A Theory of Large Fluctuations in Stock Market Activity∗

Xavier Gabaix† , Parameswaran Gopikrishnan‡ , Vasiliki Plerou‡ , H. Eugene Stanley‡

†
MIT, Economics Department and NBER
‡
Boston University, Physics Department, Center for Polymer Studies
August 16, 2003

Abstract
We propose a theory of large movements in stock market activity. Our theory is motivated
by growing empirical evidence on the power-law tailed nature of distributions that characterize
large movements of distinct variables describing stock market activity such as returns, volumes,
number of trades, and order flow. Remarkably, the exponents that characterize these power
laws are similar for different countries, for different types and sizes of markets, and for different
market trends, suggesting that a generic theoretical basis may underlie these regularities. Our
theory provides a unified way to understand the power-law tailed distributions of these variables,
their apparently universal nature, and the precise values of exponents. It links large movements
in market activity to the power-law distribution of the size of large financial institutions. The
trades made by large financial institutions create large fluctuations in volume and returns. We
show that optimal trading by such large institutions generate power-law tailed distributions for
market variables with exponents that agree with those found in empirical data. Furthermore,
our model also makes a large number of testable out-of-sample predictions.

Contents
1 Introduction 4
1.1 The power-law distribution of returns, volume, number of trades . . . . . . . . . . . 4
1.2 Outline of the theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 The empirical facts that motivate our theory 7

2.1 Methodology to estimate power laws . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Normalizations used to make different assets comparable . . . . . . . . . . . . . . . . 7
2.3 The cubic law of price fluctuations: ζr ' 3 . . . . . . . . . . . . . . . . . . . . . . . . 8

∗ xgabaix@mit.edu, gopi@cgl.bu.edu, plerou@cgl.bu.edu, hes@miranda.bu.edu. We thank Tal Fishman, David

Kang, Alex Weisgerber, and especially Kirk Doran for outstanding research assistance. For helpful comments, we
thank Tobias Adrian, Jonathan Berk, Olivier Blanchard, Jean-Phillipe Bouchaud, John Campbell, Emanuel Derman,
Ken French, Joel Hasbrouck, Soeren Hvidjkaer, Harrison Hong, Ivana Komunjer, David Laibson, Augustin Landier,
Ananth Madhavan, Lasse Pedersen, Thomas Philippon, Marc Potters, Jon Reuter, Gideon Saar, Andrei Shleifer,
Didier Sornette, Dimitri Vayanos, Jessica Wachter, Jiang Wang, Jeff Wurgler, and seminar participants at Berkeley’s
Haas School, Delta, Harvard, John Hopkins University, MIT, NBER, NYU’s Stern School, Princeton, Stanford GSB,
the 2002 Econophysics conferences in Indonesia and Tokyo, and the 2003 WFA meetings. We thank the NSF for
support. X.G. thanks the Russell Sage Foundation and NYU Stern School for their wondeful hospitality during the
year 2002-2003.

1
 2.4 The half cubic law of volume: ζQ ' 3/2 . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 The cubic law of number of trades: ζN ' 3.3 . . . . . . . . . . . . . . . . . . . . . . 10
2.6 The power law distribution of the size of large investors: ζS ' 1 . . . . . . . . . . . . 10

3 The theory, assuming a power law price impact of trades 11

3.1 Sketch of the theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3.2 Notations and some elementary links between exponents . . . . . . . . . . . . . . . . 12
3.2.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.2 Link between aggregate and trader-based exponents . . . . . . . . . . . . . . 13
3.2.3 Link between the exponents of return and volume . . . . . . . . . . . . . . . 13
3.3 A main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 An optimizing model that illustrates Theorem 3 . . . . . . . . . . . . . . . . . . . . . 16

4 The square root of price impact of trades: Evidence and a possible explanation 19
4.1 Evidence on the square root law of price impact . . . . . . . . . . . . . . . . . . . . . 19
4.2 Sketch of our explanation for the square root price impact . . . . . . . . . . . . . . . 19
4.3 A microstructure model for the power law impact of a block . . . . . . . . . . . . . . 21
4.3.1 The behavior of liquidity providers . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3.2 Permanent vs Transitory price impact . . . . . . . . . . . . . . . . . . . . . . 21
4.3.3 The behavior of the large trader . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 The distribution of individual trades q vs distribution of target volumes Q . . . . . . 23
4.5 An optimizing intertemporal model that joins power laws and microstructure . . . . 24
4.6 Some comments on the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.6.1 Arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.6.2 Plausibility of very large volumes and price impact . . . . . . . . . . . . . . 26

5 Assessing some further empirical predictions of the model 26

5.1 Contemporaneous behavior of several measures of trading activity . . . . . . . . . . 26
5.1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.1.2 Comparison between theory and facts . . . . . . . . . . . . . . . . . . . . . . 27
5.2 Some untested predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

6 Related literature 28
6.1 Alternative theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
6.1.1 The public news based (efficient markets) model . . . . . . . . . . . . . . . . 30
6.1.2 A mechanical “price reaction to trades” model . . . . . . . . . . . . . . . . . 30
6.1.3 Random bilateral matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
6.2 Related empirical findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
6.2.1 Other models for returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
6.2.2 Buy / Sell asymmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.3 Link with the behavioral and excess volatility literature . . . . . . . . . . . . . . . . 31
6.4 Link with the microstructure literature . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.5 Link with the “econophysics” literature . . . . . . . . . . . . . . . . . . . . . . . . . 32

7 Conclusion 32

8 Appendix A: Some power law mathematics 34

9 Appendix B: Generalization of Theorem 3 to non i.i.d. settings 35

2
10 Appendix C: The model with general trading exponents 37

11 Appendix D: A model for the linear supply function (38) 39

12 Appendix E: Algorithm for the simulations 40

13 Appendix F: Confidence intervals and tests when a variable has infinite variance 41
13.1 Construction of the confidence
£ intervals
¤ for Figure 7 . . . . . . . . . . . . . . . . . . 41
13.2 Test of the linear relation E r2 | Q = α + βV . . . . . . . . . . . . . . . . . . . . . 41

3
1 Introduction

Even ex post, stock market fluctuations are very hard to explain by movement in fundamentals.
Trading per se seems to move prices. This leads to excess volatility of stock market prices. Even
crashes seem to happen for no good reason. Trading volume is very high, and its large variations
are also hard to relate to fundamentals. The present paper wishes to present a theory of those large
movements in trading activity1 .
For this is useful to have precise empirical facts. Otherwise, theories are too unconstrained. This
is why we use a series of sharp empirical facts, established using dozens of millions of data points.
They are quite precise, hence they constrain any theory of large movements of stock market activity.
We first outline some of those key empirical regularities before presenting our theory.

1.1 The power-law distribution of returns, volume, number of trades

Our theory is motivated by the following empirical findings on the power law distribution of (i)
returns, (ii) volumes, and (iii) the number of trades, (iv) the power law of price impact, and (v) the
power law distribution of the size of large investors.
(i) The power law distribution of returns. Let Pt denote the price of a stock or an index, and
define return over a time interval ∆t as rt = ln(Pt /Pt−∆t ). Empirical studies (Gopikrishnan et al.
(1999), Plerou et al. (1999)) show that the distribution function of returns for 1000 largest U.S.
stocks and several major international indices decays as
1
P (|r| > x) ∼ with ζr ' 3 (1)
xζr
Here, ∼ means asymptotic equality up to numerical constants2 . This holds for positive and negative
returns separately.
Figure 1 shows that P (|r| > x) on a bi-logarithmic scale displays a linear asymptotic behavior
with slope −ζr = −3.1. Since linear behavior on a log-log plot means a power-law decay, we conclude
that ln P (|r| > x) = −3.1 ln x + constant, i.e. (1). There is no tautology that implies that this
graph should be a straight line, or that the slope should be -3. A Gaussian would have a concave
parabola, not a straight line. The remarkable nature of the distribution (1) is that it holds for stock
indices and individual stocks different sizes, different time periods (see Section 2 for a systematic
exploration), and we always find ζr ' 3. In the following, we shall refer to Eq. (1) as “the cubic law
of returns”.
(ii) The power law distribution of trading volume. We call volume the number of shares traded3 .
We find that the distribution is:
1
P (q > x) ∼ with ζq ' 3/2. (2)
xζq
The finding holds both for the volume of individual trades and aggregate volumes:
1
P (Q > x) ∼ with ζQ ' 3/2. (3)
xζQ

1Acondensed version of some elements of the present paper appeared in Gabaix et al. (2003a).
2 Formally
f (x) ∼ g (x) means f (x) /g (x) tends to a positive constant (not necessarily 1) as x → ∞.
3 The dollar value traded yields very similar results. This is expected, as, for a given security, the variations of

number of shares traded and the volume traded are essentially proportional.

4
Figure 1: Empirical cumulative distribution of the absolute values of the normalized 15 minute
returns of the 1000 largest companies in the TAQ database for the 2-yr period 1994—1995. This
represents 12 million observations. In the region 2 < x < 80 we find an OLS fit ln P (r > x) =
−ζr ln x+b, with ζr = 3.1±0.1. This means that returns are distributed with a power law P (r > x) ∼
x−ζr for large x.

We had initially found this in U.S. data (Gopikrishnan et al. 2000), and we have recently
confirmed the same on French data (Plerou et al. 2003). Figure 2 illustrates this. It shows the
density satisfies p (q) ∼ q −2.5 , i.e. (2). The exponent of the distribution of individual trades is close
to 1.5. In the following, we refer to Eq. (2)-(3) as the “half-cubic law of trading volume”.
As before, the exponents describing these power-laws seem to be stable across different types of
stocks, different time periods and time horizons etc. (see Section 2).
(iii) The power law distribution of the number of trades. A similar analysis for the number of
trades shows that
1
P (N > x) ∼ ζN with ζN ' 3.3 (4)
x
(iii) The power law of price impact.
Building on the concave nature of the impact functions in stock markets (Hasbrouck 1991, Has-
brouck and Seppi 2001, Plerou et al. 2002), we will give evidence that the price impact ∆p of a
trade of size V scales as:
∆p ∼ V γ with γ ' 1/2. (5)
(v) Power law distribution of the assets of large investors. Our empirical analysis of the size
distribution of mutual funds shows evidence for a power law distribution. As Section 2.6 reports the
number of funds with size (asset under management) greater than x follows:
1
P (S > x) ∼ with ζS ' 1 (6)
xζS
We will present a theory where those five facts are intimately related. It is largely orthogonal to
the existing finance theory. Indeed, much of finance is about the risk premia commanded by various

5
Figure 2: Probability density of the individual trade sizes q for 30 large stocks in the Paris Bourse
from January 1995 to October 1999. This represents 35 million observations. OLS fit yields a density
∼ q −(1+ζq ) , for ζq = 1.5 ± 0.1. Source: Plerou et al. (2003).

shocks, but not about the origins of the shocks themselves. For instance, if rMt is a market return,
and εit is an idiosyncratic return, a typical finance model explains why an asset i will have a return
of the factor type:
rit = βi rMt + εit
what causes variations in the loadings βi , or even what the relevant factors rM should be. In contrast,
our goal here is not to explain risk premia, but rather to propose a structured theory for the origin
of the shocks rMt , εit . For this we are guided by the regularities in their distribution and in the way
they relate to other market quantities such as volume and volume “imbalance”. In other words, we
wish to theorize on the origins of those movements, especially to the extent that they cannot easily
be related to fundamentals.

1.2 Outline of the theory

In our theory, large volumes and large returns are created by the trades of large investors. To
see this qualitatively, suppose that a large investor decides to trade in a stock due to some news
announcement, or some recent analysis he made himself. For a sufficiently large investor, the desired
quantity of the particular stock will be comparable to the daily turnover4 . So if he executes the
trade quickly he will suffer a large price impact. However, the alternative of performing the trade
very slowly over time is not attractive either since the mispricing cannot be expected to remain
indefinitely, so he does not wish to wait too long before realizing the trade. We show that under
plausible conditions, his “optimal” trading behavior generates power law exponents for returns,

4 Please see Section 4.6.2 for empirical support of this assertion.

6
volumes and number of trades, i.e. the exponents ζQ = ζq = 3/2 for the volume, and ζr = ζN = 3
for the return and number of trades.
Section 2 presents the data and our empirical findings in more detail. The provide a self-contained
tour of the empirical literature on power laws. For modularity’s sake, we chose to present the model in
two steps. In Section 3, we present the model while assuming the microstructure part, i.e. assuming
a power law price impact. We show how his generate the above power laws5 . In Section 4, we present
one possible model for the square root price impact of trade. Section 5 gathers the out-of-sample
predictions of our model and our empirical results for some of these predictions. Section 6 discusses
our finding in the context of the relevant literatures. Section 7 presents concluding remarks and
directions of future work.

2 The empirical facts that motivate our theory

2.1 Methodology to estimate power laws

There are two basic methodologies to estimate power laws exponents. We illustrate them with
the examples of returns. In both methods, one selects a cutoff of returns, and orders the top n
observations as r(1) ≥ . . . ≥ r(n) , up to a certain cutoff.
The simplest method is to do a “log rank log size regression”, where an estimate of ζ is the the
OLS coefficient on r(i) in the regression of log of the rank i (or of cumulative frequency i/n) on the
log size:
ln i/n = A − ζbOLS ln r(i) + noise
This method is the simplest, and yields a visual goodness of fit with the power law. This what we
do for instance in Figure 1.
A second method is Hill’s estimator
n−1
ζbHill = Pn−1 ¡ ¢.
i=1 ln r(i) − ln r(n)

Those two methods have pitfalls. In particular, they underestimate the true standard errors. Some
of those pitfalls are discussed in Embrechts et al. (1997, 330—345) and Gabaix and Ioannides (2004).
Most of the available statistical literature assumes i.i.d. draws. In contrast, there is a clear
autocorrelation of volatility and trading activity in financial data. This does not bias our estimates
of the unconditional distributions, but makes the assessment of the standard errors more difficult.
Before the statistically rigorous theory of how to proceed is developed, our pragmatic solution
is to use a large amount of data — several million observations. The stability of the estimates
across different periods, countries and classes of assets give us anyway confidence that the empirical
estimates we report here are robust.

