCTION financial industry must look for

Market events of 2008 will be significant improvements in the
remembered as much for their extreme performance of risk estimates. This
volatility as for a widespread failure of the analysis proposes a fundamental change
risk management, which contributed to to the way that historical data are used in
near collapse of many firms, had various the estimation of risk. It will show how
causes and went far beyond deficiencies the accuracy of one common metric
in risk measurement techniques. called value-at-risk (VaR) can be
Nevertheless, underestimation of risk significantly improved with the
played a real part and the obvious lesson application of this new method, which is
of that disaster should be that the motivated by the idea of self-sustaining
366 Journal of Risk Management in Financial Institutions Vol. 3, 4 366–379 # Henry Stewart Publications 1752-8887 (2010)
When swans are grey: VaR as an early warning signal
boom-bust patterns and their relation to historical VaR are all different measures
pricing of risk. While the paper will of the risk metric called VaR.
focus on VaR, the ideas presented here VaR has a statistical definition (it is a
are general and can be applied to the quantile of a distribution), and nothing
estimation of any risk statistic. in this definition implies that it has to be
based on a bell curve calibrated
exclusively on recent returns as is the
PROBLEMS WITH CURRENT case with most accepted models today. In
VAR IMPLEMENTATION other words, there are different ways of
VaR is the most commonly used metric measuring the metric called VaR.
of market risk. However, VaR is Extreme value theory methods do not
frequently the subject of criticism, and assume any particular overall distribution
this paper will consider some of the and Monte Carlo simulation VaR can be
arguments against it that carry the most based on many distributions, including
validity. various power law distributions that
The most often voiced objection to Taleb favours.1 These distributions can
VaR stems from the obvious model the effects of the ‘fat tail’, which
inappropriateness of the normal frequently and painfully asserts itself in
distribution for the modelling of the financial markets, although, as will be
medium- to short-term financial returns. seen, there are practical problems with
The fact that the short-term returns are those versions of VaR as well. Lack of
not distributed according to a bell curve awareness of the basic distinction
can be hardly disputed at this point. It is between the measure and the metric
enough to consider the simple fact that sometimes leads to definitions of VaR
daily returns of the S&P 500 display a which state that it is only valid during
kurtosis of greater than 7 over the three ‘normal conditions’. It is astonishing that
years even prior to August 2008. A anyone would even consider using a risk
popular version of this argument against measure that only functions when one
VaR can be found in Nassim Taleb’s does not need it. The goal in this study
Black Swan: The Impact of the Highly is not to discard VaR, but rather to make
Improbable.1 However reasonable it sure it better accounts for the drastic
appears, this argument should not be market reversals that seem to always catch
levelled at VaR as a metric of risk, but quantitative risk systems by surprise.
rather at one particular way of calculating
VaR, namely the parametric or normal THE REAL PROBLEM: EQUAL
distribution VaR based on a particular WEIGHTED AND DECAY TIME
way of incorporating historical data. WEIGHTED ESTIMATORS
This misunderstanding is based on a The main problem that makes the
confusion of the concepts of measure current risk measurement dangerously
and that of the metric. Measure is an inadequate lies not in a distributional
operation for assigning a number to assumption, but rather in the way that
something; there could be many ways of the historical data is incorporated in the
doing so. Metric is the interpretation; for risk estimates. The current methods for
example, VaR is a metric, see Holton.2 using past observations gradually became
Parametric VaR, Monte Carlo VaR and axioms that are completely divorced from
# Henry Stewart Publications 1752-8887 (2010) Vol. 3, 4 366– 379 Journal of Risk Management in Financial Institutions 367
Satchkov
the market realities or any new economic estimate leaves the risk analyst completely
thinking. Currently, there are two unaware of what extreme conditions
common methods for incorporating the could do to their portfolio. The
past observations into the estimates of historical data gathered during the period
risk metrics. This paper will call them of moderate or low volatility that
the equal weighted (EW) and decay time persisted just before the crisis will
weighted (DTW). The equal weighted severely understate the risk potential in
estimator of (co)variance (for a normal the market. Problem 2: In addition, the
distribution, which is being used here, equal-weighted estimator adjusts very
the risk and variance are assumed to be slowly to changes in the market
synonymous) is derived directly from conditions and may indicate low risk
statistics where an unbiased estimator of long after it is obvious to even the most
(co)variance is equal to: casual observer that the high-volatility
period has commenced.
1 X
n
EW 2
s j;k ¼ ½r j;i rk;i ð1Þ The DTW estimator is an attempt to
n 1 i¼1 correct the second of these problems by
making the estimate more sensitive to the
where: recent market conditions. The DTW
r j;i — return of asset j or k at time i (co)variance estimate could be written as:
n — number of observations. TDW
The VaR is then calculated as: s2j;k
2 3
VaR ¼ ks j;k ð2Þ Xn 6 li1 r r
6 j;i k;i 77
¼ ð1 lÞ 4Pn 5
where: i¼1 ð1 lÞ l l1
k — scaling based on the confidence l¼1
level of VaR. n h
X i
1
In terms of economics the implicit ¼P
n li1 r j;i rk;i
assumption in the use of this method for ll1 i¼1
l¼1
modelling variability of the returns is that
a system is being observed that is ð3Þ
relatively stable over the time period of
observation and is projected to continue
l — exponential decay factor
l,i — indicators of time
this behaviour into the future period
covered by the risk forecast. This rj,i — return of asset j or k at time i.
assumption seems to be inspired by the It can be seen that in this calculation
economic theories which purport the the more recent periods are meant to
existence of a long-term equilibrium in carry progressively more information
the economic and financial system. and are weighted accordingly. The
Nevertheless, there are two obvious implicit economic assumption can be
practical problems with this approach for understood in relation to the EW
the purposes of the risk estimation. estimator and consists of allowing the
Problem 1: If a sudden crisis erupts market and the financial equilibrium to
after a period of moderate volatility and vary more rapidly with time. The
low or moderate correlations, the risk DTW approach does indeed solve the
slow adjustment problem mentioned warning of the impending 2008

