and artificial financial markets minimal spanning tree ... · 0.35 0.55 0.75 0.95 1.15 1.35 d<...

05/09/03 Midterm Cosin Conference, Rome

Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly

Minimal spanning tree networks in real and artificial markets

Minimal spanning tree networks in realand artificial financial markets

05/09/03 Midterm Cosin Conference, Rome 2

Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Overview

• Hierarchical structures can be detected in financial portfolios;

• One filtering procedure is a correlation-based clustering procedure;

• I discuss examples observed in the equity market and for financial markets located around the world;

• Artificial market models;

• Topology of MSTs in empirical data and market models;

• Conclusion.


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


The Collaboration

• Rosario N. Mantegna

• Fabrizio Lillo

• Giovanni Bonanno

• Guido Caldarelli

• Salvatore Miccichè


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Hierarchical treeA meaningful taxonomy can be extracted from returns time series by using a correlation-based clustering procedure†:

A hierarchical tree is obtained

†R.N. Mantegna, Eur. Phys. J. B 11, 193-197 (1999)

0.35

0.55

0.75

0.95

1.15

1.35

d<

i,j


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Minimal spanning tree

Together with a minimal spanning tree

This is the set of the 100 most capitalized US stocks in 1998.


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Cross-correlation

Cross-correlation between stock returns are well-known

2222jjii

jijiij

SSSS

SSSS

−−

−=ρ

)()1(ln)(ln tRtYtYS iiii ≅−−≡

They may be quantified by the correlationcoefficient ρij


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


A metric similarity measure

10.00 20.00 30.00 40.00 50.00 60.00 70.00 80.00 90.00 100.00

10.00

20.00

30.00

40.00

50.00

60.00

70.00

80.00

90.00

100.00

0.60

0.70

0.80

0.90

1.00

1.10

1.20

1.30

1.40

The function

( )ijijd ρ−= 12

verifies the axioms of a metric distance.


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Filtering procedure

The use of the single linkage clustering method using the distance dij as a similarity measure is one possible filtering procedure selectinginformation about the hierarchical structure of the portfolio


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


The ultrametric distance matrix

10 20 30 40 50 60 70 80 90 100

10

20

30

40

50

60

70

80

90

100

0.600.700.800.901.001.101.201.301.40

This filtering israther severe byretaining only(n-1) correlationcoefficients fromthe original n (n-1)/2


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


The (ordered) distance matrix

10 20 30 40 50 60 70 80 90 100

10

20

30

40

50

60

70

80

90

100

0.600.700.800.901.001.101.201.301.40

10 20 30 40 50 60 70 80 90 100

10

20

30

40

50

60

70

80

90

100

0.600.700.800.901.001.101.201.301.40

A direct comparison shows the amount of informationnot considered after the filtering procedure


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


The global financial market

1 G. Bonanno, N. Vandewalle and R. N. Mantegna, Physical Review E62, R7615 (2000)

We investigate51 MSCI stockindices both inlocal currency and in US dollars


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly



A hierarchical tree showing regionalrelations among stock indices of major stock exchanges has be detected1

1 G. Bonanno, N. Vandewalle and R. N. Mantegna, Physical Review E62, R7615 (2000)

1 US, Canada and Europe; 2 South-America;3 Japan; and 4 Asia


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly



1996

MSTs show a clear regional clustering


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


VolatilityThe same approachcan of course be applied also at volatilitytime series1

1 Salvatore Miccichè Giovanni Bonanno, Fabrizio Lillo, Rosario N. Mantegna, Physica A 324, 66-73 (2003)

The set investigated isthe set of 93 highly capitalized stocks tradedat NYSE during thetime period 87-98


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Associated volatility MST


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


High frequency data† (stocks)† B

onan

no, L

illo,

Man

tegn

a, Q

uant

itativ

e Fi

nanc

e 1,

96-

104

(200

1)

∆t=d/2 ∆t=d

∆t=d/5∆t=d/20


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


MSTs at different time horizons†

†Bonanno, Lillo, Mantegna, Quantitative Finance 1, 96-104 (2001)

∆t=19 min 30 sec ∆t=6 h 30 min


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Topology of MSTs

Topology† of MSTs in

• empirical data;• in an (unrealistic) random marketand• in a (realistic) one-factor model

†Bonanno, Caldarelli, Lillo and Mantegna, cond-mat/0211546, PRE (in press 2003)


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Empirical data

The color code refers tothe SIC index:• finance • manufacturing• construction • utilities• wholesale trade • mining• retail trade • services• public administration

1071 stocks continuouslytraded at the NYSE during the period 1987-1998 (3030 tradingdays)


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Empirical data: degree distribution

It is approximatelypower-law decayingfor low k values and a few nodes with very high degreeare present.


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


A random (unrealistic) model

( ) ( )ttR ii ε=

εi(t) is a Gaussian zero-mean i.i.d. random variable


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


MST of a random artificial market


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Topological properties of the random model

Degree distribution


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Theoretical prediction of the random model

Theory developedin M.D.Penrose, The Annals of Probability 24,1903 (1996)


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Eigenvalue spectrum

It may be worth mentioning that the random modeldescribes a significant part of the spectrum ofcorrelation matrix eigenvalues

From Laloux et al, PRL 83, 1467 (1999)


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


One-factor model We compare our empirical results with numerical simulations based on the one-factor model

where αi and βi are two real parameters, εi is a zero-mean noise term characterized by a variance equal to σi

2

RM(t) is the market factor. We choose it as the SP 500 index.

In our investigation, we estimate the model parameters and generate an artificial market.

( ) ( ) ( )ttRtR iMiii εβα ++=


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Correlation in the one-factor model

The one-factor model explains more than 85% of the elements of correlation matrix


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


MST of a one-factor model


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Topological properties of the one-factor model

Degree distribution


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Topological properties ofMSTs are different in the degree pdf f(k) and areenough to falsifya widespread financial model


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Topological properties are able to falsify the model

Other topologicalquantities, as forexample, thein-degreecomponentf(a) allow us toconclude thesame way


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Can we use our results for predictive purposes?

• Are empirical results suggesting that one-factor model better describes high-frequency returns rather than daily returns?

SOETR

AEP

BNI

NSC

AIT

UCM

GTE

BELBHI HAL

SLB

USB

JPM

WFCONE

BACMER

FLR

AXP

JNJ

MRK

BMY

CL

PG

MOB

ARCHM

CHVPEP

KO

GM

F

AIG

CI

HWP

IBM

UIS

BCC CHA

IP

WY

NSM TXN

INTC

ORCL

CSCO

SUNW

MSFT

GE

XON

∆t19 min30 sec

A preliminaryanswer is

YES!


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


Conclusion

•The topology of MSTs obtained by filtering the information present in the time dynamics of returns of an asset portfolio traded in a financial market allows to falsify simple models of markets dynamics.

•Topological properties of MSTs obtained at very short time horizons suggest that one-factor model is more appropriate for intraday time horizons rather than for daily time horizons of stock prices, weekly index returns of stock exchanges and volatility .


Obs

erva

tory

of

Complex

Sys

tems,

Paler

mo,

Ita

ly


The OCS website

http://lagash.dft.unipa.it

and artificial financial markets minimal spanning tree ... · 0.35 0.55 0.75 0.95 1.15 1.35 d<...

Documents