2.2 Normalizations used to make different assets comparable

To compare quantities across different stocks, we normalize variables such as r, V and N by the
second moments in they exist, otherwise by the first moments6 . For instance, for a stock i, we
consider the returns rbit = (rit − ri ) /σr,i , where ri is the mean of the rit and σr,i is their standard

5 Technical extension of the present paper are available in Gabaix et al. (2003b).
6 This is explained more systematically in Appendix F.

7
Figure 3: Zipf plot for the daily fluctuations in the Nikkei (1984-97), the Hang-Seng (1980-97),
and the S&P 500 (1962-96). The apparent power-law behavior in the tails is characterized by the
exponents ζr = 3.05 ± 0.16 (NIKKEI), ζr = 3.03 ± 0.16 (Hang-Seng), and ζr = 3.34 ± 0.12 (S&P
500). The fits are performed in the region |r| > 1. Source: Gopikrishnan et al. (1999).

deviation. For the volume, which has an infinite standard deviation, we consider the normalization:
b it = Qit /Qi , where Qi is the mean of the Qit .
Q

2.3 The cubic law of price fluctuations: ζr ' 3

The cubic distribution of returns has been uncovered in a series of studies (Lux 1996, Gopikrishnan
et al. 1999, Plerou et al. 1999).We show here the robustness of the power-law exponent ζr ' 3.
It seems to hold internationally. We analyze data for 10 other developed stock markets7 . We
find a mean of ζr = 2.9 with a standard error of 0.10. Figure 3, for instance, shows how the return
distribution of three different countries are very similar.
The cubic law predicts the existence of a few very large returns over a century. In Gabaix et al.
(2003c) we show that in the U.S., stock market crashes, including the 1929 and 1987 crashes, are not
outliers to the cubic law. Hence a theory of the cubic law of returns might be a theory of crashes.
Mandelbrot (1963) argued that financial return followed a Lévy stable distribution, and offered
the theory that this might come from a large number of shocks with infinite variance. Lévy distri-
bution imply that ζr ≤ 2. But our data are inconsistent with ζr ≤ 2, hence we are lead to reject
Mandelbrot’s “stable Paretian hypothesis” (Fama 1963).

Having checked the robustness of the ζ = 3 finding across different stock markets, we look at
firms of different sizes. Small firms have a higher volatility than big firms, as is verified in Figure

7 We took the stock market indices of Australia (1/1/73-11/27/01) (1/2/84—11/17/01), Canada (1/1/69—11/20/01),

France (1/2/84-11/27/01), Germany (1/1/73, 10/18/01), Japan (1/1/69-11/30/01), Hong-Kong, (1/1/73-11/30/01),

Netherlands (1/1/80-11/27/01), South Korea (1/7/75-11/27/01), Spain (1/2/74-11/27/01), and United Kingdom
(12/25/68-11/20/01). We calculated the Hill estimator for the extreme 1% of the distribution.

8
Figure 4: Cumulative distribution of the conditional probability P (r > x) of the daily returns of
companies in the CRSP database, 1994-5. We consider the starting values of market capitalization
K, and define uniformly spaced bins on a logarithmic scale and show the distribution of returns for
the bins, K ∈ (105 , 106 ], K ∈ (106 , 107 ], K ∈ (107 , 108 ], K ∈ (108 , 109 ]. (a) Unnormalized returns
(b) Returns normalized by the average volatility σK of each bin. The plots collapsed to an identical
distribution, with ζr = 2.70 ± .10 for the negative tail, and ζr = 2.96 ± .09 for the positive tail.
Source: Plerou et al. (1999).

4(a). But the same figure also shows a similar slope. Indeed, when we normalize the distribution by
a common standard deviation, we see that the plots collapse, and the exponents are very similar,
around ζ = 3 again.

Some studies have quantified the power law exponent of foreign exchange fluctuations. The most
comprehensive is probably Guillaume et al. (1997), who calculate the exponent ζr of the price
movements between the major currencies. At the shortest frequency ∆t = 10 minutes, they find
exponents with average hζr i = 3.44, and a standard deviation 0.30. This tantalizingly close to the
stock market findings, though the standard error is too high to draw sharp conclusions8 .

2.4 The half cubic law of volume: ζQ ' 3/2

Trying to understand the origins of the cubic law for returns, Gopikrishnan et al. (2000a) looked
at the distribution of volume9 in the U.S. and found an exponent around 1.5: ζQ = 1.53 ± .07.
The same value ζq ' 3/2 held for the distribution of individual trades. Maslov and Mills (2001)
find likewise ζq = 1.4 ± .1 for the volume of market orders. To test the universality of this finding,
we examine 30 large stocks of the Paris Bourse from 1995—2000, which contains approximately 35
million records. We reported the finding in Figure 2 and found ζq = 1.5 ± .1.

8 Suppose
¡ ¢2
returns are distributed according to a cubic law with a the density proportional 1/ 1 + r2 . Examining
the top 1% of 10,000 points, the standard deviation of measured exponents is about 0.4, for a mean of about 3.
9 For stock i, we look at the fluctuations of V /V , where V is the mean volume V . This makes stocks comparable,
it i i it
and the “number of shares” and “dollar volume” would give equivalent measures of volume.

9
2.5 The cubic law of number of trades: ζN ' 3.3
Plerou et al. (2001) find an exponent ζN = 3.4 ± .05 for the number of trades in a 15 min interval in
the U.S. In the Paris Bourse data mentioned above, we find ζN = 3.17 ± .1. Our theory will predict
the exponent of 3. This may due to a flaw in the theory10 , or to simply to measurement error11 .

Figure 5: Cumulative distribution of the normalized number of transactions n∆t = N∆t / hN∆t i . Each
symbol shows the cumulative distribution P (n∆t > x) of the normalized number of transactions
n∆t for all stocks in each bin of stocks sorted according to size. An analysis of the exponents
obtained by fits to the cumulative distributions of each of the 1000 stocks yields an average value
ζN = β = 3.40 ± 0.05.

2.6 The power law distribution of the size of large investors: ζS ' 1
The size distribution of many entities follows a power law, and often with exponent 1, something
called a Zipf’s law. This is true for cities (Zipf 1949, Gabaix and Ioannides 2003) and business
firms. Indeed Axtell (2001) and Okuyama et al. (1999) have shown that the distribution of firms
(respectively in the US and Japan) follows Zipf’s law. As they show that the distribution of firms
follows Zipf’s law, it is tempting to hypothesize that the distribution of firms that manage money
follows Zipf’s law. We found data for an important subset of those firms, mutual funds12 . From
CRSP one gets the size (dollar value of the assets under management) of all the mutual funds13

10 A scaling N ∼ V ξ gives a value ζN = ζV /ξ, so that a scaling of the number of trades N ∼ V 0.44 rather that the
theoretical N ∼ V 0.5 would correct this misfit. This would mean that the broker can do slightly better than predicted
by the model. For very large trade, the broker could have an extra incentive to limit the number of people to whom
the trade is offered, perhaps to limit information leakage.
1 1 Power law relations are notoriously hard to measure, especially for positive variables with high exponents with

positive quantities. The reason is that high exponents have fewer very large events, so the measurement can be
affected by other sources of noise. For instance, suppose that the true process for the number of trades N is N =
N0 + a, where N0 has power law 3 with P (N0 > x) = kx−3 , and a is a constant. The interpretation is that a
is a constant “background noise” number of trades, and the N0 represents the trades of large investors as in our
model. Then the CDF of N is G (x) = P (N > x) = k (x − a)−3 , and the measured power law exponent will be:
ζc 0
N = −xG (x) /G (x) = / (1 − a/x). It tends to 3 for very large values of x, but in finite samples, there is an upward
bias equal to a/N0 , where N0 is the order of magnitude of typical value of N0 in the sample.
1 2 Here we sketch the main findings. Gabaix, Ramalho and Reuter (2003) present many more details.
1 3 The say 200 funds of Fidelity, for instance, count as 200 different funds, not as one big “Fidelity” fund.

10
from 1961-1999. For each year t, we do an OLS estimation of the power law exponent ζ of the
distribution. We find an average coefficient hζt i = 1.10, with a standard deviation across years14
of 0.08. The Hill estimator technique gives a mean estimate hζt i = 0.90 and a standard deviation
D E1/2
ζt2 − hζt i2 = 0.07. All in all we conclude that, to a good approximation, mutual fund sizes
follow a power law distribution with exponent:

ζS ' 1 (7)

For the purpose of this paper, one can take this distribution of the sizes of mutual funds as
a given. It is in fact not difficult to explain. One can transpose the explanations given for cities
(Gabaix 1999, and the references in Gabaix and Ioannides 2004) to mutual funds. A log normal
process with small perturbation to ensure convergence to a non-degenerate steady-state distribution,
explains the power law distribution with an exponent of 1. Gabaix, Ramalho and Reuter (2003)
develops this idea and shows that those assumptions are verified empirically.
It is only recently, say in the past 30 years, that mutual funds have come to represent a large
part of the marketplace. For the earlier periods, the theory will work if the distribution of large
agents still follow Zipf’s law with a unit exponent. We do not have direct evidence for this, but a
very natural candidate would be the pension funds of corporations. It is very likely that the size of
the pension fund of a firm with S employees is proportional to S, so that their pension funds also
follow Zipf’s law.

3 The theory, assuming a power law price impact of trades

3.1 Sketch of the theory

Assuming a power-law functional form for price impact, we will show how the power law distribution
of large investors generates, through intelligent15 , tactical behavior of the traders, the exponents
ζQ = ζq = 3/2 for the volume, and ζr = ζN = 3 for returns, volumes and number of trades. The
origins of the power law distribution of large investors are fairly well understood16 . In Section 4, we
will provide an explanation for the power law price impact.
In our theory, large volumes and large returns are created by trades made by large agents
(investors). When a large agent wants to trade with reasonable speed, he moves the market. Trading
slowly decreases the direct price impact, but has other drawbacks, as the profitability of the trading
opportunity may diminish in part due to information leakages. We will show that optimal trading
by large investors generates the exponents ζr = ζN = 3 and ζq = 3/2.
Figure 6 displays the modular structure of the paper. It indicates the logical relations of the
different parts of our theory. Its different parts are fairly independent, which confers some robustness
to our theory. For instance, even if our explanation for the square root price impact ∆p ∼ V 1/2
is erroneous, the rest of the paper still holds. The relations indicated in the Figure 6 are those
articulated in the paper, except for the link between Gibrat’s law and Zipf’s law, which comes from

1 4 We cannot conclude that the standard deviation on our mean estimate is 0.08 (1999 − 1961 + 1)−.5 . The estimates

are not independent across years, because of the persistence in mutual fund sizes.
1 5 Our traders, though intelligent (they try to avoid too high trading costs), are not necessarily hyperrational i.e.,

they may trade too much, as in Daniel et al. (1999) and Odean (2000).
1 6 See Gabaix (1999), Gabaix, Ramalho and Reuter (2003), and Gabaix and Ioannides (2004), and the references

therein.

11
 Figure 6: The modular structure of arguments in this paper

Gabaix (1999), Gabaix and Ioannides (2004) and the references therein. It is logically independent
of the paper, which is why the link is a dashed line.

3.2 Notations and some elementary links between exponents

3.2.1 Notations
We call ζX the power law exponent of a random variable X, i.e. the number such that for a large
real x, P (X > x) ∼ x−ζX . If all moments of X are finite, the convention is ζX = +∞.
Suppose a large investor wishes to Ptrade a large volume V . The large investor divides V into n
parts of sizes q1 , ..., qn , so that V = ni=1 qi . We call ∆p the price impact due to the entire trade.
Over a given time interval, we call r the return, Q the aggregate volume, and N the number of
trades. The variables Q, qi , N , r are directly measurable while V , n, and ∆p are not. We have:
X
r= ∆pj + u (8)
j traded in that interval
X
Q= Vj (9)
j traded in that interval
X
N= nj (10)
j traded in that interval

where u reflects other sources of price movement, such as news.

12
3.2.2 Link between aggregate and trader-based exponents
Our generates power law behaviors for the distribution of ‘unobservable’ variables such as V , n
and ∆p in (8 )-(10). As we show in the proposition below, we relate the power-law exponents for
measurable quantities to the unobservable variables, e.g., we show that the measurable power law
exponent ζQ = ζV .

Proposition 1 We make the maintained assumption that the Vj are independent, that the nj are
independent, and ∆pj and u are independent, and that ζ∆p ≤ ζu . Then

ζr = ζ∆p (11)
ζQ = ζV (12)
ζN = ζn (13)

Proof. We use results (55) from Appendix A.

Assumption ζ∆p ≤ ζu means that the power law of returns is due to the features of our model,
not to news. The Proposition that the sum of power laws with the same exponent ζ is also power
law with the same exponent ζ is very general (see Appendix A). In particular, our results will hold
under much milder assumptions of independence; as shown in Appendix B, they hold even if there is
autocorrelation in market liquidity and trade arrival. For the sake of clarity, we present our results
under the simplest assumption of independence.
Next, we discuss an equally simple but more substantive link.

3.2.3 Link between the exponents of return and volume

A large literature explores the relation between a trade of size V and a price impact r = f (V ).
The initial finding of Hasbrouck (1991) is an increasing, concave price impact. This is usefully
parametrized by a power law for large volumes V :

∆p = hV γ (14)

The following Proposition links the exponents of volume, return and price impact.

Proposition 2 Suppose that we have a power law distribution of volumes with exponent ζQ and a
power law price impact (14) with exponent γ. Then we have a power law distribution of returns with
exponent:
ζr = ζQ /γ.

Proof. As ∆p = hV γ ,
³ ´
P (∆p > x) = P (hV γ > x) = P V > (x/h)1/γ
³ ´−ζV
1/γ
∼ (x/h) ∼ x−ζV /γ .

so ζ∆p = ζQ /γ. We use Proposition 1 to conclude.