above. Nevertheless, it cannot fix the meltdown.
more important first problem and still The new paradigm of risk
leaves a risk manager unprepared for measurement must minimise the
the possibility of a sudden change in extrapolation from quiet periods to the
the market structure after a period of tails of the distribution. A market
moderate volatility and correlations. In structure during the extreme event is
other words, although it allows for likely to be drastically different from the
changes in the market structure, it is one observed in the trading range or even
still only a reflection of the presently in the moderate volatility periods before
observed volatility and assumes a the crisis. Incorporating this fact does not
certain continuity that is simply not necessarily mean that a financial firm
observed in practice. should be permanently braced for the
As an example, consider Figure 1. In it crisis, an approach that would be neither
can be seen S&P 500 returns and the practical nor healthy for the economy.
two methods of estimating 99 per cent Instead one should look for ways to
VaR for the S&P 500, one using the EW correct the assumptions embedded in the
method and the other using the DTW current process to improve the accuracy of
method. The exponential weight is 0.94, the risk estimates.
which was the best performing weight in
this sample for the DTW. It has been ENDOGENOUS FINANCIAL
used because it allowed complete benefit INSTABILITY
of the doubt to the existing method. To proceed further one must identify
Figure 1 confirms something that is some theoretical framework that
widely recognised, namely that both resembles reality to a greater degree than
existing methods gave absolutely no the aforementioned assumptions of a
Figure 1: Currently accepted methods in a run-up to 2008 crash
Satchkov
relatively stable ‘equilibrium’ state. There unstable. The second theorem of the
are a number of economic approaches financial instability hypothesis is that over
that can be helpful in this regard. De periods of prolonged prosperity, the
Long et al. 3 show a model of market economy transits from financial relations
bubbles in which rational traders who that make for a stable system to financial
follow positive-feedback strategies are relations that make for an unstable
buying with rising prices and selling system.’5
with falling prices, thus producing
self-sustaining trends which ultimately Soros6 extensively discusses his view of
end in a crash. The situation of demand boom-bust sequences. His description,
rising with the price not infrequently although less rigorous than Minsky’s, gives
encountered in financial markets upsets a useful view into the mechanism of
traditional supply-demand relationships self-sustaining bubbles via the mechanism
and makes traditional equilibrium he calls ‘reflexivity’. Economy can deviate
approaches incapable of dealing with the very far from a theoretical equilibrium for
real world fluctuations. De Long et al.’s long periods of time, because many
model formalises a permanent theme in so-called ‘fundamentals’ under certain
the literature on self-reinforcing bubbles conditions can become highly intertwined
which goes back as far as Bagehot.4 with prices, which are supposed to reflect
Going further, Hyman Minsky5 them. The views of both Minsky and
identified the key features of the credit Soros explain how the risk gets built up to
cycle which tend to drive large an extremely high level owing to purely
boom-bust sequences. According to the endogenous market forces. The stage is
financial instability hypothesis (FIH), then set for a dramatic reversal. In an
fundamental relationships in the real exceptionally lucid explanation of systemic
economy and financial markets change risk, Danielsson and Shin7 write:
with the change in the behaviour of the
participants and particularly with the ‘One of the implications of a highly
change in the behaviour of the financial leveraged market going into reversal is that
intermediaries. For example, after a a moderate fall in asset value is highly
period of prosperity, an increase in the unlikely. Either the asset does not fall in
risk-taking activities takes place and value at all, or the value falls by a large
rising leverage builds, setting up the amount.’
potential for a violent downturn. Some
interruption will expose the Let us now summarise these insights in a
unsustainability of leverage levels leading framework that will help to develop a
to a credit contraction and a potential risk modelling approach, which can truly
collapse in the asset values. estimate risks, that is provide an early
Minsky summarised his insights in two warning of the instability potential.
theorems of the financial instability:
FRAMEWORK FOR THE
‘The first theorem of the financial INSTABILITY RISK APPROACH
instability hypothesis is that the economy (1) Financial market behaviour can be very
has financing regimes under which it is roughly divided into two states: a stable
stable and financing regimes in which it is state where a financial economy is an
equilibrium-seeking system producing extreme value distribution (GEVD). It

