Video 3

Review of Economic Studies (2002) 69, 117–146 0034-6527/02/00050117$02.
00
ß 2002 The Review of Economic Studies Limited
An Optimal IPO Mechanism

BRUNO BIAIS
University of Toulouse
PETER BOSSAERTS
CALTECH
and
JEAN-CHARLES ROCHET
University of Toulouse
First version received March 1998; final version accepted July 2001 (Eds.)
We analyse the optimal Initial Public Offering (IPO) mechanism in a multidimensional

adverse selection setting where institutional investors have private information about the market
valuation of the shares, the intermediary has private information about the demand, and the
institutional investors and intermediary collude. Theorem 1 states that uniform pricing is
optimal (all agents pay the same price) and characterizes the IPO price in terms of conditional
expectations. Theorem 2 states that the optimal mechanism can be implemented by a non-linear
price schedule decreasing in the quantity allocated to retail investors. This is similar to IPO
procedures used in the U.K. and France. Relying on French IPO data we perform a GMM
structural estimation and test of the model. The price schedule is estimated and the conditions
characterizing the optimal mechanism are not rejected.
1. INTRODUCTION
Academic literature as well as the financial press have recently offered evidence consistent
with strategic or collusive behaviour between financial institutions participating to IPOs.
Non-competitive behaviour can arise between underwriters. For example, Chen and
Ritter (1999) document that in the U.S. at least 90% of deals raising between 20 and 80
million dollars have underwriting spreads exactly equal to 7% and relate this to the lack
of competition between investment bankers. Also based on U.S. data, Ellis, Michaely
and O’Hara (1998) find that underpricing or underwriting fees cannot be viewed as a way
to compensate investment banks for losses they would incur in stabilizing prices after the
issues. Indeed, they find that, while underwriters do play an active market making role,
this activity is profitable, especially when underpricing is large. Non-competitive beha-
viour can also arise in the relation between investment bankers and investors. Siconolfi
and McGheelan (1998) and Reimer (1998) describe situations where investment bankers
treat institutional investors and retail investors quite differently. Siconolfi and
McGheelan (1998) report evidence pointing at securities firms allowing big institutional
investors to sell back quickly the stocks they purchased in hot IPOs while deterring retail
investors from doing the same. Reimer (1998) discusses hot German IPOs where insti-
tutional investors received a disproportionately large fraction of the shares on sale. This
is consistent with the empirical finding by Cornelli and Goldreich (1999, p. 2) that
‘‘bidders who participate in many issues receive favourable treatment especially in the
117
118 REVIEW OF ECONOMIC STUDIES
more successful (i.e. oversubscribed) issues’’.1 Biased allocations, favouring institutional

investors in good deals, exacerbate the winner’s curse problem faced by small uninformed
investors, first analysed by Rock (1986). Note that it would be very surprising if
investment bankers were to favour large institutional investors, since they are engaged in
long-term repeated interactions, likely to give rise to tit-for-tat behaviour.2 An internet
based investment bank (WR Hambrecht+CO) has recently created a new IPO auction
mechanism: Openipo.com. The marketing stance taken by Openipo.com is indeed con-
sistent with potential collusion between investment bankers and large professional
investors in IPOs.3
Previous theoretical models of IPOs have often assumed that investment bankers act
in the best interest of their customer (the seller) and do not collude with institutional
investors. Yet, the above mentioned evidence suggests that it could be informative to
analyse the alternative case where investment bankers act non-competitively and where
collusive behaviour between financial institutions arises. The goal of the present paper is to
analyse the optimal design of IPO mechanisms under that alternative hypothesis.