Thus the empirical evidence ζQ = 3/2 and ζr = 3 is consistent with a price impact exponent
γ = 1/2. We will give empirical and theoretical support for γ = 1/2 in Section 4. In the rest of this
Section, however, we assume a general γ.

13
3.3 A main result

We start with the following puzzle. If each fund i of size Si traded, at random, a volume Vi
proportional to Si , Vi = ai Si , then the distribution of volumes would follow ζV = ζS . Then by (7)
and Proposition 1, we would have ζQ ' 1, in contrast to the empirical value ζQ ' 3/2.
The larger value of ζQ compared to ζS means that distribution of volumes is less fat-tailed that
the distribution of size of investors. This means that large traders trade less often than small traders,
or that, when they trade, they trade in volume less than proportional to their sizes. How can we
understand this behavior?
Intuitively, a likely reason is that large traders have large price impacts, and have to moderate
their trading to avoid paying too large costs of price impacts. This suggest that an important
quantity is
Annual amount lost by the fund in price impact
c (S) =
Value S of the assets under management of the fund
For example, if funds of size S pay on average 1% in price impact a year, c (S) = 1%.
The above arguments motivate the following theorem which constitutes a central result of this
paper:

Theorem 3 If the following conditions hold:

(i) The size (assets under managements) of large financial market participants follows a power
law with exponent ζS ;
(ii) A large volume V has a price impact:

∆p ∼ V γ (15)

with γ > 0;
(iii) Funds trade in volumes
V ∼ Sδ (16)
for some δ > 0;
(iv) Funds adjust trading frequency and/or volume so as to pay an amount transaction costs,
relative to their assets:
c (S) = C (17)
(survival constraint);
Then returns and volumes are power law distributed with the exponents:
1 ζS − 1
ζr = 1 + + (18)
γ γδ
ζS − 1
ζQ = 1 + γ + (19)
δ
We state a useful Corollary.

Corollary 4 In additions to the assumptions of Theorem 3, assume: (i) Zipf ’s law for investors,
ζS = 1, and (ii) the square root price impact, ∆p ∼ V 1/2 (i.e. γ = 1/2), then returns and volumes
are power law distributed with with exponents resp. 3 and 3/2:

ζr = 3
ζQ = 3/2.

14
 We view Theorem 3 and its Corollary 4 as describing the essential features of large movements
in market activity. The power law distribution of returns and volumes comes from the power law
of large investors (Condition i) and the power law impact of trades (Condition ii). Traders want to
avoid paying too much in price impact (Condition iv), which moderates those power laws.
Corollary 4 is relevant, as empirically it seems that ζS ' 1 and γ ' 1/2. Section 4 will discuss
reasons for γ = 1/2. In general, the fatter the tails of the distribution of large agents, the fatter
the tails of the distributions of volumes and returns. This may be relevant in highly regulated or
immature markets, which might not have ζS ' 1.
We comment on the hypotheses before proceeding to the proof. We have discussed Condition
(i) in Section 2.6. The distribution of many other social entities, such as firms, cities and invidual
wealth, follows power law distributions.
Condition (ii) is consistent with most empirical work (Hasbrouck 1991, Hasbrouck and Seppi
2002, Plerou et al. 2002, Lillo et al. 2003). We endogenize it in Section 4.
Condition (iii) is a weak technical assumption which we endogenize in Section 3.4. Condition
(iii) can be further weakened considerably. The result would still hold, up to logarithmic corrections,
if we had V (S) ∼ exp (aS α ) for positives a, α. Condition (iii) assumes that V (S) “fast enough” in
S, i.e., more rapidly than logarithmic.
Condition (iv) means that funds in the upper tail of the distribution pay roughly similar annual
price impact costs c (S) reaches an asymptote for large sizes. We provide a foundation for this in
Section 3.4. We interpret condition (iv) as an evolutionary “survival constraint”. Funds that would
have a very large c (S) would have small returns and would be eliminated from the market. The
average return r (S) of funds of size S is independent of S (Gabaix, Reuter and Ramalho 2002).
Since small and large funds have similarly low ability to outperform the market (see e.g. Carhart
1997), c (S) is also independent of S.
We now proceed to the proof of the proposition.
Proof. Each time an investor trades, he incurs a price impact proportional to V ∆p. Given (ii),
this cost is hV 1+γ . If F (S) is the an annual frequency of trading, the total annual dollar lost in
transactions costs is F (S) h[V (S)]1+γ , i.e. a fraction

c (S) = F (S) h[V (S)]1+γ /S (20)

of the value S of this portfolio. Hypothesis (iii) gives that V (S) and/or F (S) will adjust so as to
satisfy:
−(1+γ)
F (S) ∼ S · V (S) . (21)
By condition (i) there is a number G (S) ∼ S −ζS of traders of size greater than x . So the density
of traders of size S is ρ (S) = −G0 (S) ∼ S −1−ζS . They trade with frequency F (S) given in (21).
Volumes V > x correspond to traders of size S such that S δ > x. Putting all this together:
Z Z
P (V > x) ∼ F (S) ρ (S) dS ∼ S 1−(1+γ)δ S −1−ζS dS.
S δ >x S>x1/δ
h i∞
∼ −S −(1+γ)δ−ζS +1 1/δ = x−(1+γ)−(ζS −1)/δ .
x

which gives (19). The value of ζr comes from this and Proposition 2.
One can state a result independent of a particular value of γ.

Corollary 5 (Reciprocity law between the exponents of returns and volume). Under the assumptions

15
of Theorem 3, assuming also ζS = 1 (Zipf ’s law for large investors), one has, independently of γ:
1 1
+ = 1. (22)
ζr ζQ

Proof. This comes from direct calculation from (18) and (19):
1 1 1 1
+ = + = 1.
ζr ζQ 1 + 1/γ 1+γ

The “reciprocity law” (22) is likely to be particularly robust, as it does not depend on a specific
value of γ. It is verified for our empirical values ζr = 3, ζQ = 3/2.

3.4 An optimizing model that illustrates Theorem 3

It is useful to illustrate the “bare-bones” model in Theorem 3 by an economic model. To this end,
we introduce a motivation to trade and a reason for why large funds moderate their price impact
as in condition (iii). For the motivation for trade, we suppose that fund managers perceive i.i.d.
mispricings that are independent across different funds. We will consider a world made of K assets
indexed by i, but the intuition works for K = 1 asset. Suppose a manager believes that there is
a mispricing Mi , i.e. that the asset will have excess returns of size Mi . Without loss of generality,
we here take Mi > 0, so that the manager want to buy the asset. Suppose that the fund managers
trade on perceived mispricings which occur them Fi ≤ Fimax times a year17 . Given a mispricing Mi ,
and a volume Vi invested in it, the dollar profit from a trade of size Vi is the mispricing in excess of
the transaction cost: Vi (Mi − ∆pi ), where the price impact ∆pi is

∆pi = hi Viγ . (23)

In the condition (iii) of Theorem 3: Traders do not wish to lose more than C% of their funds
in transaction costs. Intuitively, this arises because if they follow their instincts to trade too much,
their returns may be too small. To capture this, we assume that the trader believes with probability
π the mispricings are real, while with probability 1 − π the mispricing is a false perception. If the
trader’s estimated mispricings are reliable his expected returns are
1X
r+ = Fi Vi (Mi − ∆pi )
S i

1 7 Before we proceed, we comment on possible interpretations of M . The simplest is that the trader thinks that
i
the market price is wrong. There is a lot of evidence that investors do not have all the same assessment of the fair
valuations, which give rise to the perceived mispricing Mi . Bagwell (1992) offers early evidence. Diversity in forecasts
offers direct evidence for this, and the very high trading volume offers strong indirect suggestion of heterogenous
valuations. There is evidence for return predictability (Barberis and Thaler 2003, Hirshleifer 2001, Shleifer 2000) —
so mispricings get progressively corrected, captured in our model by the µT term. A second interpretation for the
trading motive Mi is that a given point in time, the investor may hold his optimum portfolio. The difference between
his current portfolio and his optimum portfolio costs him µ per unit of time, and is proportional to the dollar amount
at stake, V . Hence taking T units of time to realize his trade costs him µT V . So, the derivations in the paper go
through with that interpretation. We now proceed to examine the tradeoff that the manager faces between taking
advantage of perceived opportunites M and the trading costs.

16
But if his intuitions are wrong, and markets are efficient, his returns are:
1X
r− = − Fi Vi ∆pi
S i

The traders has an objective function U on returns, so that his expected utility is:
¡ ¢ ¡ ¢
W = πUΓ,r r+ + (1 − π) UΓ,r r− . (24)

We will consider the class of utility functions:

UΓ,r (r) = u (r + r) − Ze−Γ(r+r) .

for some positive Z, Γ and some increasing weakly concave function u. UΓ,r is increasing and concave
and satisfy:

lim UΓ,r (r) = u (r + r) if r + r > 0

Γ→∞
= −∞ if r + r < 0

In other terms, the trader has a standard utility function for returns r above a threshold −r, but
wants to avoid getting returns below −r. We can think of −r < 0 as some low underperformance
threshold below which the fund managers lose their job. Γ is an aversion to underperformance.
We state the Proposition.

Proposition 6 In the setup above, suppose that the “cost cap” r is low enough, so that the survival
constraint (iv) in Theorem 3 holds (c.f. Eq. (33)). Then, as the aversion to underperformance Γ
become very large, hypothesis (ii) and (iii) of the Theorem 3 hold for each stock i, with

δ = 1/ (1 + γ) . (25)

The size of trades is:

µ ¶1/γ
Mi
Vi = k S 1/(1+γ) (26)
hi
for a constant k given in (34). The total amount spent in transaction costs is the value of the
portfolio times r.

Thus this economic model generates the conditions (iii) and (iv) of Theorem 3. It makes a
prediction (25) on the value of δ that could in principle be tested. The volume traded increases with
the size of the trader, but a in less than linearly. The elasticity ∂ ln V /∂ ln S = 1/ (1 + γ) is higher
in markets in markets with low elasticity of price impact γ.
We also see that, in term of Theorem 3, if the constraint (iv) holds for the whole portfolio, it
will hold stock by stock too.
It was expected that, as in (26), funds trade in larger volume when they are bigger, when the
perceived profit opportunities Mi are larger, and the prefactor hi of price impact is small.
Proof. We study directly the case Γ = ∞, as the problem is continuous in Γ. The trader’s
problem is:
XK
max Fi Vi (Mi − ∆pi ) (27)
Fi ,Vi
i=1

17
 K
X
s.t. Fi Vi ∆pi ≤ rS , (28)
i=1
Fi ≤ Fimax , and (29)
∆pi = hi Viγ

The Lagrangian is:

à !
X X X
L= Fi Vi (Mi − ∆pi ) + λ rS − Fi Vi ∆pi + κi (Fimax − Fi )
i i i

and the first order conditions are:

1 ∂L κi
= Mi − (1 + λ) hi Viγ − =0 (30)
Vi ∂Fi Vi
1 ∂L
= Mi − (1 + λ) (1 + γ) hi Viγ = 0 (31)
F ∂Vi

Eq. (31) gives Vi > 0, and (30)—(31) give κi = (1 + λ) γhi Viγ > 0, i.e. (29) binds. So:

Mi
Viγ = (32)
(1 + λ) (1 + γ) hi

The left-hand side of (28) is:

−1−1/γ
X 1+1/γ −1/γ
H (λ) = [(1 + γ) (1 + λ)] Fimax Mi hi
i

and (28) binds iff H (0) > rS, i.e.

1 X 1+1/γ −1/γ
r< (1 + γ)−1−1/γ Fimax Mi hi (33)
S i

As we assume this, λ solves H (λ) = rS. With (32), this yields (26) where:
" #1/(1+γ)
r
k= P 1+1/γ −1/γ
(34)
i Fimax Mi hi

The total cost constraint for stock i is Ci = Fimax hi Vi1+γ /S, i.e.
1+1/γ −1/γ
F max Mi hi
Ci = P i 1+1/γ −1/γ
r
max
j Fj Mj hj
P
In particular i Ci = r.

18
4 The square root of price impact of trades: Evidence and a
possible explanation
At this stage, we understand how a power law price impact of trading, and a power law distribution
of traders sizes, gives rise to power law distributions of volume and returns. We now explore why a
particular value of the price impact, γ = 1/2, is favored on empirical and theoretical grounds. The
initial motivation is Proposition 2, which shows that a square root impact γ = 1/2 precisely gave
the right link between ζr = 3 and ζQ = 3/2. So we first present empirical evidence that supports
γ = 1/2, and then propose a model that can explain this particular value.

4.1 Evidence on the square root law of price impact

We provide evidence here for the γ = 1/2 in (14), i.e. a price impact function:

∆p ∼ V 1/2 (35)

We start from the benchmark where, in a given time interval, J i.i.d.P blocks are traded, with volumes
V1 , ..., VJ , with i.i.d. signs εj = ±1. The aggregate volume is Q = Jj=1 Vj , and the aggregate return
is:
XJ
1/2
r =u+h εj Vj
j=1

where u is some other source of price movement, assumed to be uncorrelated. Then,

 
£ 2 ¤ X X
E r | Q = σu2 + h2 E  Vj + εj εk Vj Vk | Q
j j6=k

= σu2 2
+ h Q + 0.

We gather: £ ¤
E r2 | Q = σu2 + h2 Q (36)

Figure 7 shows the empirical result18 . It reveals an affine relation, rather than any clear sign of
concavity or convexity. A formal test that we detail in Appendix F confirms this. We thus name
(35) as a good candidate for an empirical law, and now propose a theory for it.

