See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221991812

Is Art Market Behaviour Predictable?

Article in European Economic Review · April 1995

DOI: 10.1016/0014-2921(94)00058-8

CITATIONS READS

73 170

1 author:

Olivier Chanel
Aix-Marseille Université
92 PUBLICATIONS 2,302 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Taking citizens' views into account in public decisions View project

APHEKOM View project

All content following this page was uploaded by Olivier Chanel on 28 September 2018.

The user has requested enhancement of the downloaded file.

 EUROPEAN
M!%mc
ELSEVIER European Economic Review 39 (1995) 519-527

Is art market behaviour predictable?

Olivier Chanel
G.R.E.Q.A.M., 2, rue de la Charit;, F-13002 Marseille, France

Abstract

In this paper we look for relationships between art and financial markets through
econometric methods. The main results indicate that financial markets influence the art
market, with a lag of about one year, which seems plausible due to the relative sizes of both
markets. The Vector Auto Regressive (V.A.R.) model shows that lagged financial variables
help predicting art prices, even if the lag does not allow for systematic profits.

Keywords: Art market; Financial markets; V.A.R. model; Causality

JEL classification: C53, Zl

1. Introduction

It may seem reducing, at least for a non-economist, to use econometrics in

dealing with the art market. Indeed, econometrics is an objective science interested
in measurable data. Nevertheless, one may consider that a work of art has a
subjective value and a price, resulting from a general agreement on its value.
During the last twenty years, this distinction seems to have progressively
disappeared due to the rising presence of the financial sphere and speculators.
Return and risk, up to then unfamiliar to traditional collectors, became decisive

* I am indebted to Majorie Gassner, Victor Ginsburgh and Marcos De Bustamante Monteiro for
very helpful comments in Brussels. Financial support from the Belgian Government under contract PAI
no. 26 is gratefully acknowledged.

0014-2921/95/$09.50 0 1995 Elsevier Science B.V. All rights reserved

SSDI 0014..2921(94)00058-l
520 0. Chanel/European Economic Review 39 (1995) 519-527

criteria motivating most ’ of art purchases, which as a consequence, may be

viewed as a speculative asset.
In Section 2, we explore the financial aspects of investment in art, emphasizing
similarities and differences with financial assets. Then, in Section 3 we focus on a
particular segment of investment in art - contemporary and modern paintings -
and explain how to build an index for such heterogeneous ‘economic goods’. In
Section 4, we examine the causality links between this index and those of four
major stock markets (London, New York, Paris and Tokyo) through econometric
methods. In Section 5, we use these links in order to improve art market
predictability.

2. Financial aspects of art investment

It is well known that investment in art is strongly influenced by income and

other economic factors (see Frey and Pommerehne, 1989; Goetzmann, 1993), and
it could be interesting to study the links between art and stock exchange
movements, which certainly capture quite well the forces at play in the economy.
If we are only interested in the financial value of a work of art, considering that
its aesthetic or its social return cannot be expressed in monetary terms, the two
relevant choice criteria between assets are - besides returns - liquidity and risk.
Art works are not very liquid assets and anyway are not divisible; transaction
costs are high and there are unavoidable delays between the decision to sell and
the sale.
Risk is due to specific factors, like theft, fire, or possible reattribution of the
work to another artist. Moreover, ‘values’ are determined by a complex and
subjective set of beliefs on present and future prices, individual tastes or fashion
effects.
On the contrary, financial assets appear as very liquid and allow for diversifica-
tion thus reducing risk. Moreover, they are relatively transparent: fundamentals do
exist and can be analyzed. However, some analyses focus on art as a possible asset
in a Capital-Asset-Pricing-Model; they conclude that art possesses a low system-
atic risk. Indeed, Chanel (1993) and Pesando (1993) find p coefficients ranging
from 0.03 to 0.43, and a non-significant risk-adjusted excess-return.
In order to establish whether arbitrage (and therefore a source of systematic
profits) between these two assets is possible, we explore the dynamics of the
indices.