random deviations and an unstable one relies on the asymptotic properties of the
where it becomes a deviation- independent and identically distributed
amplifying system. (Of course, there are (i.i.d.) sample maxima and in that sense
many different specifications possible is similar to the central limit theorem for
but the aim is to capture only the most sums of i.i.d. variables. If one now makes
important aspects to build tractable an assumption that extreme observations
models.) gathered from different periods over a
(2) Currently used weighting schemes long time are independently and
allow the latest observed data to identically distributed, then one can
dominate the sample in any conditions. gather enough historical outliers to
This leads to a severe understatement of estimate the parameters of the GEVD
risks and an overstatement of any using the maximum likelihood methods.
diversification benefits during the low A good summary of the EVT approaches
volatility/low correlation stable phase can be found in Embrechts et al. 8 (1997).
and to an overstatement of risks in the EVT has been shown to be a powerful
bottom of the cycle. tool, see for example, Longin9 or Phoa,10
(3) An unstable state is characterised by the with a much better predictive power in
multiple feedback loops of liquidation the tail than attained by the current
resulting in a unique environment approaches. Still, there are some serious
which has no linear relation to the obstacles to its practical use in everyday
stable state. Therefore, data gathered in risk management and it is that use that
stable periods carry exceedingly little one is most concerned with here. EVT
value for the estimation of risks. presupposes unconditional distribution of
(4) As the risk taking behaviour in the the extreme values and is a purely
economy grows, a risk statistic should statistical framework with no linkages to
assign a greater weight to the data the economic conditions at large. Thus,
gathered during the unstable periods. even if it accurately represents the risk of
a tail event, it leaves the risk manager
with a risk estimate that is more or less
It is important to note that the first three constant through time. (This is because
assumptions, while contrary to the EVT methods use all available extreme
currently prevailing paradigm, are not observations in history to satisfy the
new, since they are implicitly asymptotic properties and fill the sample,
incorporated in the extreme value theory so that new extreme observations usually
(EVT) approach to financial risk do not significantly change the estimate.)
estimation. EVT is a statistical This constant estimate is bound to be
framework, which has been extensively quite high. The result is that, if this
used by financial risk researchers to estimate is acted on, it will produce a
overcome the deficiencies of the EW and permanent bracing for the crash, a state
DTW approaches and the resulting focus which was explicitly alluded to earlier, as
on stable periods. EVT relies on a neither feasible in a competitive financial
theorem, which postulates that extreme economy nor desirable for the economic
values from different distributions can be development. In other words, it may
modelled as coming from the generalised correctly estimate the tail, but it says
Satchkov
nothing about the potential changes in are met in relation to the risk
its likelihood. This is why it is necessary (mis)pricing metrics.
to draw on the ideas outside of a purely In the interest of the simplicity of
statistical framework to estimate the presentation, one can take the most basic
instability potential of the economy. The and perhaps most widely used type of
author here argues that statistics should VaR, the parametric VaR, and start with
be strictly subordinate to the economic the EW and DTW methods of using the
reasoning in risk modelling and that past data for its estimation. One will then
heuristics should be used where they are introduce an instability estimator of
necessary. parametric VaR. The testing of the
instability estimator will show that, even
INSTABILITY ESTIMATOR OF with an obviously simplistic parametric
THE PARAMETRIC VAR VaR, it can do a good job of capturing
The conclusions in the above framework tail risks.
lead to the new paradigm for the
inclusion of the past data in the risk
estimation process. As the excess Construction of the instability
risk-taking persists, a financial economy estimator of parametric VaR
becomes increasingly vulnerable to all EW and DTW estimators of VaR have
types of shocks. This framework already been described above. This paper
motivates the following definition of a will now describe the instability estimator
new class of risk estimates: and run back tests comparing the three
of them. Since parametric VaR is being
Instability estimator of risk (IER) is used, the instability estimator of
calculated by assigning progressively (co)variance will take the following
greater weight to the observations from form:
the extreme portion of the sample when
c h i
two conditions are present: INST 1 X
s2j;k ¼ WEX r EX r EX
n t¼1 j;t k;t
(1) Risk is mispriced for a period of