4.2 Sketch of our explanation for the square root price impact
We now present a sketch of our explanation for (35) before developing it further. The basic mi-
crostructure models (e.g. Kyle 1985) give a linear19 price impact ∆p ∼ V . Virtually all the empirical
evidence (starting with Hasbrouck 1991), however, shows a concave price impact. Put succinctly,
the argument is the following. In our model, a trader who is willing to wait for an amount of time

1 8 Theconstruction of the confidence interval requires special care, as r2 has infinite variance.
19 A concave price-volume is given by Seppi (1990), Barclay and Warner (1993), and Keim and Madhavan (1996).
These models do not predict a square root, however.

19
 10

E (r | Q)
5

0
0 10 20
Q

£ ¤
Figure 7: Conditional expectation E r2 | Q of the squared return r2 given the volume Q. The bands
represent 95% confidence intervals. The theory predicts, for large Q’s, a relation E[r2 |Q] = α + βQ,
the “square root” price impact of volume. r is in units of standard deviation, and Q in units of the
first moment. The time interval is ∆t = 15 min, and the results are averaged over N = 116 liquid
stocks. Formal tests reported in Appendix F show that one cannot reject E[r2 |Q] = α + βQ large
enough (Q ≥ 3).

T to realize his buy trade will incur a price impact:

V
∆p = (37)
T
so that Kyle (1985)’s linearity of the price impact in V holds, but for a given T . But more patient
traders who are willing to wait for a larger time T , will be have smaller price impact. Suppose that
asset’s price is increasing at a rate µ. The total benefit from the trade of size V will be:

B = V (M − µT − ∆p)

The optimal ∆p maximizes B subject to (37), so that

µ ¶
V
∆p = arg max V M − µ − ∆p
∆p ∆p

which gives
∆p = (µV )1/2 ∼ V 1/2
and T ∼ V 1/2 . In words, the price impact of a trade of size V scales like less than linear (∆p ∼ V )
because large traders are willing to wait for a longer time T to moderate their price impact. In other
words, in order to execute a trade of size V , traders are on the average more patient and wait for a
time T ∼ V 1/2 , so that Eq. (37) gives Eq. (35). We now proceed to the model that gives (37).

20
4.3 A microstructure model for the power law impact of a block
We will present a search model for the way in which a trader finds counterparts for a large trade.
Our model is a search model. Other models include He and Wang (1995), Saar (2001), Vayanos
(2001), Duffie, Garleanu and Pedersen (2002).

4.3.1 The behavior of liquidity providers

We will present a model where a seller of size si who received from the buyer a price concession ∆p
is willing to supply a number of shares

qi = αsi |∆p| (38)

where α is some proportionality factor. This linear supply could be the linearization of a host of
models. We present one such model in Appendix D.

4.3.2 Permanent vs Transitory price impact

In most microstructure analyses, the price impact has a temporary and a permanent component.
The typical implementation is:
Assumption A on the temporary and permanent price impact. If the mid-quote price before the
trade was pt−1 , and the transaction price was pTt rade = pt−1 + ∆pt , then the next mid-quote price
is on average pt+1 = pt−1 + b∆pt with b ∈ (0, 1).
Various ways to model this are offered in the literature (Biais et al. 2002, O’Hara 1995). Here
weadopt a very simple justification for this. Our results would not change if bt had a seasonal (e.g.,
time of the day) component, or was random with power law ζb ≥ 3.

Proposition 7 Suppose that in the model there are risk neutral arbitrageurs who think that, when
a traded price differs from the previous quoted price, the trade comes from probability b from an
informed, risk neutral trader, and with probability 1 − b from uninformed trader. Then, they will will
enforce Assumption A above.

Proof. Call ft the fundamental price, Ft the information at time t, and It = 1 if the trade is
informed and 0 otherwise, so that b = P (It = 1). Arbitrageurs impose pt−1 = E [ft−1 | Ft−1 ], and
the after-trade price is:
£ ¤
pt+1 = Et ft+1 | pTt rade and Ft−1
£ ¤ £ ¤
= bEt ft+1 | pTt rade , It = 1 and Ft−1 + (1 − b) Et ft+1 | pTt rade , It = 1 and Ft−1
= b (pt−1 + ∆p) + (1 − b) pt−1
= pt−1 + b∆p.

The above justification is not particularly interesting, as it does not make any prediction. Its
only merit is to close our model in a simple way.

4.3.3 The behavior of the large trader

We consider a large trader A who wants to buy a large number of shares V . To accomplish the
trade, the trader will search for counterparts. When all the counterparts are found, the trade is

21
executed, i.e. the trades with the counterparts are executed at the same time. Our description is a
stylization of the “upstairs” market, and is in the spirit of other search models in the literature.
At time t = 0 trader A decides to buy a quantity V and looks for a counterpart. Liquidity
providers, indexed by i, are found stochastically20 . Because of (38), liquidity providers are willing
to sell the active trader a quantity of shares qi = αsi ∆p. So after a time t, the active trader A can
buy X
αsi ∆p.
traders i that have app eared b etween 0 and t

The search process stops when the desired quantity is reached, i.e. at a time Te such that21
X
αsi ∆p = V (39)
i arrives b etween 0 and Te

Let f denote the (Poisson22 ) frequency of arrival

h of liquidity
i providers, s their average size, and the
average time for realizing the trade be T = E Te | ∆p, V . Taking expectations on (39), we obtain
f T αs∆p = V , and:
V
T = (40)
αf s∆p
The trader’s goal is to minimize two quantities: the execution time T , and the execution cost
|∆p|. From above, his objective function for the given trade is:

B = V (M − µT − ∆p) (41)
Indeed, as above, the trader expects the asset to have excess returns M . But after a wait of T , the
price has gone up by µT , so that the remaining mispricing is only M − µT. The total dollar profit
B is the realized excess return M − µT minus the price concession ∆p, times the dollar volume
transacted V . Given (41) and (40)
µ ¶
V
∆p = arg max V M − µ − ∆p . (42)
∆p αf s∆p

This give ∆p ∼ V 1/2 , the square root price impact (35). We present this formally:

Proposition 8 In the above setting, γ = 1/2. Traders incur the square root law of price impact:

∆p ∼ V 1/2 , (43)

wait for an mean amount of time

T ∼ V 1/2 ,

2 0 An interpretation is that the active trader A has, perhaps via his broker, calls successively potential buyers. After

time t, the number of potential liquidity providers are proportional to t.

2 1 Here we omit integer constraints. Taking them into account would complicate but not change the analysis.
2 2 To increase descriptive realism, we could assume that the Poisson arrival rate f itself increases with the price

offered |∆p|, as in f (∆p) . It is easy to verify that our results for the power laws will not change if the function
f increases, for large |∆p|, like less than a power law (e.g. the arrival rate tends to saturation maximal value as
|∆p| → ∞, or say f (∆p) ≤ a ln (b + |∆p|)k for some finite a, b, k).

22
and trade with a mean number of liquidity providers

n ∼ V 1/2

The full expressions are:

µ ¶1/2
µV
∆p = (44)
αf s
h i µ V ¶1/2
E Te | V = (45)
αf µs
µ ¶1/2
fV
E [n | V ] = . (46)
αµs

Proof. The loss function (39) gives (44) while (45) comes from (40). As liquidity providers
h i
arrive with frequency f , we have E [n | V ] = f E Te | V .
So large traders create large price impacts, and have to accept a delay in the execution of their
trades. This is qualitatively consistent with the stylized facts found in the empirical literature :
the price impact |∆p| is an increase and concave function of the trade size V (Hasbrouck 1991,
Holthausen et al. 2000, who find that a large part of the impact is permanent), impatient traders
(high cost of time µ) will have a bigger price impact ∆p (Chan and Lakonishok 1993, 1995, Keim and
Madhavan 1996, 1997, Lo, MacKinlay and Zhou 1997, Breen, Hodrick, Korajczyk 2002). Our theory
gives a precise quantitative hypothesis for this price impact and delay, ∆p ∼ V 1/2 and T ∼ V 1/2 .
Future research might examine directly whether those square root predictions (for large trades) hold
in the data.

4.4 The distribution of individual trades q vs distribution of target vol-

umes Q
Theorem 3 gives reasons for distribution of target volumes with ζV = 3/2. The “upstairs” mechanism
above provides a way to divide them into a several smaller trades of size qi . We describe the resulting
distribution below23 .

Proposition 9 In the search process above, if distribution of target volumes has power law exponent
ζQ > 1, and the distribution of the sizes of liquidity providers has power law exponent ζs = 1, then
the distribution of individual trades has power law exponent

ζq = ζQ .

Proof. We will actually derive the Proposition in a more general form, as here we assume the
following supply function for the number of shares supplies by a liquidity provider of size si to a
large trader who wants to buy a volume V :

qi = αsi ∆p = α0 si V γ

for a γ ∈ (0, 1), not necessarily γ = 1/2. Individual trades of size x come from large trades of size
V (> x). There is a density ρ (V ) ∼ V −(ζV +1) of them. Each large volume V generates a number

2 3 More technical, rigorous results are given in Gabaix, Gopikrishnan, Plerou and Stanley (2003b).

23
n (V ) = V /E [q] ∼ V 1−γ of trades. So we find:
Z
P (qi > x) ∼ ρ (V ) n (V ) P (si V γ > x) dV
V >x
Z
∼ V −(ζV +1) · V 1−γ P (si V γ > x) dV.
V >x

As the liquidity providers’ size distribution is power law with exponent ζs = 1,

P (si V γ > x) = P (si > x/V γ ) ∼ (x/V γ )−1

and
Z ³ x ´−1
P (qi > x) ∼ V −(ζV +1) V 1−γ dV
V >x Vγ
Z
= x−1 V −ζV dV ∼ x−1 · x−(ζV −1) = x−ζV
V >x

so we have ζq = ζV .
A very similar proof would show that ζq = ζV + (ζs − 1) (1 − γ) for ζs 6= 1.

4.5 An optimizing intertemporal model that joins power laws and mi-
crostructure
We now join our model of intertemporal trading with the model of Section 3.4, and our microstructure
model of Section 4.3. It turns out that a key ingredient is price leakages. In their absence, traders
would trade very slowly, creating very little price impact. Leakage is an important concern in trading
large blocks. Keim and Madhavan (1996) shows that in their sample leakage more than doubles the
price impact of large trades24 .
We assume that, during the search process for the block trade by the large trader A, the price
increases at a rate µ + ν, where ν represents the rate of leakage. There is leakage when an investor
B, after hearing that A is going to buy, try to “front run” A and buy before him. Given that leakage
affects the execution price, it should be included in the transaction costs, i.e. count in the constraint
(iii) of Theorem 3.
For the ideas above, we propose the following simplified formalization. Suppose that with a
probability π liquidity provider i will front run A, i.e. trade before him. Liquidity provider i will
have a mean price impact I (si ). Hence the price pressure associated with trading the block is an
increase of the price per unit of time:
ν = πI (si )f ,
where f is the number of liquidity providers contacted per unit of time. The fact that ν > 0 is
important, but the exact expression of ν does not very important.
We adopt the same structure of beliefs and objective function as in Section 3.4. In the trader’s
opinion, the dollar profit from a trade of size V is: V (M − (µ + ν)∆p). So if he has true intuitions
about the market, his the excess returns are:

r+ = F · V (M − (µ + ν)T − ∆p) /S

If his intuitions about the market are false, his excess returns are:

2 4 Without leakage the price impact if 4.3%, while with leakage it is 10.2%.

24
 r− = −F · V (νT + ∆p) /S
Hence, the price impact includes not only ∆t, but also the price increment due to leakage νT .
The trader maximizes W in (24) subject to F ≤ F max and (39), which we rewrite as:

V
∆p = a (47)
T
The solution is given by the next Proposition.

Proposition 10 In the market setup above, we assume that the trader maximizes the objective
function W given in (24). In the regime where Γ À 1 (large aversion to underperformance) and
M À S 1/3 µ (which can be more concretely expressed M À µT for the optimal solution), the solution
of problem above yields V ∼ S δ with δ = 2/3. Thus, the conditions of Theorem 3 hold with δ = 2/3
and the square root price impact γ = 1/2.

Proof. It is easy to see that the constraints bind. Consider the limit µ = 0. The problem
becomes:
max F max V M − rS
∆p,V
µ ¶
V
s.t. F max νa + ∆p V = rS (48)
∆p
1/2 1/2
Optimizing over ∆p, we get ∆p = (νaV ) , and (48) gives 2V 3/2 F max (νa) = rS, hence V ∼
S 2/3 .
The proof is now complete. In addition, it may be instructive to see which variables matter for the
1/2 −1/2 1/2
limit µ → 0. For this define x = (νaV ) /∆p, B = rF max −1 (νa) , η = B 1/3 M −1 S 1/3 (a/ν) µ,
−3/2
and y = BSV . The problem becomes:

max y −2/3 − ηxy −1 s.t. x + x−1 = y

³ ¡ ¢1/2 ´
This forces y ≥ 2, and leads to x = ξ (y) = y − y 2 − 4 /2, and y (η) = arg max y −2/3 −
ηξ (y) y −1¡.We have
¡ y (0) = 2, and
¢¢ the solution is continuous in η. Thus we get y (η) = 2 + o (η), i.e.
V = S 2/3 1 + o µr1/3 S 1/3 M −1 .
Hence our predictions are confirmed by this framework. The cubic law arises if large traders are
forced to trade quickly because of price leakage.

4.6 Some comments on the model

4.6.1 Arbitrage
Our model is does not leave scope for short-term arbitrage, as the trading decisions are unpredictable.
However, it offers some scope for predictability of long term returns. Indeed as volume moves prices
without necessarily informational reasons, prices deviate from fundamentals, and they will revert in
the long term. There is a evidence for this phenomenon, see e.g. Campbell and Shiller (1988). So
our model requires limited arbitrage of long term returns. There are good reasons why long term
arbitrage is difficult, see e.g. Shleifer and Vishny (1997) and Shleifer (2000).