’ In a 1953 survey, 63% of art buyers admit that speculation is their most important criterion (see
Hoog and Hoog, 1991).
 0. Chattel/European Economic Review 39 (1995) 519-527 521

3. Method and data

First, we have to construct an index representing the behaviour of the art

market.
Prices of the art market are constructed using hedonic regression instead of
Repeat Sales Regressions (R.S.R.) which can hardly be used to construct quarterly
series of prices.
A hedonic equation can be written as:

lnP,,=f(X,,,,..., -Gk,,..., X,,,) +g(t) +ek,. (1)

e~N(0,&@Zr) and k=l,..., K;m=l,..., M;t=l,..., T.

where Pkr represents the price of good k at time t, Xmk, is a measurable

characteristic of good k at time t, g(t) is a function of time.
The representation finally adopted, after experimentations is the following:

E - N(0, & 63IT) (2)

m=l t=1

where St is a dummy variable which takes the value 1 for a sale occurring in
period t, and 0 otherwise. The coefficients /3, lead to an index of prices.
We apply this method using a database ’ containing 25.300 transactions of
paintings (excluding drawings, collage,. . . ) concerning 82 well-known artists. The
data sources - which determine the number and type of available characteristics -
are the Mayer compendia (1963-1993).
The characteristics include width, and height, their squares, the surface of the
painting, the place of sale (18 places) and a painter dummy. The exp( p,) terms
lead to a quarterly art market index, covering 1961:4 to 1992:4. See also Chanel et
al. (1992, 1994).
The four financial indices used are the Standard and Poors (New York), the
F.T. Actuaries (London), the Nikkei (Tokyo) and the I.N.S.E.E. (Paris), corre-
sponding to countries where 81% of the sales of our database took place. They all
include dividends. We chose to deflate each of them by the consumer price index
of the corresponding country, and the art index by the arithmetic mean of these
CPIS.
The indices, expressed in logarithms, are reproduced in Figs. 1 and 2. One
clearly sees the impact of the 1974 oil-crash (somewhat later for the art index), the
growth of prices until 1990, followed by the crash after 1990 both for the Nikkei
and art prices.

’ I am very grateful to Sophie Docclo for allowing me to use her database

 0. Chattel/European Economic Review 39 (1995) 519-527

‘t+‘+‘-
3.5

,961 1 ,963:l 1965.1 19671 19691 1971 1 19731 1975.1 19771 19791 ,981 1 1963:l 1965.1 1967.1 ,969, ,991 1 ,993,

Fig. 1. Logarithms of deflated indices.

7,5

I I
~ Nlkkel (Tokyo)

7 L

I-
6,5

1961.1 1963.1 1965:1 19673 1969.1 1971.1 19731 1975.1 19771 19791 1961 1 1963’1 19651 19871 19891 1991 119931

Fig. 2. Logarithms of deflated indices.

 0. Chanel/ European Economic Review 39 (I 995) 519-527 523

4. Relations between art and finance: Empirical investigations

We first checked for stationarity of the series used. Unit root tests indicate that
each series is first-order integrated (I(1)) - without quadratic and linear trend or
constant - so that first-order differences are stationary. Note that accepting the
random walk hypothesis is a sign of weak efficiency of the markets.
The study of short term relations is done through causality tests and measures.
Granger-causality and Geweke-Meese-Dent-causality (GC and GMD for short) are
first computed between the five variables. Intuitively, a variable Y ‘causes’ a
variable X if Y helps improving predictions of X. Formally, we use the following
tests for each pair of variables:

Granger-Causality : AX,=a+ t/3;AX,_i+ ty,Ay,_,+~,,

i=l i= 1
E, - I.I.N(O, (r’), (3)

GMD-Causality: AY,=a+ 5 Y~AX,_~+ &;Ay,_i+~,,

i= -p i= 1
E, N I.I.N(O, a2), (4)
where q are lags, p are leads, X, is the caused and Y, the causal variable.
We conclude that ‘Y, causes X,’ if the yi jointly differ from 0 (using an
F-test). Tests have been conducted with p = 4 leads and q = 12 lags (GC w
F(12,87) and GMD N F(4,78)), and are summarized in Table 1.