time þ ð1 WEX Þ ðEW s2j;k Þ
(2) The trend of worsening risk ð4Þ
mispricing stops and shows signs
of reversal. where:
WEX — weight assigned to the extreme
Risk mispricing can potentially be observations (similarly to EVT methods,
measured by a variety of more or less in order to find enough of these extreme
easily observed metrics. In this class of observations one must widen the
metrics one can include available sample as much as possible. In
price-to-earnings ratios, junk credit this study, the available sample for
spreads, housing price-to-rent ratios, extreme observations used starts on 31st
sovereign spreads, financial sector December, 1930)
leverage and others. The extreme sample c — is the number of observations that
observations will carry more weight satisfy the criteria to be chosen as
when two conditions in the definition extreme data points
r EX
j;t — selected extreme period return of the concept to be tested without delving
the asset j or k at time t into the issues of the generalisation of the
and the instability estimator of parametric IER concept to the multiple asset/
VaR for asset j: multiple factor model.
With the above considerations in mind,
VaR ¼ k ðINST s j;j Þ the weighting function can be introduced:
where: WEX;i ¼ MIN

k — scaling based on the confidence 0 0 1
level of VaR. DCPE;i þDCJS;i
B B C !
In practice there are two key choices that B B þ2 DCDJS;i C
need to be made: B
BMAX B
B ;1C
C1; 0:3
@ @ 20 A
(1) How to separate the stable from
unstable periods?
ð5Þ
(2) What is the form and parameters of
the weighting function in the
where:
formula above, in other words, how
i — time at which the risk (mis)pricing
to measure the risk-taking activity
and its reversal?
metric is observed and VaR is estimated
Risk mispricing metrics:
The answer to the first question can be DCPE;i — reverse decile of one-year
fairly straightforward. Since one cares average PE in the historical sample of all
about the tail events that are not well such average PEs (where deciles are made
captured by the existing methods, one fractional through multiplying decimal
can label as ‘unstable’ all of the past percentage ranks by ten, ie a rank of one
observations that were outside the 2.33 (highest possible value) becomes a decile
deviation band (that would be equivalent rank (that is reverse decile) of ten
to violating 99 per cent parametric VaR), (highest possible contribution to weight))
where standard deviation is calculated DCJS;i — decile of one-year average junk
using the EW method using all of the spread
history available. Risk mispricing reversal metric:
The second problem is considerably DCDJS;i — reverse decile of 180-day
trickier. The goal is to test the IER change in the average one-year junk
concept on a broad market index, spreads (sources of data are Standard and
specifically the S&P 500, so one needs to Poors, Merrill Lynch, and Haver
choose only the most broadly applicable Analytics)
measures of risk pricing. For the
risk-pricing part of the definition a It is clear from this weighting function
trailing one-year average price to that the weight of extreme periods
earnings ratio of the S&P 500 and a should be highest when smoothed PE is
trailing one-year average credit junk high, smoothed junk spreads (JS) are low
spread will be used. For the metric of and 180-day change in smoothed junk
risk-pricing reversal part a 180-day spreads (DJS) is high (ie spreads are
percentage change in the average credit trending up after a period of risk
junk spread will be used. This will allow mispricing). The combined weight of PE
Satchkov
and JS as representing risk mispricing is VISUAL TESTING

equal to one half, while the weight of Figures 2 –5 summarise in graphical form
DJS, as representing the warning of the some of the key advantages of the IER
end of mispricing, is also equal to one method. In Figure 2 it is seen that, just as
half. All this follows the definition of the in 2008, both EW and DTW methods
IER above. The combined deciles are did not provide any early warning about
divided by 20 and one is subtracted from the Long Term Capital Management
the total. The highest possible value of (LTCM) crash. Instability VaR on the
the sum of deciles is 40, so the weight contrary showed sharply increased risk
of the extreme sample will vary from after the Asian currency crisis and kept
zero to one directly proportionally with the risk elevated leading up to August of