25
4.6.2 Plausibility of very large volumes and price impact
The literature finds a big impact of big traders. Chan and Lakonishok (1993, 1995) find a range
of 30-100 bps, Keim and Madhavan (1996), looking at smaller stocks, find 400 bps. There are also
many colorful examples of large investors moving the market. Some are reported by Corsetti, Pesenti
and Roubini (2002), Coyne and Witter (2002) and the Report on the 1987 crash (United States’
Presidential Task Force on Market Mechanisms 1988).
As it is difficult to detect exogenous trades, measuring a price impact is difficult. There is a more
“intuitive” way to convince oneself that a large fund is likely to move prices. The typical yearly
turnover of a stock is 50% (Lo and Wang 2001), so that with 250 trading days per year, its daily
turnover is 0.5/250=0.2% of the shares outstanding. Consider a moderate size fund, e.g. the 30th
largest fund. At the end of 2000, such a fund held 0.1% of market25 . So, on the average, it will hold
0.1% of the capitalization of a given stock. So if the fund wants to sell its holdings, it will create
an additional 0.1% in turnover of stock, while the regular turnover is 0.2% per day, i.e., the fund
will have to absorb half of the daily turnover. So, the size of the desired trade of this fund is quite
important compared to the normal turnover. This supports the idea that large funds are indeed
large for liquidity of the market, and supports the assumption that big traders will pay attention to
their trading strategy to moderate their price impact.

5 Assessing some further empirical predictions of the model

5.1 Contemporaneous behavior of several measures of trading activity
5.1.1 Definitions
The above model proposes an explanation for the power laws of financial market activity. We
now show how it correctly explains another series of patterns in trading activity, pertaining to the
joint distributions of returns, volumes, and their variants. More precisely, we fix a time interval
∆t = 15min. We will consider the return rt , the volume26 Qt and number of trades Nt in interval
(t − ∆t, t]. We also partition those quantities in buys and sells. Suppose each trade i occurring at
time i, with a quantity of shares exchanged qi , and define εi = +1 is the trade can be identified as
buy initiated, εi = −1 if it was sell-initiated, εi = 0 if no identification is possible27 . We define Nt0
as the net number of trades, i.e. number of “buy” trades minus the number of “sell” trades. Finally
we define Q0t as the net volume, i.e. number of shares exchanged that come from a buy order minus

2 5 It had $19 billion in assets under management. The total market capitalization of The New York Stock Exchange,
the NASDAQ and the American Stock Exchange was $18 trillion.
2 6 The dollar value of the shares traded, or the number of the shares traded divided by the number of shares

outstanding give the same results, as they are proportional in the short term.
2 7 We us the Lee-Ready (1991) algorithm. We identify buyer and seller initiated trades using the bid and ask

quotes PB (t) and PA (t) at which a market maker is willing to buy or sell respectively. Using the mid-value PM (t) =
(PA (t) + PB (t))/2 of the prevailing quote, we label a transaction buyer initiated if P (t) > PM (t), and seller initiated
if P (t) < PM (t). For transactions occurring exactly at PM (t), we use the sign of the change in price from the previous
trade to determine whether the trade is buyer or seller initiated, while if the previous transaction is at the current
trade price, the trade is labelled indeterminate. We use the prevailing quote at least 5 s prior to the trade. 17% of
the trades are left indeterminate.

26
number of shares exchanged that come from a sell order. Formally:
X
Nt = 1 (49)
trade i happ ening in (t−∆t,t]
X
Qt = qi (50)
trade i happ ening in (t−∆t,t]
X
Nt0 = εi (51)
trade i happ ening in (t−∆t,t]
X
Q0t = εi qi (52)
trade i happ ening in (t−∆t,t]

We compute the graphs E [Y | X] for many combinations of X, Y = r, Q, Q0 , N , N 0 , N 0 /N ,

Q/N . The choice was guided by an assessment of how informative such a graph would be, and how
non-trivial the shape predicted by the theory would be. The model reproduces all the shapes of
those graphs, as can be seen in Figure 8 below.
We use simulations to plot the model’s predictions. We compare the empirical and theoretical
conditional expectations E [Y | X], for X, Y = Q, N, Q0 , N 0 , r, N 0 /N, r2 , Q/N . We show our results
only for selected pairs of X, Y that seem nontrivial.

5.1.2 Comparison between theory and facts

In Figure 8a-e we present the empirical values on the left panel, and the model’s predictions on
the right panel. We obtained the model’s predictions using Monte Carlo simulations detailed in
Appendix E. The graphs show that the model broadly matches28 all the E [Y | X] that we construct
empirically. Consider for example E[r | Q0 ] in Figure 8a. Our model reproduces on the left panel the
increasing, concave shape (in the positive region). The aggregation over several trades flattens the
shape of the price impact function. Plotting the symmetrical version, E [Q0 | r], yields a surprising
finding. The empirical graph, in Figure 8b is now linear. The model matches this in the right panel
of Figure 8b. The cause of the linearity, again, is the aggregation over several trades.
The model predicts that periods of large volume are periods when large traders execute trades
in smaller pieces, so that, for a large buy trade, one observes large number of buy smaller trades. A
compact way to say that is that N is high, and |N 0 /N | is high too when the volume is high29 . This
is expressed in Figure 8c.
It is an important feature of the model that large trades beget more trades. Indeed, in our model

|N 0 | ∼ N (53)

for large N and is dominated by one large fund manager who desires to trade a volume Vj , and
1/2 1/2
creates a number of orders Vj , so that nj , N , and |N 0 | have the same order of magnitude, Vj .
Relation (53) means that most trades have the same sign, i.e., move the price in the same direction,

2 8 Given we have 1.5 million observations, it is likely that statistical tests would show that the fit is not perfect.

Our intention here is just to illustrate that the fit is good. To obtain a perfect fit, one would need to complicate the
model, in particular to introduce autocorrelation in the propensity to trade. Here we chose to show that the simple,
non-optimized, i.i.d. version of our model can get a pretty good fit.
2 9 A very different relationship would happen under another plausible model. Suppose that there are N trades
0
PN 0
with i.i.d. signs i )i=1...N and sizes (V³
³√ (ε´ ´, so that Q = i=1 εi Vi . Then a large N corresponds to an N =
i )i=1...N
PN √
i=1 εi = O N , so that N 0 /N = O 1/ N is small. Likewise a large Q0 corresponds in part to a large N, and
a small |N 0 /N|, counterfactually.

27
the sign of the trade of the large fund manager. Equation (53) is indeed consistent with empirical
data shown in Figure 8d. This contrasts with a simple alternative model where each desire to trade
would Pcreate only one trade, as in a competitive market. In this alternative model we would have
N
N0 = 0
i=1 εi , where εi = ±1, leading to |N | ∼ N
1/2
in the tail events, or E[N | N 0 ] ∼ N 02
in contrast to the data in Figure 8d. Figure 8e supports the view that in periods of high volume
imbalance V 0 , most trades change the price in the same direction. Indeed, the data and the model
exhibit a similar sharp transition of N 0 /N as V 0 changes sign.

We tested other cases too and we find good agreement between the theory and empirical data30 .
Here, we report the curves with the shapes we deem most informative. Given that the model was
not constructed to match those facts, we view this as a very reassuring feature of the model.

5.2 Some untested predictions

The model makes a number of untested predictions, which we present in this Section.
• As per equation (21), a large fund of size S will do “price-moving” block trades31 of size V ∼ S δ
with a frequency that depends on size as S/V 1+γ , where γ is the exponent of the price impact
(14).
• Expression (26) for the volume traded in an individual stock, in particular the dependence in
S and the prefactor of the price impact hi (23)
• The expressions (44)-(46) for the price impact, the time to execution, and the number of
trades. In particular, it would be interesting to test their “square root” dependence in S, and
the differential impact of µ (the impatience of the trader), α, f, s.
• The linear supply function (38), saying that a trader willing to wait an amount of time T and
to pay a price increment ∆p, will get an amount of shares S (∆p, T ) that will be proportional
to ∆p · T in the limit of large ∆p and T.
• We anticipate the values of the four trading exponents α, β, κ and τ in Appendix B to be 0,
1, 0 and 1 respectively.
Hence our theory makes predictions about the value of 24 power law exponents32 . We can make
also graphical predictions E [Y | X] graphs that we did not build33 , which yields 56 predictions. We
leave the investigation of those values to future research.

6 Related literature
6.1 Alternative theories
The general theme is that most theories (certainly those that we are aware of) do not explain (and
cannot explain with a simple modification) the cubic and half-cubic laws. At best, they have free
parameters which can be artificially tuned to replicated the ζr = 3.

3 0 The graphs we tested where E [V | r], E [N | r], E [r | V ], E [V 0 | N 0 ], E [N 0 | V 0 ], E [V /N | r], and empirical

finding and theory matched very well.

3 1 So looking at the fund turnover will not be the right thing to do. To provide liquidity, or avoid looking like

a “closet indexer” a large fund may maintain a fairly large turnover, though the “price moving” trades will indeed
happen less often, as they move prices a lot.
3 2 The above bullets make respectively 2, 2, 5 · 3 = 15, 1 and 4 predictions.
3 3 The printout of the 56 theoretical relationships E [Y | X] involving V, N, V 0 , N 0 , r, N 0 /N, r 2 , V /N is available from

the authors upon request.

28
Figure 8: Conditional expectations for (a) E[r | Q0 ], (b) E[Q0 | r], (c) E[N | Q0 ], (d) E[N | N 0 ],
and (e) E[N 0 /N | Q0 ]. We form, for each interval ∆t = 15 min, the quantities (i) r, the returns, (ii)
QB (resp. QS ), the number of shares exchanged in a buyer (resp. seller) initiated trade, (iii) NB
(resp. NS ), the number of buyer (resp. seller) initiated trades, Q0 ≡ QB − QS , and N 0 ≡ NB − NS .
The left panel shows the empirical values for the 116 frequently traded stocks in the Trades and
Quotes database for 1994—1995. This represents 1.5 million observations. Variables are normalized
to unit variance after setting the mean to zero; for variables such as volume for which the variance
is divergent, we have by normalized the first moment instead. The right panel shows the model’s
predictions, which agree well with the empirical data.
29
6.1.1 The public news based (efficient markets) model

In this model, price movements reflect without frictions public pieces of news. One would need to
make large number of assumptions to make this model fit the empirical power laws. First, because
returns reflects news, we have to assume that the distribution of news has ζr = 3.
This is not the only difficulty. In the benchmark where volumes do not move prices, there is no
reason for big traders to trade less than small ones. With no price impact, essentially all models that
traders should trade in amount proportional to their sizes, so that the volume Vi traded by trader i
will be equal to Vi = ai Si for ai a random variable that do not scale with size. Then, we would get
ζV = ζS = 1.This prediction is a major difficulty with the news approach. Otherwise, to work with
the public news model, one would have to assume ζV = 3/2 , for some inexplicable reason. Hence
the values of ζr and of ζQ are just assumed in that model and do not arise as a consequence of a
theory.

6.1.2 A mechanical “price reaction to trades” model

In this model, the desired N trades are i.i.d. The price has a reaction f (εj qj ) to each trade of size
P
qj and sign εj . So r = Jj=1 f (εj qj ) and the other variables are defined as in (49 )-(52).
This model is a variant of our model, and so it will tend to generate qualitatively right E [Y | X]
curves. However there are problems with this model. Firstly, we need to assume ζQ = 3/2. However,
given ζS = 1, the natural prediction would be ζQ = 1. Secondly, we need to assume ζN ' 3. It
could be any other value. Thirdly, we also need to assume a price impact ∆p ≈ V 1/2 to get the
shape of E[r2 | Q].

6.1.3 Random bilateral matching

This has been proposed by Solomon and Richmond (2001). The model says that two traders of
size S, S 0 meet, and the resulting volume is max(S, S 0 ) and the price impact is min(S, S 0 ). While
this approach is reasonable, it is not without problems. As the initial size distribution is ζS = 1,
ζmin(S,S 0 ) = 2, so that one would predict ζQ = ζr = 2. The authors rely on a scaling exponent of the
size of agent ζS = 3/2, as is approximately the case in the US and the UK in the late 1990s, but
not in other periods. A major difficulty for this theory is that empirically the exponent is individual
wealth is highly variable. Feenberg and Poterba (1993) show that the exponent for the wealth of
U.S. individuals had large variations from 1.5 to 2.5 in the last 30 years. This large variability is
confirmed in a series of papers on U.S. and international data by Piketty and Saez (e.g. Piketty and
Saez 2002).

6.2 Related empirical findings

6.2.1 Other models for returns
The empirical literature has proposed other distributions. We are more confident about our findings
rely on much larger number of data points, hence quantify the tails more reliably. We can also explain
previous findings in light of ours. Andersen et al. (2001) show that the bulk of the distribution of
realized volatility is lognormal. In independent work Liu et al. (1999) show that while this is true,
the tails are power law. Gopikrishnan et al. (1999) also report how, with a small sample that looks
only at the center rather than the tails of the distributions, one would find Lévy or truncated Lévy
distributions, as was found by Mantegna and Stanley (1995).
P t
Finally Ané and Geman (2000) propose a model which is in essence rt = N j=1 ui where ui are
normal shocks. This is a variant of the Clark (1973) model where the clock is driven by the number

30
of trades rather than the volume. This model has an important empirical problem: from ζN ≥ 3, as
|r| ∼ N 1/2 in this model, we conclude ζr = 2ζN ≥ 6, so that model fails to reproduce the empirical
cubic law of returns. In fact, in our model, results similar to Ané-Geman would be found: for
instance, E[r2 | N ] looks almost linear (though it is not really linear). Under the null of our model,
their model is a good, but ultimately inaccurate, representation of reality.

