You are on page 1of 8

Journal

Article
Journal 36

Multi-fractal Identification of
Sick Markets: The
LIBOR Scandal Case

Journal
The Capco Institute Journal of Financial Transformation

Recipient of the Apex Awards for Publication Excellence 2002-2012

Global Finance and Regulation

#36

02.2013

Sylvain Michael Prado


Ian Rawlinson

SECTION 2

Multi-fractal
Identification of
Sick Markets:
The LIBOR Scandal
Case
Sylvain Michael Prado Consultant, Capco
Ian Rawlinson Managing Principal, Capco

Abstract
Multi-fractal analysis is now widely used in medicine to distinguish healthy and pathological conditions (i.e., healthy
and cancerous tissues). We follow the same approach for
financial markets: fractal tools disclose hidden information
from time series and allow the identification of market abnormalities. Financial regulators can apply these tools to detect fraud and market manipulation. Here we use the LIBOR
scandal as an illustration.

We express our sincere gratitude to Espen A. F. Ihlen for making the Matlab
codes of the multifractal detrended fluctuation analysis (MDFA) available.
The algorithms were implemented in Matlab R2012a using the codes from
http://www.ntnu.edu/inm/geri/software.

113

Introduction: Fractal analysis

fractals to identify abnormalities. Fractal tools allow practical diagnos-

Price behaviors are not smooth they follow booms and busts in healthy

tics and useful prognostics in medicine.2 Patients with congestive heart

markets. Benoit Mandelbrot (19242010) was the founder of fractal anal-

failure and a very high risk of sudden cardiac death, for instance, have

ysis in economics. He argued that the main feature of price records is

heartbeats with altered scaling properties and a breakdown in long-range

roughness and that the proper language of the theory of roughness in na-

correlations [Peng et al. (2000)].

ture and culture is fractal geometry [Mandelbrot (2005), 1]. In 1963, Mandelbrot already identified an erratic but regular pattern in cotton price.

In finance, the fractal analysis of currency markets by Fisher et al. (1997)

Particular price changes cannot be forecasted but there are scale invari-

accidentally identified a consequence of activism in exchange rate poli-

ances. Through an hourly, daily, or monthly scale of time, the shapes of

cies. While the DM/USD exchange rate had a multi-fractal signature, the

price curves remain the same [Mandelbrot (1963)].1 The cause may be

JPY/USD had no multi-fractal properties the JPY/USD appeared much

that even if there are many bits of information in the market at any given

smoother and lacking in sharp local peak. The explanation being that,

time, the price changes for individual transactions may depend solely on

even if the DM/USD market was much larger than JPY/USD, G3 central

what the traders regard as the most important piece of information. In

bank interventions on the Japanese yen were twice the dollar value of the

addition, the individual bits of information may follow a specific distribu-

interventions on the Deutschmark.

tion with properties leading the price changes to move asymptotically in


Fractals provide a new perspective on time series analysis, especially on

a specific way [Fama (1963)].

autocorrelations revealing much more information than correlograms of


Sharing the concerns of a growing population of experts [Shojai et al.

classical econometrics. In the next sections we aim to evaluate the con-

(2010)], the fractal perspective highlights the inaccuracy and the dangers

sequences of the LIBOR frauds on interest rates.

of modern portfolio theory (MPT).

The LIBOR scandal: a fractal analysis perspective

Scale independence of fractal patterning

Background

By definition fractals are self similar: some properties do not change if

The LIBOR scandal could be the largest insider trader scandal of all time.

scales of length are multiplied by a common factor. Therefore, a time

The LIBOR is the rate at which an individual contributor panel bank

series y(t) is self similar if a properly rescaled process

aHy(t/a)

presents

could borrow funds, were it to do so by asking for and then accepting interbank offers in reasonable market size just prior to 11:00 London

an identical probability distribution.

time.3 An average of submissions made by a number of selected banks


To be precise, we have y(t) aHy(t/a). The symbol means that there
are similar probability distributions on both sides of the equation.

aHy(t/a)

constitutes the LIBOR, which is calculated for ten different currencies


and 15 borrowing periods ranging from overnight to one year.

is the time series rescaled on the x-axis by a factor a (tt/a) and on the
y-axis by a factor of aH (yaHy).