the increase in the sum of deciles of the 1998, just as LTCM loaded up on
original signals. The maximum function short-volatility trades they viewed as
ensures that when the sum of signals is exceptionally attractive, perhaps because
below 20, the weight is zero and not they were using the standard methods of
negative, while the minimum function VaR calculation.
caps the total weight of the extreme Figure 3 shows a similar story prior to
periods to 30 per cent of the sample to the dot-com bubble crash. Both EW and
avoid calculating VaR based on only a DTW methods showed low risk and
small number of observations in the tail. DTW even showed a significant decrease
(The results are not changed significantly in VaR at the worst possible moment
with small or even moderate changes in just a few weeks leading up to the
these settings, thus indicating robustness.) beginning of meltdown in April 2000.
It should be clear that formula (5) Instability VaR on the other hand
occupies in the IER paradigm the same showed a dramatic increase in risk in
place occupied by the exponential decay January 2000.
weighting in the presently accepted one, While early warning should be a
that is, its purpose is to indicate which critical objective for risk models, it is
periods carry more valuable information. important that financial firms put
The present paradigm assumes that recent themselves in a position to benefit from
periods are always more valuable, while the boom and not be permanently
the IER is based on the assumptions that braced for crisis. (Of course, the issue is
there are other factors that govern this being discussed from the perspective of
relative importance of data points. the individual user given the financial
Let us summarise what would be system that is in place now. The
expected from the instability estimator aggregate effects on the economy are a
based on the formulae (4) and (5). The different story and regulators would do
instability estimator is essentially a well to consider other uses for this
weighted average of the EW estimator model to possibly smooth out the
and a similar calculation, but based only risk-taking cycle.) Figure 4 shows that
on the extreme periods (the first term on this would indeed be the case with the
the right-hand side of formula (4)). The instability VaR approach. For the
weight between those two terms varies period from the beginning of 2002 to
with the signals for the risk-taking the end of 2006, the risk would
activities of the boom-bust cycle. largely be the same as under EW and
Figure 2: Run-up to the LTCM crash (7/1997–12/1998)
DTW methods. The two exceptions author does not view these false alarms
are a few months at the end of 2004 as failures of the model; rather they
and the second half of 2005. The represent ‘tests of the bubbles’, as
Figure 3: Run-up to the dot-com meltdown
Satchkov
Figure 4: Benefiting from the boom
discussed by Soros.6 The only way to Basel VaR testing. The traffic light test,
distinguish a burst of the bubble from also called the binomial test, is based
a ‘test’ is ex post. on the idea that, with a large enough
Figure 5 repeats Figure 1, but with sample, an adequate VaR model should
the instability VaR added. Instability have the percentage of its violations
VaR clearly showed increased risk being close to 1 per cent for the 99
starting in July 2007 and did not per cent confidence VaR.
decrease despite numerous minor rallies It is obvious from Table 1 that both
and some ‘lull before the storm’. The EW and DTW methods show
performance of the EW and DTW in significant deficiencies in all periods.
this period was already discussed. Instability VaR, on the other hand,
shows good results for all periods, even
STATISTICAL TESTING the most challenging ones for a risk
The results of the testing are model, such as the one from the end of
summarised in Tables 1 and 2. Table 1 2006 to the end of 2008.
shows the percentage of time that the Finally, Figure 3 shows the statistical
actual return for the S&P 500 breaks test based on the likelihood ratio (LR)
through the 99 per cent VaR band. suggested by Kupiec.12 LR is given by
This test is based on Basel traffic light the following formula:
test11 but we will use a much bigger
sample that encompasses different
portions of the cycle to avoid the ð1 qÞnv qv
LR ¼ 2 LN ð6Þ
deficiency of the short sample in the ð1 rÞnv r v
Figure 5: Run-up to the 2008 crash
where: This statistic is the chi-square distributed