6.2.2 Buy / Sell asymmetry

Our basic model has built-in full symmetry between positive and negative tails. This is not a bad
approximation of reality: the exponents are roughly identical, around 3. However, it is true that
the exponents are slightly lower (fatter tails) in the negative tail that in the positive one. More
generally, one finds a bit more skewness in the negative tail. One reason may found in the “leverage
effects”. Volatility goes up after negative returns. Mechanically, this skews to the left the returns
over a non-infinitesimal ∆t, and the effect grows with ∆t.
Our framework suggests another hypothesis. Suppose the frequency of arrival f of liquidity
providers is higher for desired sells (so the liquidity providers have to buy the asset) than buys.
Chan and Lakonishok (1993) report the “Street” explanation for this34 . When one looks to buy a
large amount, one can only deal with liquidity providers who already own the assets, if shorting is
difficult. But all liquidity provider can absorb a sell order. So it is easier to meet counterparts for
sells than for buys, i.e. f+ < f− , where f+ (resp. f− ) is frequency of arrival of liquidity providers
that are willing to be counterpart of a buy (resp. sell) order, f− . By plugging f+ or f− in (70)
we see that |∆p+ | > |∆p− |, i.e. the theory predicts higher price movements for buys than for sells,
though also identical exponents.

6.3 Link with the behavioral and excess volatility literature

There is a growing literature suggesting that many movements in the markets are hard to explain by
any kinds of news. Cutler, Poterba, Summers (1989) look at daily returns, Fair (2002) looks at the
largest 1 to 5 min S&P500 returns. Both studies find that many are not explainable by news. The
puzzle is stated in related ways by French and Roll (1986), Roll (1988), and Shiller (1989). Trading
per se seems to create volatility (Evans and Lyons 2002), and returns, even ex post, are hard to
explain. Our theory would predict that, on those events of large movements with no news, trading
volume is abnormally high. It would also predict that those returns should mean revert.

6.4 Link with the microstructure literature

The literature from which this paper draws most is perhaps the microstructure literature. This
literature is extremely vast (for reviews see e.g. O’Hara 1997 and Biais et al. 2002), so we will only
mention the most directly relevant pointers. Our theory makes predictions above the relationships
between volume and price impact, and we reviewed in Section 4 the papers directly pertaining to
our theory. Other relevant papers, that show a long run impact of trading on prices include Evans
and Lyons (2002), and the analysis of volume in Wang (1994), Lo and Wang (2001). Rather than
the fine analysis proposed in the literature (e.g. Madhavan and Cheng 1997), we adopted a fairly
stylized view of trading institutions. This was motivated by the fact that the cubic exponents arise
for different market structures, so that our explanation should not depend too finely on a specific
market structure.
Our goal is to explain the cubic power laws, and the patterns in trading activity at a fairly
high degree of aggregation (1/2 hour as opposed to trade by trade). Hence we highlight some

3 4 Saar (2001) provides an alternative explanation based on asymmetric information.

31
traits of reality — mainly the heterogeneity between the size of agents, and the trade-off between
execution cost and execution time — that are typically not central to the microstructure literature.
Another feature is that we are agnostic about the importance of information asymmetry in the
determination of asset prices. We highlight more “macro” phenomena that arise from aggregation,
rather the detailed understanding of micro situations that tend to be the focus of the microstructure
literature. Hence, our model and the models of this literature appear quite different. Still, it would
be desirable, in further research, to merge the rich informational and institutional understanding
from the microstructure literature and the more macro approach of the present paper.

6.5 Link with the “econophysics” literature

This paper is part of a broader movement utilizing tools from physics in for the study of economic
issues. In the same way that the application of physics to earth sciences and biology are called
geophysics and biophysics, this literature is sometimes called “econophysics”. This literature was
pioneered by Mandelbrot (1963). This literature is growing, and includes Bak, Chen, Scheinkman
and Woodford (1993), Bouchaud and Potters (2000), Canning et al. (1998), Gabaix (1999, 2003),
Plerou et al. (1999), Lux and Sornette (2002), Lévy, Lévy and Solomon (2000), Mantegna and
Stanley (2000).

7 Conclusion
This paper proposed a theory of power law behavior of several variables that describe financial
market activity. It explains the cubic and half-cubic exponent of returns and volume, and also
makes many predictions about other quantities. Section 5.2 presents numerical predictions about 24
other power law exponents. This offers ample room for future research. We have examined some of
its predictions, in particular in Figure 8. Further research35 is made is by the modular structure of
our arguments, as shown in by Figure 6.
This theory can also be used for policy analysis. The cubic law predicts the existence of very
large returns, and Gabaix et al. (2003c) report that stock market crashes such as the 1929 or 1987
crashes do do not seem to be outliers to the cubic law. Hence if we understand the origin of the
cubic law, we may understand the origin of stock market crashes. Within the present framework,
Gabaix et al. (2003d) examine the performance of traditional proposals against crashes, such as
the Tobin tax or circuit breakers. It shows that they not dampen crashes, and works out different
proposals that in principle would achieve that goal.
On the substantive front, our model provides one possible quantitative theory of excess volatility
in asset markets: it is simply due to the desire to trade of large traders (perhaps stimulated by news).
More precisely, it is the power law 1 of the distribution of sizes that generates, through an intelligent
though not hyperrational (people trade too much) trading process, the power laws of 3 in returns
and 3/2 in volume. On the methodological front, we introduce some new questions that finance
theories should answer. Matching, as we do, the quantitative empirical regularities established here
(in particular explaining the exponents of 3 and 3/2, not merely assuming them), in particular the
cubic laws, would be a sine qua non criterion for the admissibility of a of volume and volatility. We
hope that the regularities we established will sharply constrain, and guide, future theories. Given

3 5 Our model has very little to say about the low frequency structure of trading. Our model contains a series of

independent trading decisions, and is silent about the time-correlations in market activity. It is straightforward to
propose simple extensions of it, such as an GARCH-type model for the propensity to trade Jt , so as to qualitatively
account for the well-known long term memory in volatility (where Jt+1 depends positively on past Jt0 s and |rt |). The
proper analysis of this, however, is outside the scope of this paper. We are investigating this in ongoing research.

32
its empirical success and its simple structure, the present model might a useful point of departure
to think about those issues.

33
8 Appendix A: Some power law mathematics
We present here some basic facts about power law mathematics, and show how their great aggre-
gation properties makes them especially interesting for both theoretical and empirical work. They
also show how our predictions are robust to other sources of noise.
A random variable X has power law behavior if there is a ζX > 0 such that:
1
P (X > x) ∼
xζX
so that the probability density p (x) follows:
1
p (x) ∼
xζX +1
A more general definition is that there is a “slowly varying”36 function L (x) and a ζX s.t.
p (x) ∼ L (x) /xζX +1 , so that the tail follows a power law “up to logarithmic” (by some abuse of
language) corrections.
ζX < ζY means that X has fatter tails than Y , hence the large X 0 s are (infinitely, at the limit)
more frequent than large Y 0 s.
α α
The definition implies that with α > 0, E [|X| ] = ∞ for α > ζX , and E [|X| ] < ∞ for α < ζX .
α
As an example, if for returns ζr = 3, then E [|r| ] for α > 3. In particular, the kurtosis of returns
in infinite37 , and their skewness borderline infinite. For Y a normal, lognormal, exponential, the
formal power law exponent is ζY = ∞, given Y dies out faster than any power laws.
Power laws have great aggregation properties. The property of being power law is conserved
under addition, multiplication, polynomial transformation, min, max. The motto is that, when we
combine two power law variables, “the biggest (fattest=smallest exponent) power law dominates”.
Indeed, for X, Y independent variables, we have the formulaire:

ζX+Y = ζX·Y = ζmax(X,Y ) = min (ζX , ζY ) (54)

= power law of fattest variable
ζmin(X,Y ) = ζX + ζY

For instance, if X is a power law for ζX < ∞, and Y is power law variable with an exponent
ζY ≥ ζX , or even normal, lognormal or exponential variable (so that ζY = ∞), then X + Y, X · Y ,
max (X, Y ) are still power laws with the same exponent ζX . So multiplying by normal variables,
adding non fat tail noise, summing over i.i.d. variables preserves the exponent. So (i) this makes
theorizing with power law very streamlined; (ii) this lets the empiricist hope that those power laws
can be measured, even if there is a fair amount of noise in the data. One does not need to carry
around the additional noise, because though it will affect variances etc., it will not affect the power

3 6 L (x) is said to be slowly varying (e.g. Embrechts et al. 1997, p.564) if

lim L (tx) /L (x) = 1 for all t > 0.
x→∞

Prototypical examples are L = a and L (x) = a ln x for a non-zero constant a.

3 7 This makes the use of the kurtosis invalid. As the theoretical kurtosis is infinite, empirical measures of it are

essentially meaningless. As a symptom, according Lévy’s theorem, the median sample kurtosis of T i.i.d. demeaned
³P ´ ³P ´2
T 4 T 2
variables r1 , ..., rT , with κT = i=1 ri /T / i=1 ri /T , increases to +∞ like T 1/3 , as the sample size T
increases. The use of kurtosis should banished. As a simple diagnostic for having “fatter tail than from normality”,
we would recommend, rather than the kurtosis, quantile measures such as P (|(r − hri) /σr | > 1.96) /.05 − 1, which is
positive if tails are fatter than a normal.

34
law exponent. Power law exponents carry over the “essence” of the phenomenon: smaller order
effects do not affect the power law exponent.
For instance, say a theory, for instance ours, gives a mechanism for R, with ζR = 3. Other things
are going on, so that in reality, we observe:

rf f
it = a
e f
it Rit + bit

For instance, the liquidity of the market varies, so that af f

it is random, and news can affect prices bit
without affecting volume. But even then, we will have ζr = ζR = 3 if e a, eb are smaller order effects,
i.e. have thinner power laws (ζa , ζb ≥ 3). If the theory of Rit capture the first order effects (i.e.
those with dominating power law), its predictions for the power law exponents of the “noised up”
empirical counterpart rfit will still be true.
Proof. See Breiman (1965) and Gnedenko and Kolmogorov (1968) for rigorous proofs, and
Sornette (2000) for intuitive, heuristic derivations. Here we just indicate the proofs for the min and
max operations. For ζmin(X,Y ) , we have:

P (min (X, Y ) > x) = P (X > x and Y > x) = P (X > x ) P (Y > x)

k k0 kk 0
= ζ ζy
= ζX +ζY
x xX x
Also,

P (max (X, Y ) > x) = 1 − P (max (X, Y ) < x) = 1 − P (X < x and Y < x)

= 1 − P (X < x ) P (Y < x)
µ ¶µ ¶
k k0 k 00
= 1 − 1 − ζX 1 − ζy ∼ min(ζ ,ζ )
x x x X Y

where k 00 = k if ζX < ζY , k00 = k0 if ζX > ζY , and k = k + k 0 if ζX = ζY .

This generalizes to K independent power law variables:

ζX1 +...+XK = ζX1 ·...·XK = ζmax(X1 ,...,XK ) = min (ζX1 , ..., ζXK ) (55)
ζmin(X1 ,...,XK ) = ζX1 + ...ζXK

Finally, a useful formula is:

ζX
ζX α = for α > 0 (56)
α

The proof is the same as in Proposition 2.

9 Appendix B: Generalization of Theorem 3 to non i.i.d. set-

tings
We sketch here how our main results, including Theorem 3, extend without difficulty to non i.i.d.
settings. For instance, there can be autocorrelation in market liquidity and trade arrival. Also,
different traders can have different trading frequencies. We index large traders by a. We say that
τat = 1 if a does trade at time t in that interval, τat = 0 otherwise. So, the total volume and returns

35
created by the fund are respectively:

Vat = τat kat S δa (57)

1/2
rat = ±τat λat Vat (58)

for some possibly time-varying, correlated prefactors kat , λat . While the paper is, for simplicity,
i.i.d., this formulation allows the prefactors kat , λat and other prefactors to be correlated between
stocks and across time. For instance, a period with high kat and low λat has large typical volume
and low price impact. This captures the presence of time-varying liquidity and volume that have
been detected in empirical data (see e.g. Lobato and Velasco 2000).
The aggregate volume and return are:
X
QAgg
t = Vat (59)
a
X
rtAgg = rat (60)
a

as each trader creates a volumes Vat and moves prices by an amount rat . They correspond to the
observed return and volume, simply called rt and Vt in the paper.
We need to show that ζrAgg = 3 and ζQAgg = 3/2 in that context also. Showing this is a bit
involved. The difficult part is to show that rat and Vat have respectively power law 3 and 1.5. The
general properties of power laws make rtAgg and QAgg
t inherit the power laws of rat and Vat , i.e.
ζrAgg = 2ζQagg = 3.
We now proceed to show rat and Vat have respectively power law 3 and 1.5. The result is stronger
than the one in the body of this paper:

Proposition 11 We index funds by a, and dates by t. We call θa the “trading style” of fund a,
and εat indexes market—wide and fund—specific time-varying characteristics. We assume (i) fund’s
sizes follow Zipf ’s law; (ii) the price impact follows
1/2
|rat | = λat Vat ;

(iii) fund a, of size Sa , trades in volumes

Vat = kat S δa ;

(iv) fund a’s total transaction costs are equal to Ca Sa . We assume the following stochastic structure:

λat = λ (θa , εat )

kat = k (θa , εat )
δa = δ (θa )
Ca = C (θa )

for some functions λ, k, δ and C with compact support in (0, ∞), and θa , εat are (possibly multidi-
mensional) random variables, independent of the sizes (Sa )a=1...n , identically distributed across a.
(θa )a=1..n and (εat )t;a=1...,n are independent, but the εat , assumed to be stationary, can be correlated
across t’s and a’s. The probability that fund a will trade at t has the form

Fat = G (Sa , θa ) H (θa , εat )

36
so that it is dependent on shocks εat and the size Sa but in a form that is weakly separable. Then, the
distribution of the individual (before aggregation) volumes and returns follow power law distributions
with exponents:
ζQat = 3/2 and ζrat = 3.