Financial institutions set their own rates according to the LIBOR in order
to price mortgages, credit, and derivatives products. More than $800 tril-

Determining that two processes are statistically identical may be difficult.

lion in securities and loans are linked to it.4

Initially, practitioners approximated the equality by only examining the


mean and the variance (the first and second moments) on both sides

In June 2012, the Financial Services Authority (FSA) imposed a financial

of the equation. The exponent H, called the self-similarity parameter or

penalty of 59.5 million on Barclays Bank Plc. An investigation revealed

Hurst exponent [Hurst (1951)], is commonly used as a measure of the

fraud and collusions by the banks involved in the rate submissions. Addi-

geometric (fractal) scaling in the time series. It can be viewed as an indi-

tionally to the manipulation of the euro and the Japanese yen LIBOR, the

cator of the roughness of the original time series: the larger the value of

FSA has identified that between January 2005 and May 2009, at least

H, the smoother the time series.

173 requests for U.S. dollar LIBOR submissions were made to Barclays

The generalized Hurst exponent denoted Hq is commonly used in fractal


analysis. Extending the analysis of H to higher moment can reveal multi-

fractal properties. In a multi-fractal system, a continuous spectrum of


exponents is required to describe its dynamics.

Applications of fractal analysis


114

Significant discoveries in medicine and finance illustrate the power of

2
3
4

In 1900, Louis Bachelier already proposed a statistical model with self similar properties:
the Brownian motion. Sadly, Bachelier (1900) went quite unnoticed at the time. Although the
Brownian does not have fat tails, it has fractal properties. Today the Brownian constitutes
a particular case (with an exponent of ) of the more general fractional Brownian motion
developed by Mandelbrot and van Ness (1968).
See Ihlen (2012) for a review of literature on medical applications.
http://www.bbalibor.com/bbalibor-explained/definitions
http://blogs.wsj.com/source/2012/07/16/libor-qa/

The Capco Institute Journal of Financial Transformation


Multi-fractal Identification of Sick Markets: The LIBOR Scandal Case

submitters (including 11 requests based on communications from traders

Until Jan 05

at other banks) [FSA (2012)].

Jan 05-Jun 09

After Jun 09

q=0.1

0.681

0.720

0.645

q=1

0.607

0.571

0.710

Methodology, results and discussion

q=1.5

0.578

0.409

0.728

A focus on the three-month U.S. dollar LIBOR form January 1986 to Sep-

q=2

0.552

0.285

0.718

tember 2012 highlights that interest rate behaviors have been fundamen-

q=2.5

0.528

0.202

0.694

tally impacted by frauds and financial crises. We compare three periods:

q=3

0.505

0.144

0.667

from the creation of the U.S. LIBOR (P1: starting in January 1986) to the

q=4

0.462

0.070

0.620

period of falsification reported by the FSA (P2: May 2005 June 2009) and

q=5

0.427

0.025

0.585

the last period assumed free of fraud (P3: June 2009 September 2012).
Table 1

Although the rates were manipulated, irregularities from one period to another are not obvious through a direct view of the curve (Figure 1).

0.8

As previously mentioned, the Hurst exponents of integrated time series

0.7

are linked to various autocorrelation functions. The list below synthesizes

0.6

different situations [Peng et al. (2000)].

0.5

Case A: H(q=2) is equal to 0.5; the series are white noises: the value

0.2

C(T)=0 for any lag T.


Case B: H(q=2) is slightly higher than 0.5 and H(q>2) are close to 0.5;

0.1

there are short-term correlations and the autocorrelation functions will

0.0

decay exponentially C(T)=exp(-T/T0).

Case C: 0.5<H(q=2)<1; there are persistent long-range power-law cor-

q=1

q=1.5

q=2

q=2.5

q=3

q=4

q=5

Figure 2

relations; C(T)~TY with Y=2-2H.

Jan 05-Jun 09
After Jun 09

0.3

the integrated series are random walk; the autocorrelation function

Until Jan 05

0.4

at one instant is completely uncorrelated with any previous values;

Case D: 0<H(q=2)<0.5; power-law anti-correlations exists; small values


have a higher probability of being followed by larger values; large val-

period to another, revealing specific correlations during the fraud period

ues are more likely to be followed by small values.

P2. To be specific, we focus on the bold values of Table 1. We recognize

Case E: H(q=2)>1: the correlations are not following a power law any-

Case B for period P1, meaning that, before the fraud period, the market

more.

used to react with few jerks and relatively quickly to new economic information. The fraud period P2 is dramatically different and contains Case

We apply the MDFA method to estimate the Hurst exponents. The clas-

D characteristics: there are anti-correlations that can be explained by

sical Hurst exponent (with q = 2) shows a significant variance6 from one

successive artificial readjustments by the submitters, which may be re-

lated to the scams. The last period P3 is typical of an unstable market


12

related with a Case C characteristic.