In the book building IPO process, used in the U.S. or Germany, the allocation of
shares is entirely left at the discretion of the investment bankers. Assman (1998) under-
scored that this discretion facilitates priviledged allocations of shares by investment
bankers to large institutional investors. This suggests that, in presence of potential col-
lusion between investment bankers and institutional investors, the efficiency of the IPO
process could be improved by allowing the IPO mechanism to define more precisely the
allocation rule. Examples of IPO mechanisms where the allocation of shares is not left at
the discretion of the underwriter include the standard market clearing, uniform price
Walrasian auction, which is used in Israel (see Kandel, Sarig and Wohl, 1999) and pro-
posed by Openipo.com, as well as other types of uniform price auctions such as the British
offers by tender (see Brennan and Franks, (1995)) or the French Offre à Prix Minimum
(see Biais and Faugeron-Crouzet (2002), and Derrien and Womack (1998)).4 In the latter,
the firm sets a reservation price and investors submit bids. The net demand, resulting from
aggregating all the investors orders and the financial intermediary’s demand is submitted
to the auctioneer. Given this aggregate demand, the price is set by the auctioneer,
representing the Stock Exchange and the seller. The stronger the aggregate demand, the
higher above the reservation price the IPO price is adjusted. Yet, the price is deliberately
set below the market clearing level. The oversubscribed shares are allocated prorata. As
argued below, the optimal mechanism we derive can be interpreted as an abstraction of
this auction-like IPO method.
In spite of the potential conflict of interests between the seller and financial inter-
mediaries, the former could hardly dispense from the services of the latter. Industrial firms
going through IPOs very rarely are knowledgeable about financial markets, the willingness
of institutional investors to hold their shares, or where and how to contact potentially
1. Note however that Cornelli and Goldreich (1999) offer a different interpretation of their findings, which
they relate to the analyses of Benveniste and Spindt (1990) and Benveniste and Wilhelm (1991) where the financial
intermediary is assumed to act in the interest of the seller.
2. Hanley and Wilhelm (1995) analyse empirically allocations of shares by investment bankers to
institutional investors in the U.S. Their focus is very different from ours. Still, they provide evidence that the
investment bank and the institutional investors are engaged in a long term ‘‘tit for tat’’ relation.
3. For example it states (on its website: www.openipo.com) that: ‘‘Openipo treats a bid from an individual
the same as a bid from an institution’’ and that ‘‘IPO shares are allocated in an even-handed way by an auction
system rather than preferentially by investment bankers and brokers.’’
4. Consistent with the analysis proposed in the present paper, Derrien and Womack (1998) show
empirically that underpricing is larger for IPOs relying on the book building method than for the IPOs conducted
through the Offre a Prix Minimum auction.
BIAIS, BOSSAERTS & ROCHET OPTIMAL IPO MECHANISM 119
interested retail investors. IPO mechanisms must therefore be designed to protect the
interests of the selling firm, while eliciting information from the financial intermediaries
and professional investors about market conditions. In the present paper we take the
stance that the Stock Exchange designs the IPO mechanism in order to cope with these
problems. This is consistent with the observation that established exchanges such as the
Paris Bourse as well as new internet markets such as Openipo.com actually design and
operate auction-like IPO mechanisms.
To capture the above discussed stylized facts we analyse a theoretical model where
investment bankers have private information about the retail investors demand (because
they collect retail orders through their network and from their customers), institutional
investors have private information about their valuation of the asset, investment bankers
and institutional investors collude, and the IPO mechanisms is designed to maximize the
proceeds from sale. The main characteristics of the optimal mechanism are the following:
. The optimal mechanism is a simple schedule, specifying the IPO price as a function
of the quantity allocated to each uninformed retail investor. The optimal schedule
is decreasing, this enables some information extraction: the better the private signal
of the intermediary about the future market valuation of the shares, the more it
desires to purchase, the smaller the amount of shares left for the retail investor, and
hence the higher the price.
. In addition to eliciting information revelation, the fact that the price is decreasing
in the quantity allocated to the uninformed investors works against the winner’s
curse they face. When the market valuation of the stock is low and the
intermediary allocates a large quantity to the uninformed, the price is also set at
a relatively low level. In fact, we show that, in the optimal mechanism, the winner’s
curse is completely eliminated.
. Yet, in the optimal mechanism there is underpricing, reflecting the informational
rents earned by the informed agents.
. A priori, the general class of mechanisms we consider could allow for
discriminatory pricing, whereby the price at which the intermediary buys the
shares from the firm, would be different from the price at which the intermediary
sells to the retail investors. We show, however, that such discriminatory pricing is
not optimal.