We comment on the hypotheses. There are four main improvements upon the text, which assumes
constant functions λ, k,δ and C, and Fat = G (Sa ).
Firstly, δa and Ca need not be the same across funds: they can depend on the “trading style”
θa of fund a, though we keep the assumption that they are independent of the size Sa . Secondly,
the probability of trading, Fat , the volume traded Vat , and the price impact |rat | can vary from
trade to trade. This can reflect changing “market conditions,” such as the liquidity 1/λ. For
instance, the probability of trading Fat could increase with the probability of trading. Thirdly, we
allow intertemporal correlation in the noise structure via εat . This allows for autocorrelation in the
trading frequency. Fourthly, the probability of trading Fat , the volume traded Vat = S δa k (θa , εat ),
1/2
and the price impact, |rat | = λ (θa , εat ) Vat can be correlated. For instance, the volume k could be
higher when the liquidity 1/λ is higher. The fact that εat is multidimensional allows a large class
of imperfect correlations: one component of εat can be a market-wide term dependent on t, and
another can be an idiosyncratic term dependent on a and t. Also, the price impact can vary on the
“nature” of the fund, represented by θa (for instance degree of impatience represented by µ in the
paper can be part of a multidimensional θa ).
The crucial hypothesis is that the distributions of (δa , kat , λat , Ca ) are “scale invariant,” i.e. do
not depend on Sa . The essence of the proof is similar to the one in the paper, though the notations
are more involved. The proof is fairly technical, and would be too long to reproduce here. It is
available from the authors upon request.

10 Appendix C: The model with general trading exponents

We present here the general structure of the model. It allows to see which assumptions are crucial
to get the “cubic” exponents. Also, natural quantities, the “exponents of trading”, emerge. Their
direct measurement is an interesting task for future research. We call ζS and ζL the Pareto exponents
of the large traders and the liquidity providers, respectively. We make the following Ansatz for the
behavior of our variables in the large S, |∆p|, and T limit:

• A large trader of size S will make large position of size V = S δ at an annual frequency S −φ .
• If a large traders makes an order at a price concession ∆p, then the frequency of arrival of
liquidity traders is f = ∆pα , and each liquidity provider of size s supplies a number of shares
q = s · ∆pβ .
• A trader of size S will adjust the trading frequency so as to pay a proportional amount of
transactions costs S κ .
• The large trader, in a given trade of size V , wants to minimize a loss function ∆p+µT τ where T
is the mean time needed to find enough liquidity providers to complete the trade. For instance
τ < 1 would correspond to a concave arrival time of information to the market.

The original model corresponds to a value of 0 for α and κ, and a value of 1 for ζS , ζL , β, and τ .
The next Proposition describes the results.

37
Proposition 12 Under the above conditions, which generalize the model in the text, the exponents
of returns, volumes, individual trades, and number of trades are:
µ ¶
ζS + φ 1
ζr = +α+β (61)
δ τ
ζS + φ
ζQ = (62)
δ
ατ + 1
ζq = ζQ + (ζL − 1) (63)
(α + β) τ + 1
µ ¶
ζS + φ βτ
ζN = 1+ (64)
δ ατ + 1

The measured price impact function is ∆p ∼ V γ with

τ
γ= (65)
(α + β) τ + 1

and the exponent of the trading frequency is:

φ = δ (1 + γ) − κ − 1 (66)

Hence the exponents of returns, volume and number of trades increase with the exponent of the
distribution of traders ζS . Somehow forcing large traders to trade less, other things equal, is going
to decrease the fat-tailness of returns, i.e. increase ζr . Decreasing the time pressure to trade, in
the sense of lowering τ , also decreases the fat-tailness of volumes. Equation (65) shows how to get
values for γ different from 1/2, the value we mostly used in the paper.
Proof. The probability that a target volume V is > x is:
Z
ζS +φ
P (V > x) ∼ S −1−ζS S −φ dS ∼ x− δ
S δ >x

so that the distribution of target volumes is (62).

After a time t, the large trader has received on average ∼ ∆pα+β t shares. So the mean time
needed to realize a trade of size V is T = h0 V /∆pα+β , and the objective function is:

Vτ
min ∆p + µh0
∆p ∆pτ (α+β)

which yields ∆p ∼ V γ with γ given by (65). This gives (61) by (56).

Each transaction costs V ∆p (V ) ∼ V 1+γ ∼ S δ(γ+1) , so the annual amount will be ∼

S −φ · S δ(1+γ) = S 1+κ

which gives (66).

The number of trades is n ∼ V /∆pβ , which gives
ατ +1
n∼V (α+β)τ +1 . (67)

and (64) by (56).

38
 ¡ ¢
An individual trade > x involves a target volume V > x, and a fraction equal to P si ∆pβ > x
of the n (V ) ∼ V /∆pβ liquidity traders who meet his demand. So:
Z Z
P (qi > x) ∼ V −1−ζQ · n (V ) dV · s−1−ζ
i
L
dsi
V >x si ∆p(V )β >x
Z µ β·γ
¶ζL
V V
∼ V −1−ζV βγ dV
V >x V x
Z
= x−ζL V −ζQ −βγ+ζL ·β·γ
V >x
−ζQ −(ζL −1)(1−βγ)
∼

so that ζq = ζQ + (ζL − 1) (1 − βγ), which gives (63).

11 Appendix D: A model for the linear supply function (38)

We present a possible model equation (38), where a liquidity provider of size si gives qi for a price
concession ∆p. As we try to keep the theory modular, the sub-model is essentially independent of
the rest of the theory. Our model is a simple version of the inventory cost models of e.g. Grossman
and Miller (1988) and Stoll (1978). For simplicity, we drop the subscript i in this Appendix. We
use a very stark model to side-step the complexities identified by O’Hara (1995).
The liquidity provider has a net assets s, and a mean variance objective function on his returns
at some horizon38 H, with a zero riskless rate.
ρ
U = E [WH ] − varWH
2s

The factor ρ/s ensure that the risk taking behavior is independent of wealth, as U is homogenous
of degree 1 in (WH , s). If he holds A shares, with expected return m and volatility σ, his utility is:
³ ρ ´
U ∗ = W + H mA − σ 2 A2 (68)
2s
We will assume that he is allowed to supply the liquidity once before he can sell it again, costlessly,
to the market. On average it takes a time L to do this. We assume that he cannot rebalance his
portfolio in the meantime. Formally, we could model L as the average time before some other large
liquidity provider arrives, who would absorb liquidity costlessly. It could also be the amount of time
to sell the block at essentially no cost.
At the optimum, the liquidity provide holds A∗ share of the asset. By optimization of (68)
ms
A∗ = .
ρσ 2
If he supplies q shares to the trader, he will hold A∗ − q shares between times 0 and L and the
optimum quantity, A∗ , between times L and H. He pockets the price concession q∆p. So his utility
is: h i h i
ρ ρ
U ∗∗ = W + q∆p + L m (A∗ − q) − σ2 H (A∗ − q)2 + (H − L) mA∗ − σ 2 A∗2
2s 2s

3 8 Formally, this could be a Poisson time at which his portfolio is evaluated.

39
We assume that the large trader has full bargaining power, so that the liquidity provider must be
just indifferent between supplying the shares or keeping them. This happens if U ∗∗ = U ∗ , i.e.

∆p = ρσ 2 Lq/s (69)

So we get the supply function (38) with

1
α=
ρσ 2 L
This means that the liquidity, α, is higher when assets have low volatility σ, short resell time L, and
when the effective risk aversion ρ of the liquidity providers is low. Those are the traditional features
of inventory models.
The argument above holds much more generally. Costs of a suboptimal holdings are second
order, hence equal to Λq 2 for some constant Λ. The excess cash pocketed by the liquidity provider,
q∆p, has to compensate that, so that: q∆p = Λq 2 , and ∆p = Λq. The form ∆p = λsq comes from
the scale-invariance in s.

12 Appendix E: Algorithm for the simulations

In the period ∆t considered, there will be J such “trading rounds” where a big trader creates one or
0
more trades. Each round j has a sign εj = ±1, and gives rise to rj , Vj , nj , Vj0 , and nj . We define:

J
X J
X 1/2
r= rj = εj Vj (70)
j=1 j=1
J
X J
X
Q= Vj , N = nj
j=1 j=1
J
X J
X
Q0 = εj Vj , N 0 = εj nj
j=1 j=1

1/2
We have E [nj | Vj ] = Vj up to numerical prefactors. In reality, empirical returns the r above, plus
some noise due e.g. to news and other market events outside the model. However, those extraneous
events do not affect the large events and do not affect our graphs in major ways.
One simulates T time intervals indexed by t = 1, ..., T in the following way.

1. One draws an integer value Jt drawn at random, as the integer part of 10egt , with gt a standard
normal.
2. One draws vt = e.5ut and ηt = e.5vt with u, v i.i.d. standard normals. For j = 1...Jt , one
draws Qj with ζQ = 3/2, and εj = ±1 with probability 1/2, and sets:
1/2
rj = εj νt ηt−1 Qj (71)
1/2
Nj = νt ηt Qj (72)
Nj0 = εj Nj (73)

Note that νt , ηt are constant across the Jt rounds j of the time interval.
3. Applying (70), one gets a five values (r, Q, N, Q0 , N 0 ) which are stored as (rt , Qt , Nt , Q0t , Nt0 ) .

40
 One calculated the various E [Y | X] graphs on the data set made up of those values T vectors
(rt , Qt , Nt , Q0t , Nt0 ). To facilitate comparisons across stocks, one rescales the variables — demean
them, and divide them by their standard deviation if this standard deviation exists39 , otherwise by
their absolute mean.

Equations (71) and (72) are intended to be the analogues to the theoretical (44) and (46) resp.
−1/2
r and N vary like Q1/2 , but there are varying “market conditions” νt = (αs) (a measure of the
“depth” of each liquidity provider of the market at the given point in time), and ηt = µ1/2 (the
arrival rate of liquidity providers). While those quantities must have some randomness, we do not
have strong prior on the extent of their randomness, and any way they play only a minor role in
the model. So, largely arbitrarily, we chose make them equal to e.5u , with u a standard normal.
The .5 factor captures a relatively small randomness, but the results are largely insensitive to that
choice. Finally, the value of J around 10 was so that the average number of trades in a given half
hour match roughly the empirical one, without further effort to calibrate the model..

13 Appendix F: Confidence intervals and tests when a vari-

able has infinite variance
13.1 Construction of the confidence intervals for Figure 7
£ ¤
In a given bin conditioned by Q = Qi , with k elements r12 ,...,rk2 , the point estimate of E r2 | Q = Qi
is naturally the sample mean of the rj2 , which we call m. Getting a confidence interval for m is
delicate, as r2 has infinite variance, so the standard approach relying on asymptotic normality is
invalid. But the theory of self-normalizing sums of Logan et al. (1973) shows that if µ is the true
mean and σr2 is empirical standard deviation of the rj2 in the bin, then Student’s t = k 1/2 (m − µ) /σ
tends to a non-Gaussian distribution that depends only on the tail exponent of r2 , which is 2ζr = 3
in our case, and the fact that r2 is a positive variable. So a 95% confidence interval is mi ± ∆ri2
with ∆ri2 = χσri,j
2 k
−1/2
for some real number χ. Monte-Carlo simulations show χ = 3.23. Thus
we can use the standard procedure for confidence intervals, with a factor χ = 3.23 rather than the
Gaussian χ = 1.95.

13.2 Test of the linear relation E [r2 | Q] = α + βV

£ ¤
For each bin Qi of Q, we set ri2 = E r2 | Q = Qi , and ∆ri2 equal to the half-width of its 95%
£ ¤ ¡ ¢−2
confidence interval ri2 −£ ∆ri2 , r¤i2 + ∆ri2 . By weighted least squares, with weights ∆ri2 , we fit
an affine relationship E r2 | Q = g (Q), with

g (Q) = 0.90 + 0.31V

(0.057) (0.013)
R2 = 0.98

The standard errors are in parentheses. We find that for all values Qi ≥ 3, the predicted value g (Qi )
belongs to the 95% confidence interval:
£ ¤
g (Qi ) ∈ ri2 − ∆ri2 , ri2 + ∆ri2 .

3 9 So we took the first moment for V, r2 and V /N, and the second moment for N, N 0 , r and N 0 /N.

41
 £ ¤
We conclude that, at the 95% confidence level, we cannot reject the linear form E r2 | Q = g (Q)
for Q large enough, here Q ≥ 3.

References
[1] Ané, Thierry and Hélyette Geman, “Order Flow, Transaction Clock, and Normality of Asset
Returns,” Journal of Finance, 55 (2000), 2259-84.
[2] Andersen, Torben G., Tim Bollerslev, Francis Diebold, and Heiko Ebens, “The Distribution of
Realized Stock Return Volatility,” Journal of Financial Economics, 61(2001), 43-76.
[3] Axtell, Robert “Zipf Distribution of U.S. Firm Sizes,” Science, 293 (2001), 1818-20.
[4] Bak, Per, Ken Chen, José Scheinkman and Michael Woodford “Aggregate Fluctuations from
Independent Sectoral Shocks: Self-Organized Criticality in a Model of Production and Inventory
Dynamics,” Ricerche Economiche, 47 (1993), 3-30.
[5] Barberis, Nicholas and Richard Thaler “A Survey of Behavioral Finance”, in Handbook of the
Economics of Finance, eds. George Constantinides, and René Stulz (North-Holland: Amster-
dam, 2003).
[6] Biais, Bruno, Lawrence Glosten and Chester Spatt, “The Microstructure of Stock Markets,”
mimeo (2002).
[7] Bouchaud, Jean-Philippe and Marc Potters, Theory of Financial Risks: From Statistical Physics
to Risk Management (New York : Cambridge University Press, 2000).
[8] Breen, William, Laurie Hodrick and Robert Korajczyk, “Predicting Equity Liquidity,” Man-
agement Science 48 (2002) 470-483.
[9] Breiman, Leonard “On some limit theorems similar to the arc-sin law”, Theory of probability
and its applications, 10 (1965) 323-30.
[10] Burdett, Kenneth and Maureen O’Hara, “Building Blocks: An Introduction to Block Trading,”
Journal of Banking & Finance, 11 (1987), 193-212.
[11] Campell, John and Robert Shiller “The Dividend—Price Ratio and Expectations of Future
Dividends and Discount Factors”, Review of Financial Studies 1 (1998), 195—228.
[12] Canning, David, Luis Amaral, Y. Lee, Martin Meyer, and H. Eugene Stanley, “Scaling in the
Volatility of GDP Growth Rates,” Economics Letters, 60 (1998), 335-341.
[13] Carhart, Mark, “On Persistence in Mutual Funds Performance,” Journal of Finance, 52 (1997),
57-82.
[14] Chan, Louis and Josef Lakonishok, “Institutional Trades and Intraday Stock Price Behavior,”
Journal of Financial Economics, 33 (1993), 173-189.
[15] Chan, Louis and Josef Lakonishok, “The Behavior of Stock Prices Around Institutional Trades,”
The Journal of Finance, 50 (1995), 1147- 1174.
[16] Corsetti, Giancarlo, Paolo Pesenti and Nouriel Roubini, The Role of Large Players in Currency
Crises, in: S.Edwards and J. Frankel (eds.) Currency Crises in Emerging Markets: Crisis
Prevention, NBER and Chicago University Press, (2002).