10

P1

P2
5

P3

US LIBOR
0

/1

01

1
/1

/1

01

/0

01

/0

/0

01

01

/0

01

/0

01

01

/0

/0

01

/0

/0

01

01

/0

01

01

/9

/9

01

/9

/9

01

01

01

/9

/9

01

/9

01

/9

01

/9

01

01

/9

/8

01

/8

01

/8

/8

01

01

01

Figure 1

Kantelhardt et al. (2002) proposed the multifractal detrended fluctuation analysis (MDFA) for
the identification of multi-fractal scaling properties. Ihlen (2012) extensively discusses its
application in medicine. The wavelets and partition functions (PF) are alternative methods
and they provide similar results. In finance, the sample sum formula and PF inspired from
thermodynamics are commonly used: Fisher et al. (1997) applied the PF to compare JPY/
USD and DM/USD. Wavelets and MDFA, however, are simpler to apply. In addition, the
MDFA is widely used in medicine and appeared more intuitive as well as closer to our
metaphor on the financial regulation to identify unhealthy markets.
The samples have more than 1,000 units, which is enough to obtain accurate results
[Katsev and LHeureux (2003)]. We controlled that the time series are scale invariant with a
linear regression (see Appendix) of the samples sizes and the parameter Fq of the MDFA
calculation as suggested by Ihlen (2012). Additional powerful tests can be applied [Clauset
et al. (2009)].

115

1.2

12

0.8

1.1

0.7

10

0.6

0.9

0.8

0.5

0.7

Until Jan 05

0.4

Jan 05-Jun 09

0.6
4

0.5
0.4

0.2

0.3

Local Hurst exponent

US LIBOR

Ht Trend

0.2

After Jun 09

0.3

0.1

1
/1

/1

01

/1

01

/0

/0

01

01

01

/0

01

/0

/0

01

/0

01

/0

01

/0

/0

01

01

/9

/0

01

01

/9

01

01

/9

/9

01

01

/9

/9

/9

01

01

/9

01

/9

/9

01

01

/8

01

01

/8

01

/8

/8

01

01

0.0

Figure 3

q=1

q=1.5

q=2

q=2.5

q=3

q=4

q=5

Figure 4

Extending the analysis to a multi-fractal perspective (we include all the

1.2

values of Table 1 in Figure 2 to illustrate the Hurst exponent curves for the

1.1

different moments7) discloses additional inconsistency in the scale be-

12

10

0.9

havior of interest rates: there are multi-fractal properties in each period,

0.8

however, while the patterns are quite similar for P1 and P3 (on a different

0.7

level) the multi-fractality of P2 seems very specific. It highlights again

0.6

the sickness of the second period. The high level of the curve for the

0.5

third period (P3) may reflect the growing incertitude in financial markets

0.4

US LIBOR

Ht Trend
0

1
/1

/1

01

/1

01

/0

/0

01

01

/0

/0

01

01

/0

01

01

/0

01

/0

/0

01

/0

01

/9

/0

01

01

/9

01

/9

01

/9

01

01

/9

01

/9

01

/9

/9

01

/9

/9

01

01

/8

/8

/8

/8

01

01

0.2

01

use of quantitative easing, for example) following the 2008 crisis.

Local Hurst exponent

01

0.3

01

related to bank insolvency risks and the economic instabilities (intensive

The calculation of local Hurst exponents (q=2) at each point [Ihlen (2012)]
Figure 5

reveals a positive trend for the last 30 years (Figure 3).8 The increasing
financial deregulations generating unstable markets could be an explanation of the long-term trend.

values of the time series. To control that our results are mainly related to

Validation of the results

correlation, we reshuffled9 the time series [Kantelhardt et al. (2002)]. After

The Hurst exponents reflect correlation alterations of interest rates. Multi-

reshuffling, H(q=2) becomes very close to 0.5 (Table 2) and the time series

fractality can be due to different long-range correlation of the small and

develop into white noises. These results in real life would reflect a market

large fluctuations but also to a broad probability density function for the

with a perfect equilibrium and perfect prices adjustments.