n — number of observations in the with one degree of freedom. Using the
sample chi-square tables, one can infer the lower
v — number of violations ie number of and upper bounds for the number of
times that 99 per cent VaR was breached violations that would need to be
q — desired violation rate observed in order not to reject the
r — empirical violation rate ie v/n. hypothesis that the 99 per cent VaR is
violated 1 per cent of the time (ie that it
The null hypothesis being tested here is:
is likely a 99 per cent VaR or something
H0 : r ¼ q close to it is being produced).
Table 1: Percentage of periods where 99 per cent VaR was violated

Per cent violations (Basel)
Period EW (%) DTW (%) Instability (%)
31st December, 2005 to 21st May, 2010 2.99 2.90 1.54
26th April, 1989 to 21st May, 2010 1.82 1.93 1.03
31st December, 2006 to 31st December, 2008 4.75 3.56 1.58
Satchkov
Table 2: Actual number of violations along with 10 per cent (5 per cent on each side) confidence
range
Number violations (significance)
10% confidence
Period bound EW DTW Instability
31st December, 2005 to 21st May, 2010 From 6 to 18 33 32 17
26th April, 1989 to 21st May, 2010 From 40 to 68 97 103 55
31st December, 2006 to 31st December, 2008 From 2 to 9 24 18 8
Again, it can be seen that the allow the well-known deficiencies

instability estimator dominates the group. of VaR to be addressed without
It is also seen that EW and DTW abandoning the metric known for its
estimators are consistently rejected by a ease of use and communication. Further
wide margin. For example, for the whole research into instability risk measures
period from 26th April, 1989 to 21st could proceed in a number of directions.
May, 2010 actual S&P 500 returns violate The first direction would be to adapt the
the EW VaR 97 times and DTW VaR approach outlined above to a multi-asset
103 times. The implied 90 per cent class framework by refining the proxies
confidence range for the correct for the risk-taking activity across various
binomial distribution is between 40 and asset classes. This could take the form of
68 violations. Instability VaR gives 55, looking for a more nuanced and accurate
near the middle of the range. The index of various types of leverage
same situation holds for all other (including hidden leverage) and adding
subsamples. sector and asset class specific risk-pricing
metrics, eg sector-specific credit spreads
CONCLUSIONS or house price to rent for the real estate
The failure of the financial risk market. The second direction could be
estimation to give any warnings of the implementation of advanced
danger is widely known, especially after optimisation algorithms for the data
the 2008 crash. This paper suggests a weighting function. The choice was
modified calculation of the popular made not to do that here in order not to
value-at-risk (VaR) measure with a link clutter the presentation of a novel
to the pricing of risk in the financial measure with the additional complex
economy. Comparing the accuracy of algorithm which itself must be fine
this measure over the past 21 years of the tuned and evaluated, but it would
S&P 500 returns against the commonly present an interesting topic as a separate
used equal weighted (EW) and decay study to find if improvement on the
time weighted (DTW) estimates of presented relatively simple algorithm
parametric VaR shows that it can provide could be obtained. The third extension
significant improvement in the estimation of research would be to apply the
of market risk. Implementation of the concept of instability VaR to different
instability estimator of risk (IER) will distributions like the multivariate Student
t distribution. The author is currently 7 Danielsson, J. and Shin, H. S. (2003)