42
[17] Coyne, Kevin and Jonathan Witter, “Taking the Mystery Out of Investor Behavior” Harvard
Business Review (2002).
[18] Clark, Peter, “A Subordinated Stochastic Process Model with Finite Variance for Speculative
Prices,” Econometrica, 41 (1973), 135-155.
[19] Cutler, David, James Poterba and Lawrence Summers “What moves stock prices?” Journal of
Portfolio Management 15, (1989) 4—12.
[20] Daniel, Kent, Daniel Hirshleifer and Avanidhar Subrahmanyam, “Investor Psychology and Se-
curity Market Under- and Over-reactions,” Journal of Finance, 53 (1998), 1839-1885.
[21] Duffie, Darrell, Nicolae Garleanu and Lasse Pedersen, "Securities Lending, Shorting, and Pric-
ing," Journal of Financial Economics, 66 (2002), 307-339.
[22] Easley, David, Soeren Hvidjkaer and Maureen O’Hara,“Is Information Risk a Determinant of
Asset Prices?”, forth. Journal of Finance (2003).
[23] Edelen, Roger and Jerold Warner, “Aggregate Price Effects of Institutional Trading: A Study of
Mutual Fund Flow and Market Return,” Journal of Financial Economics, 59 (2001), 195-220.
[24] Embrechts, Paul, Claudia Kluppelberg and Thomas Mikosch, Modelling Extremal Events
(Springer Verlag, New York, 1997).
[25] Evans, Martin and Richard Lyons, “Order Flow and Exchange Rate Dynamics”, Journal of
Political Economy, 110 (2002), 170-180.
[26] Fair, Ray, “Events that shook the market”, Journal of Business, 75 (2002), 713—731.
[27] Fama, Eugene, “Mandelbrot and the Stable Paretian Hypothesis” Journal of Business, 36
(1963), 420-429.
[28] Feenberg, Daniel and James Poterba, “Income Inequality and the Incomes of Very High-Income
Taxpayers: Evidence from Tax Returns,” in Tax Policy and the Economy, ed. James Poterba
(Cambridge, MA: MIT Press, 1993), 145-77.
[29] French, Kenneth and Richard Roll, “Stock Return Variance: The Arrival of Information and
the Reaction of Traders” Journal of Financial Economics 17 (1986), 99-117.
[30] Gabaix, Xavier, “Zipf’s Law for Cities: An Explanation,” Quarterly Journal of Economics 114
(1999), 739-67.
[31] Gabaix, Xavier, “Power Laws and the Origins of the Business Cycle,” mimeo (2003).
[32] Gabaix, Xavier, Parameswaran Gopikrishnan, Vasiliki Plerou and H. Eugene Stanley, “A theory
of power law distributions in financial market fluctuations,” Nature, 423 (2003a), 267—230.
[33] Gabaix, Xavier, Parameswaran Gopikrishnan and Vasiliki Plerou, and H. Eugene Stanley,
“Technical appendix” for the present paper, mimeo (2003b).
[34] Gabaix, Xavier, Parameswaran Gopikrishnan and Vasiliki Plerou, and H. Eugene Stanley, “To-
bin tax and large fluctuations in financial activity”, mimeo (2003c).
[35] Gabaix, Xavier, Parameswaran Gopikrishnan and Vasiliki Plerou, H. Eugene Stanley, “Are
stock market crashes outliers?”, mimeo (2003d)

43
[36] Gabaix, Xavier, Rita Ramalho and Jonathan Reuter (2003) “Power laws and mutual fund
dynamics”, MIT mimeo (2003).
[37] Gabaix, Xavier and Yannis Ioannides, The evolution of the city size distribution, in Handbook
of Urban and Regional Economics, 4 (2004) V. Henderson and J. Thisse ed., North Holland.
[38] Gnedenko, B. V., and A. N. Kolmogorov. Limit Distributions for Sums of Independent Random
Variables (Addison-Wesley, Reading, MA. 1968).
[39] Gopikrishnan, Parameswaran, Martin Meyer, Luis Amaral and H. Eugene Stanley “Inverse
Cubic Law for the Distribution of Stock Price Variations,” European Physical Journal B, 3
(1998), 139-140.
[40] Gopikrishnan, Parameswaran, Vasiliki Plerou, Luis Amaral, Martin Meyer and H. Eugene Stan-
ley “Scaling of the Distribution of Fluctuations of Financial Market Indices,” Physical Review
E, 60 (1999), 5305-5316.
[41] Gopikrishnan, Parameswaran, Vasiliki Plerou, Xavier Gabaix and H. Eugene Stanley “Statis-
tical Properties of Share Volume Traded in Financial Markets,” Physical Review E, 62 (2000),
R4493-R4496.
[42] Grossman, Sanford and Merton Miller “Liquidity and Market Structure,” Journal of Finance,
43 (1988), 617-33.
[43] Guillaume, Dominique M., Michel M. Dacorogna, Rakhal R. Davé, Ulrich A. Müller, Richard B.
Olsen and Olivier V. Pictet, “From the Bird’s Eye to the Microscope: a Survey of New Stylized
Facts of the Intra-Daily Foreign Exchange Markets,” Finance and Stochastics, 1 (1997), 95-129.
[44] Hasbrouck, Joel, “Trades, Quotes, Inventories, and Information,” Journal of Financial Eco-
nomics, 22 (1988), 229-252.
[45] Hasbrouck, Joel “Measuring the Information-Content of Stock Trades,” Journal of Finance, 46
(1991), 179-207.
[46] Hasbrouck, Joel and Duane Seppi, "Common Factors in Rpices, Order Flows and Liquidity",
Journal of Financial Economics, 59 (2001), 388-411.
[47] He, Hau and Jiang Wang, “Differential Information and Dynamic Behavior of Stock Trading
Volume,” Review of Financial Studies, 8 (1995), 919-972.
[48] Hirshleifer, David, “Investor psychology and asset pricing,” Journal of Finance, 56 (2001),
1533-1598.
[49] Holthausen, Robert, Richard Leftwich and David Mayers “Large-Block Transactions, the Speed
of Response, and Temporary and Permanent Stock-Price Effects,” Journal of Financial Eco-
nomics, 26 (1990), 71-95.
[50] Keim, Donald and Ananth Madhavan “The Upstairs Market for Large-Block Trades: Analysis
and Measurement of Price Effects,” Review of Financial Studies, 9 (1996), 1-36.
[51] Keim, Donald and Ananth Madhavan “Transaction Costs and Investment Style: An Inter-
Exchange Analysis of Institutional Equity Trades,” Journal of Financial Economics, 46 (1997),
265-292 .

44
[52] Koski, Jennifer and Roni Michaely, “Prices, Liquidity, and the Information Content of Trades,”
Review of Financial Studies, 13 (2000), 659-696.
[53] Kyle, Albert “Continuous Auctions and Insider Trading,” Econometrica, 53 (1985), 1315-1335.
[54] Lee, Charles and Mark Ready, “Inferring Trade Direction from Intraday Data,” Journal of
Finance, 46 (1991), 733-746.
[55] Lévy, Moshe, Heim Lévy and Sorin Solomon, Microscopic Simulation of Financial Markets:
From Investor Behavior to Market Phenomena (San Diego: Academic Press, 2000).
[56] Liu, Yanhui, Parameswaran Gopikrishnan, Pierre Cizeau, Martin Meyer, Chung-Kang Peng
and H. Eugene Stanley “Statistical properties of the volatility of price fluctuations,” Physical
Review E, 60 (1999), 1390—1400.
[57] Lillo, Fabrizio, Doyne Farmer and Rosario Mantegna, “Master curve for price-impact function”,
Nature, 421, (2003) 129-130.
[58] Lo, Andrew, Craig MacKinlay and June Zhang “Econometric models of limit-order executions,”
Journal of Financial Economics, 65 (2002), 31-72.
[59] Lo, Andrew and Jiang Wang, “Trading Volume: Definitions, Data Analysis, and Implications
of Portfolio Theory”, Review of Financial Studies, 13 (2001), 257-300.
[60] Logan, B.F., C. Mallows, S. Rice and L. Shepp “Limit distributions of self-normalized sums,”
Annals of Probability, 1 (1973), 788-809.
[61] Lux, Thomas, “The Stable Paretian Hypothesis and the Frequency of Large Returns: An
Examination of Major German Stocks,” Applied Financial Economics, 6 (1996), 463 - 475
[62] Lux, Thomas and Didier Sornette “On rational bubbles and fat tails,” Journal of Money Credit
and Banking, 34 (2002), 589-610.
[63] Madhavan, Ananth and Minder Cheng,“In Search of Liquidity: Block Trades in the Upstairs
and Downstairs Markets,” Review of Financial Studies, 10 (1997), 175—203.
[64] Madhavan, Ananth, Matthew Richardson and Mark Roomans “Why do Security Prices Fluc-
tuate? A Transaction-Level Analysis of NYSE Stocks,” Review of Financial Studies, 10 (1997),
1035-1064.
[65] Mandelbrot, Benoit, “The Variation of Certain Speculative Prices,”Journal of Business, 36
(1963), 394-419.
[66] Mantegna, Rosario and H. Eugene Stanley, “Scaling Behavior in the Dynamics of an Economic
Index,” Nature, 376 (1995), 46-49.
[67] Mantegna, Rosario and H. Eugene Stanley, An Introduction to Econophysics: Correlations and
Complexity in Finance(Cambridge: Cambridge University, 2000).
[68] Maslov, Sergei and Mark Mills, “Price Fluctuations from the Order Book Perspective: Empirical
Facts and a Simple Model,” Physica A, 299 (2001), 234-246.
[69] O’Hara, Maureen, Market Microstructure Theory (Cambridge, MA: Blackwell, 1995).
[70] Okuyama, K., M. Takayasu and Hideki Takayasu, “Zipf’s Law in Income Distribution of Com-
panies,” Physica A, 269 (1999), 125-131

45
[71] Piketty, Thomas and Emmanuel Saez, “Income Inequality in the United States, 1913-1998”,
Quarterly Journal of Economics, 118, (2003), 1-39.
[72] Plerou, Vasiliki, Parameswaran Gopikrishnan, Luis Amaral, Martin Meyer and H. Eugene Stan-
ley “Scaling of the Distribution of Fluctuations of Financial Market Indices,” Physical Review
E, 60 (1999), 6519-6529.
[73] Plerou, Vasiliki, Parameswaran Gopikrishnan, Luis Amaral, Xavier Gabaix and H. Eugene Stan-
ley “Economic Fluctuations and Anomalous Diffusion,” Physical Review E, 62 (2000), R3023-
R3026.
[74] Plerou, Vasiliki, Parameswaran Gopikrishnan, Xavier Gabaix, Luis Amaral and H. Eugene
Stanley, “Price Fluctuations, Market Activity, and Trading Volume,” Quantitative Finance, 1
(2001), 262-269.
[75] Plerou, Vasiliki, Parameswaran Gopikrishnan, Xavier Gabaix and H. Eugene Stanley, “Quan-
tifying Stock Price Response to Demand Fluctuations”, Physical Review E, 66 (2002) 027104.
[76] Plerou, Vasiliki, Parameswaran Gopikrishnan, Xavier Gabaix and H. Eugene Stanley, “Power
laws in finance: Evidence from the Paris Bourse”, mimeo (2003).
[77] Roll, Richard, “R2 ”, Journal of Finance, 43 (1988), 541-566.
[78] Saar, Gideon, “Price Impact Asymmetry of Block Trades: An Institutional Trading Explana-
tion,” Review of Financial Studies,14 (2001), 1153-81.
[79] Shiller, Robert, Market Volatility, (Cambridge, MA: MIT, 1989).
[80] Shleifer, Andrei “Do Demand Curves for Stocks Slope Down?” Journal of Finance 41 (1986),
579-90.
[81] Shleifer, Andrei, Inefficient Markets: an Introduction to Behavioral Finance (Oxford University,
2000).
[82] Shleifer, Andrei and Robert Vishny “The Limits of Arbitrage,” Journal of Finance, 52 (1997),
35-55.
[83] Solomon, Solomon and Peter Richmond “Power Laws of Wealth, Market Order Volumes and
Market Returns,” Physica A, 299 (2001), 188-197.
[84] Sornette, Didier, Critical Phenomena in Natural Sciences : Chaos, Fractals, Self-organization
and Disorder : Concepts and Tools, ( Heidelberg and New York: Springer 2000).
[85] Stoll, Hans, “The supply of dealer services in securities markets” Journal of Finance 33, (1978)
1133—1151.
[86] United States’ Presidential Task Force on Market Mechanisms, Report of the Presidential Task
Force on Market Mechanisms submitted to the President of the United States. (Washington,
D.C. Docs., U.S. G.P.O., 1988).
[87] Vayanos, Dimitri “Strategic Trading in a Dynamic Noisy Market,” Journal of Finance, 55
(2001), 131-171.
[88] Wang, Jiang, “A model of competitive stock trading volume”, Journal of Political Economy,
102 (1994), 127-168.
[89] Zhang, Yi-Cheng. “Toward a theory of marginally efficient markets”, Physica A, 269 (1999).

46