The multi-fractality of each period becomes similar (especially for P2

q=0.1

116

Until Jan 05

Jan 05-Jun 09

After Jun 09

0.532

0.609

0.931

q=1

0.506

0.573

0.694

q=1.5

0.493

0.544

0.588

q=2

0.481

0.506

0.505

q=2.5

0.469

0.458

0.442

q=3

0.458

0.407

0.393

q=4

0.437

0.316

0.323

q=5

0.419

0.251

0.276

Table 2

and P3: Figure 4),10 and the rising trend of local exponents disappears11
(Figure 5).

7
8

We chose q=1;1.5;2;2.5;3;4;5 following Fisher et al. (1997).


To deal with local fluctuation close to zero, we use an alternative methodology to the one
of removing data in the time series. We replaced RMS<0.0001 with the average of the
previous values.
9 We randomly changed the order of the data points of the time series.
10 In Figure 4 we redefined three periods using lengths equal to the periods in Figure 1 and
Figure 2.
11 Figure 5 shows the original time series of the LIBOR, similarly to Figure 3. The local
exponents in Figure 5, however, are calculated from the reshuffled time series.

The Capco Institute Journal of Financial Transformation


Multi-fractal Identification of Sick Markets: The LIBOR Scandal Case

Conclusion

Appendix

Contrary to Rypdal and Lovsletten (2011), we showed that interest rates

We confirm the scale invariance of the series by checking that the re-

present direct fractal features which can be directly collected; the time

lationship is linear between the samples sizes and the parameter Fq of

series do not need to be fundamentally converted through a model.

the MDFA calculation; it should not be curved or S-shaped [Ihlen (2012)].

Multi-fractals could be of interest for financial regulators in order to di-

Figures 6, 7, and 8 illustrate periods P1, P2, and P3.

agnose market abnormalities.12 However, people can argue that, using


the LIBOR scandal as an illustration, it is easy to find additional evidence
with hindsight. Therefore, the approach could be developed in multiple
ways, for instance, by sliding the exponents through time (on thousand
units path basis) to disclose significant frauds.

References

Bachelier, L., 1900, Thorie de la speculation, Annales Scientifiques de lcole Normale


Suprieure, 3(17): 2186

Clauset A., C. R. Shalizi, and M. E. J. Newman, 2009, Power-law distributions in empirical


data, SIAM Review, 51: 661703
Fama E. F., 1963, Mandelbrot and the Stable Paretian Hypothesis, The Journal of Business,
36: 420429 (available at: http://www.jstor.org/stable/2350971)
Fisher, A. J., L. E. Calvet, and B. B. Mandelbrot, 1997, Multi-fractality of Deutschemark / US
Dollar Exchange Rates, Working Paper, Cowles Foundation Discussion Paper No. 1166, Sauder
School of Business (available at SSRN: http://ssrn.com/abstract=78628)
Financial Services Authority [FSA], 2012, Final notice: to Barclays Bank Plc, from the Financial
Services Authority (available at: www.fsa.gov.uk/static/pubs/final/barclays-jun12.pdf)
Hurst, H. E., 1951, Long-term storage capacity of reservoirs, Transactions of the American
Society of Civil Engineers, 116: 770808
Ihlen, E. A., 2012, Introduction to multi-fractal detrended fluctuation analysis in Matlab,
Frontiers in Fractal Physiology, 3:141 (doi: 10.3389/fphys.2012.00141)
Kantelhardt, J. W., S. A. Zschiegner, E. Koscielny-Bunde, S. Havlin, A. Bundeand, and H.
E. Stanley, 2002, Multi-fractal detrended uctuation analysis of nonstationary time series,
Physica A, 316: 87114
Katsev, S. and I. LHeureux, 2003, Are Hurst exponents estimated from short or irregular time
series meaningful?, Computers & Geosciences, 29: 1085 1089
Mandelbrot, B. B., 1963, The Variation of Certain Speculative Prices, The Journal of Business,
36(4): 394419
Mandelbrot, B. B., 1997, Fractals and scaling in Finance. Discontinuity, concentration, risk, New
York: Springer
Mandelbrot, B. B., 2005, The inescapable need for fractal tools in finance, Annals of Finance,
1: 193195
Mandelbrot, B. B. and J. W. van Ness, 1968, Fractional Brownian motions, fractional noises
and applications, SIAM Review, 10(4): 422437
Peng C. K., J. M Hausdorff, and A. L. Goldberger, 2000, Fractal mechanisms in neural control:
Human heartbeat and gait dynamics in health and disease, in Walleczek, J. (ed.), SelfOrganized Biological Dynamics and Nonlinear Control Cambridge: Cambridge University Press
Rypdal M. and O. Lovsletten, 2011, Multi-fractal modeling of short-term interest rates, Working
Paper 1.5265v1 [q-fin.ST], University of Troms (available at: http://arxiv.org/pdf/1111.5265.pdf)
Shojai, S., G. Feiger, and R. Kumar, 2010, Economists Hubris The Case of Equity Asset
Management, Journal of Financial Transformation, 29: 916 (available at: http://ssrn.com/
abstract=1597685)