Table 1
Causality tests
Causal variables Caused variables

Art New York Tokyo Paris London

Art

New York GC 1.26 1.68 * 1.78 * *

GMD 3.17 * * 2.79 ** 3.94 * *

Tokyo GC 2.18 * * 1.60 *

GMD 2.85 * * 0.47

Paris GC 1.67 * - 1.59 *

GMD 0.81 1.55

London GC 1.13 0.87 -

GMD 2.19 * ’ 1.97 * -
* = significant at the 10% level.
**
= significant at the 5% level.
524 0. Chard/European Economic Review 39 (1995) 519-527

Table 2
Causality measures

Lags Art + New York New York + Art Art + Tokyo Tokyo + Art

0 12.36 * *
1 4.07 * * 2.92 * 3.45 *
2 3.90 * * 5.33 * *
3 9.34 * * 5.63 * *

5 4.29 * *
6 4.25 * *
7 3.83 * *

1 to 12 16.81 22.03 * * 7.5 1 29.64 *’

MY Art + Paris Paris + Art Art + London London + Art

2 2.82 * 6.86 ’ *
3 2.98 *

6 5.01 * *
7 3.20 *

11 2.76 *
12 5.03 * *

1 to 12 12.45 10.40 14.99 16.90

1
= significant at the 10% level.
*I
= significant at the 5% level.

We note that art never ‘causes’ the stock markets, whereas English, Japanese
and American stocks significantly cause art. Furthermore, the US market is never
caused, but causes the Japanese and the English stock markets.
We next compute measures of causality for each of the four financial-art
relations in order to determine the relevant lags. The methodology adopted
consists in testing the significance level of each yi in relation (31, with a
Chi-Square (1) test (see Gourieroux and Monfort, 1990). The main results, given
in Table 2, confirm previous findings. Moreover, we see that most of the
significant lags are between 1 and 4 quarters, with little feedback effects from art
to financial markets.
We conclude that short-run relations exist between the two spheres, and we turn
to a dynamic model in order to exploit these links, using Vector Auto Regressive
(V.A.R.) models.
 0. Chanel/European Economic Review 39 (1995) 519-527 525

5. Predicting art prices with V.A.R. models

In a critical analysis of traditional macroeconomic modelling, Sims (1980)

envisages a new approach known as V.A.R. He considers that the structural form
of a model can hardly be obtained by theory or intuition, and thinks that the data
should reveal the causal structure and exogeneity restrictions. This kind of model
consists in estimating coefficients of a model in which each variable is being
explained by its own lagged values and those of other variables.
We first use a Johansen Maximum Likelihood method to detect possible
long-term steady state relationships between the indices. If such relations exist, we
must explicitly take them into account. Otherwise (i.e., if all the linear combina-
tions of X, are I(1)) a V.A.R. can be estimated on stationary variables, and will be
stable.
Using a V.A.R. representation expressed in first differences and lagged levels,
we have

AX,= ~T,dX,,+a/YX,_, +E,, E,-N(O,d), (5)

k=l

where X, is the vector of our five indices.

In a first step, we search for the minimal number of lags ensuring normality of
the residuals of the five estimated equations. The value of k = 9 lags is chosen
using Jarque-Bera and Ljung-Box 3 tests. In a second step, we test for the number
of cointegration relations using the A-max and trace tests (see Johansen, 1991;
Osterwald-Lenum, 1992; Reinsel and Ahn, 1988). Both accept the null hypothesis
that no long-term relation exists between the art and financial markets. 4
We therefore estimate a classical V.A.R. model (in fact a particular case of (5)
with no cointegration relation):

AX,= &kAX,_k+~,, E,-N(O,a*). (6)

k=l

We first check for K, the optimal number of lags (between 1 and 121, using a
step by step Likelihood Ratio test. The restriction from 12 to 11 lags is rejected
with a high probability level, so K will be set at 12 (3 years). LR tests are also
used to check for exogeneity, which leads us to keep all five variables in the
V.A.R. representation.
An unrestricted V.A.R. involving all the variables and their lags is of little
interest, due to the large number of (often non-significant) coefficients. We then
choose to restrict the representation by using a method introduced by Hsiao
(1979): the ‘bottom-up top-down’ strategy, which consists in finding the optimal

3 Normality for the Japanese market could not be found, even with 12 lags.
4 Ginsburgh and Jeanfils (1995) obtain similar results.
526 0. Chattel/European Economic Review 39 (1995) 519-527

ALn(Art)
0.6

0,4 -c

-0.6

Fig. 3. V.A.R. model.

lag for each variable in each equation, and removing the non significant lags once
the model is completely specified.
Rather than reproducing the results for the various models obtained, we choose
to focus on the capacity of two particular models to predict the 1990 downturn in
the art market. The first model used is a V.A.R. (see Fig. 3) while the second one
is an autoregression of the art index on its twelve first lags. Table 3 clearly shows
that the V.A.R. model is superior (smaller Root Mean Square Error (RMSE),
larger adjusted correlation coefficient (i*) and better predictions). Moreover, it
allows for ex-ante predictions thanks to its dynamic structure which includes
financial indices.