pursuing all three directions. ‘Endogenous risk’ in ‘Modern
Risk Management: A History’,
References Introduced by Peter Field, Risk
1 Taleb, N. N. (2007) ‘The Black Swan: Books, London.
The Impact of the Highly Improbable’, 8 Embrechts, P., Kluppelberg, C. and
Random House, New York, NY. Mikosch, T. (1997) ‘Modelling Extremal
2 Holton, G. A. (2003) ‘Value-at-Risk: Events: For Insurance and Finance’,
Theory and Practice’, Academic Press, Springer, New York, NY.
Boston, MA. 9 Longin, F. M. (1996) ‘The asymptotic
3 De Long, B., Shleifer, A., Summers, L. distribution of extreme stock market
and Waldmann, R. (1990) ‘Positive returns’, Journal of Business, Vol. 69,
feedback investment strategies and No. 3, pp. 383– 408.
destabilizing rational speculation’, 10 Phoa, W. (1999) ‘Estimating credit spread
The Journal of Finance, Vol. XLV, No. 2, risk using extreme value theory’, Journal
June, pp. 379 –395. of Portfolio Management, Spring, Vol. 25,
4 Bagehot, W. (1872) ‘Lombard Street: A No. 3, pp. 69– 73.
Description of the Money Market’, 11 Basel Committee on Banking
Scribner and Armstrong, New York, NY. Supervision (1996) ‘Supervisory
5 Minsky, H. P. (1992) ‘The Financial Framework for Use of Backtesting in
Instability Hypothesis’, working paper Conjunction with the Internal Models
74, May, Levy Economics Institute, Bard Approach to Market Risk Capital
College, Annandale-on-Hudson, New Requirements’, Bank for International
York. Settlements, Basel, January.
6 Soros, G. (1987) ‘The Alchemy of 12 Kupiec, P. H. (1995) ‘Techniques for
Finance: Reading the Mind of verifying the accuracy of risk
the Market’, Simon & Schuster, measurement models’, Journal of
New York, NY. Derivatives, Vol. 3, No. 2, Winter.
Copyright of Journal of Risk Management in Financial Institutions is the property of Henry Stewart
Publications LLP and its content may not be copied or emailed to multiple sites or posted to a listserv without
the copyright holder's express written permission. However, users may print, download, or email articles for
individual use.

When Swans Are Grey

Enviado por

Dados do documento

Descrição original:

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

When Swans Are Grey

Enviado por

Direitos autorais:

Formatos disponíveis

When swans are grey: VaR as an

early warning signal

RiXtrema Inc., 14–17 150 Street, Whitestone, NY 11357, USA

Keywords: market crises, VaR, early warning indicators, market bubbles,

INTRODUCTION financial industry must look for

slow adjustment problem mentioned warning of the impending 2008

Figure 1: Currently accepted methods in a run-up to 2008 crash

equilibrium-seeking system producing extreme value distribution (GEVD). It

(1) Risk is mispriced for a period of

where: WEX;i ¼ MIN

and JS as representing risk mispricing is VISUAL TESTING

Figure 2: Run-up to the LTCM crash (7/1997–12/1998)

Figure 3: Run-up to the dot-com meltdown

Figure 4: Benefiting from the boom

Figure 5: Run-up to the 2008 crash

where: This statistic is the chi-square distributed

Table 1: Percentage of periods where 99 per cent VaR was violated

Again, it can be seen that the allow the well-known deficiencies

t distribution. The author is currently 7 Danielsson, J. and Shin, H. S. (2003)

Você também pode gostar