Figure 6

Figure 7

Figure 8

12 Multi-fractal studies usually include spectrum graphs. They did not provide, however,
additional insight on the LIBOR question. As a result our discussion was only around the
generalized Hurst exponents.

117

Individual Selected Article from Journal 36

Guest Editor
Prof. Damiano Brigo, Head of the Mathematical Finance Research Group, Imperial College,
London

Advisory Editors
Cornel Bender, Partner, Capco
Christopher Hamilton, Partner, Capco
Nick Jackson, Partner, Capco

Editorial Board
Franklin Allen, Nippon Life Professor of Finance, The Wharton School,
University of Pennsylvania
Joe Anastasio, Partner, Capco
Philippe dArvisenet, Group Chief Economist, BNP Paribas
Rudi Bogni, former Chief Executive Officer, UBS Private Banking
Bruno Bonati, Strategic Consultant, Bruno Bonati Consulting
David Clark, NED on the board of financial institutions and a former senior
advisor to the FSA
Gry Daeninck, former CEO, Robeco
Stephen C. Daffron, Global Head, Operations, Institutional Trading & Investment
Banking, Morgan Stanley
Douglas W. Diamond, Merton H. Miller Distinguished Service Professor of Finance,
Graduate School of Business, University of Chicago
Elroy Dimson, BGI Professor of Investment Management, London Business School
Nicholas Economides, Professor of Economics, Leonard N. Stern School of
Business, New York University
Michael Enthoven, Former Chief Executive Officer, NIBC Bank N.V.
Jos Luis Escriv, Group Chief Economist, Grupo BBVA
George Feiger, Executive Vice President and Head of Wealth Management,
Zions Bancorporation
Gregorio de Felice, Group Chief Economist, Banca Intesa
Hans Geiger, Professor of Banking, Swiss Banking Institute, University of Zurich
Peter Gomber, Full Professor, Chair of e-Finance, Goethe University Frankfurt
Wilfried Hauck, Chief Executive Officer, Allianz Dresdner Asset Management
International GmbH
Michael D. Hayford, Corporate Executive Vice President, Chief Financial Officer, FIS
Pierre Hillion, de Picciotto Chaired Professor of Alternative Investments and
Shell Professor of Finance, INSEAD
Thomas Kloet, Chief Executive Officer, TMX Group Inc.
Mitchel Lenson, former Group Head of IT and Operations, Deutsche Bank Group
Donald A. Marchand, Professor of Strategy and Information Management,
IMD and Chairman and President of enterpriseIQ
Colin Mayer, Peter Moores Dean, Sad Business School, Oxford University
John Owen, Chief Operating Officer, Matrix Group
Steve Perry, Executive Vice President, Visa Europe
Derek Sach, Managing Director, Specialized Lending Services, The Royal Bank
of Scotland
ManMohan S. Sodhi, Professor in Operations & Supply Chain Management,
Cass Business School, City University London
John Taysom, Founder & Joint CEO, The Reuters Greenhouse Fund
Graham Vickery, Head of Information Economy Unit, OECD

Layout, production and coordination: Cypres Daniel Brandt, Kris Van de Vijver and
Pieter Vereertbrugghen
Graphic design: Buro Proper Bob Goor
Photographs: Buro Proper - Bob Goor
2013 The Capital Markets Company, N.V.
All rights reserved. This journal may not be duplicated in any way without the express
written consent of the publisher except in the form of brief excerpts or quotations
for review purposes. Making copies of this journal or any portion there of for any
purpose other than your own is a violation of copyright law.

Journal
The Capco Institute Journal of Financial Transformation

Amsterdam
Antwerp
Bangalore
Bratislava
Chicago
Dsseldorf
Frankfurt
Geneva
Johannesburg
London
New York
Orlando
Paris
San Francisco
Toronto
Washington, D.C.
Zurich
CAPCO.COM

You might also like