Table 3
Models’ performance

RMSE i?2 Post-sample predictions

1990:2 1990:3 1990:4 1991:l
‘True’ values - 0.248 0.073 - 0.434 -0.174
V.A.R. 0.116 0.601 -0.212 0.109 0.146 -0.165
Autoregression 0.170 0.170 0.053 0.142 0.099 0.051
 O. Chanel/ European Economie Reuiew 39 (1995) 519-527 527

6. Conclusion

It seems likely that there exist relations between art and financial markets:
graphical analysis, causality tests and measures, and the V.AR. representation
lead to the same conclusion. Moreover, we note that, roughly, these markets move
together, in the sense that no systematic transfers exist. One should also keep in
mind that financial markets are highly correlated with economies, and that their
volume of transactions is much larger than the art market's.
It would appear, then, that financial markets react quickly to economic shocks,
and that the profits generated on these markets may be invested in art, so that
stock exchanges may be considered as advanced indicators to predict what
happens on the art market. However, art is subject to fashions, tastes and fads and
this makes long-term forecasts quite difficult.

References

Chanel, O., 1993, Apports de l'économétrie à l'étude des champs culturels: Applications au marché des
oeuvres d'art et à la demande télévisuelle, Ph.D. Dissertation (E.H.E.S.S., Paris).
Chanel, O., L.-A. Gérard-Varet and V. Ginsburgh, 1992, The relevance of hedonic price indices: The
case of paintings, Working paper 92-Al9 (G.R.E.Q.E., Marseille).
Chanel, O., L.-A. Gérard-Varet and V. Ginsburgh, 1994, Prices and returns on paintings: An exercise
on how to price the priceless, Geneva Papers on Risk and Insurance Theory 19, 7-21.
Frey, B.S. and W.W. Pommerehne, 1989, Muses and markets (Basil Blackwell, London).
Ginsburgh, V. and P. Jeanfils, 1995, Long-term comovements in international markets for paintings,
European Economie Review 39, this issue.
Goetzmann, W.N., 1993, Accounting for taste: Art and the financial markets over three centuries,
American Economie Review 83, no. 5, 1370-1376.
Gourieroux, C. and A. Monfort, 1990, Cours de séries temporelles et modèles dynamiques (Economica,
Paris).
Hoog, M. and H. Hoog, 1991, Le marché de l'art, Que sais-je?, (Presses Universitaires de France,
Paris).
Hsiao, C., 1979, Autoregressive modeling of Canadian money and income data, Journal of the
American Statistical Association 74, 553-560.
Johansen, S., 1991, Estimation and hypothesis testing of cointegration vectors in Gaussian vector
autoregressive models, Econometrica 59, 1151-1180.
Osterwald-Lenum, M., 1992, A note with quantiles of the asymptotic distribution of the maximum
likelihood cointegration rank test statistics, Oxford Bulletin of Economies and Statistics 54,
461-72.
Mayer, E., 1963-93, International auction record (Editions M., Zurich).
Pesando, J.E., 1993, Art as an investment: The market for modern prints, American Economie Review
83, no. 5, 1075-1089.
Reinsel, G.C. and S.K. Ahn, 1988, Asymptotic distribution of the likelihood ratio test for cointegration
in the nonstationary vector AR mode!, Technical report (Department of Statistics, University of
Winsconsin, Madison, WI).
Sims, C.A., 1980, Macroeconomics and reality, Econometrica 48, 1-48.
 Fig. 1. Logarithms of deflated indices

Logarithms
7

SP-500( New-York)
6.5 INSEE (Paris)
Art index

5.5

4.5

3.5
1961:31963:31965:31967:31969:31971:31973:31975:31977:31979:31981:31983:31985:31987:31989:31991:31993:3
 Fig. 2. Logarithms of deflated indices

7.5 Logarithms

7
Nikkeï (Tokyo)
F.T.-Actuaries (London)
Art index
6.5

5.5

4.5

4
1961:3 1963:3 1965:3 1967:3 1969:3 1971:3 1973:3 1975:3 1977:3 1979:3 1981:3 1983:3 1985:3 1987:3 1989:3 1991:3 1993:3

Page 1

View publication stats