From the methodological point of view, it should be stressed that our analysis delivers
a simple and intuitive solution to a technically difficult problem, involving mechanism
design in presence of multi-dimensional adverse selection, without parametric assumptions
about the distribution of random variables.5
Our theoretical analysis is in the line of Benveniste and Spindt (1989), Benveniste and
Wilhelm (1990) and Spatt and Srivastava (1991) to the extent that we analyse the IPO
selling methods as optimal mechanisms designed to elicit information from privately
informed investors. Indeed, as Benveniste and Wilhelm (1990) and Benveniste and Spindt
(1989), we show that underpricing is equal to the informational rent of these agents. In
contrast with these papers, however, we do not assume that the investment banker is
acting in the best interest of the seller. Our emphasis on the agency problem between the
seller and the intermediary is in the line of Baron (1982). Our result that discriminatory
5. Che and Gale (1999) analyse multi-dimensional mechanism design in the presence of liquidity
constraints.
pricing is not optimal contrasts with Benveniste and Wilhelm (1990) and reflects our
different assumptions.
On the empirical front, the optimal mechanism we characterize exhibits similarities
with the auction-like IPO procedures used in the U.K. (offer by tender) and in France
(offre à prix minimum), in particular the uniform price is increasing with aggregate
demand and decreasing with the quantity allocated to the retail investors. Because our
theoretical model generates moment conditions reflecting the optimality conditions of the
mechanism designer and the intermediary, and because it does not rely on parametric
assumptions on distributions, it lends itself nicely to econometric analysis. Using data
generated by 88 offres à prix minimum which took place between 1983 and 1996, we test
the hypothesis that this IPO procedure actually implements the optimal mechanism we
characterize. To execute a structural test we use Hansen’s (1982) Generalized Method of
Moments (GMM). The moment conditions implied by the optimality of the mechanism
and of the strategy of the intermediary are not rejected at the 10% level.
In the next section, the model is presented. In the third section we present our main
theorems, characterizing the optimal mechanism. In the fourth section, we offer con-
structive proofs of the theorems. In Section 5, the econometric analysis is presented. Section
6 presents a brief examination of the robustness of our theoretical model. Concluding
comments are in the last section. Proofs not given in the text are in the Appendix.
2. THE MODEL
2.1. Agents
Consider a firm selling a fixed amount of shares (Q) in an IPO.6 Without loss of generality
Q is normalized to 1. All agents are assumed to be risk neutral. The market valuation of
the shares, , is distributed over the bounded interval ½ ; . Two categories of investors
may participate in the auction for the shares: professional investors and retail investors. As
in Rock (1986), retail investors only know the distribution of , while professional
investors observe a private signal about it. For simplicity and without qualitative con-
sequences for the results, assume this signal is perfectly revealing.
There is a continuum of competitive, risk neutral retail investors. The total mass of
these investors is normalized to one. Each one of them is subject to a liquidity shock: he
has the funds to participate in the IPO (and can purchase up to d shares) with
probability 1=u.7 With the complementary probability, the retail investor is liquidity
constrained. He does not have the funds to bid in the IPO, and cannot purchase any
shares. Further, the random variables determining for each retail investor if he has the
funds to bid are i.i.d. Consequently, by the law of large numbers, the mass of retail
investors who have the funds to bid is 1=u. u has a simple interpretation: if only the retail
investors participated in the IPO, the quantity each one of them would receive would be u.
We assume that the objective function of the firm is to maximize the proceeds from
the sale.8 Also we assume that the firm cannot reach retail or professional investors
directly. It must use the services of an intermediary (an investment bank, a broker or a
6. In practice, the number of shares is indeed most of the time set a priori.
7. For reasons that will be apparent below it is more convenient to parametrize this probability as the
inverse of positive number greater than one.
8. Admittedly this assumption may be oversimplifying. Firms may pursue longer term goals when going
public, such as constructing a broad and maybe diffuse shareholder basis, (Brennan and Franks (1995) offer an
interesting empirical analysis of this aspect in the case of the U.K.), or signalling their long term profitability, (see
Allen and Faulhaber (1989) and Welch (1989)).
securities firm) who has developed a marketing and distribution network and thus
established links with potentially interested institutional and retail investors.9 We assume
that the intermediary, selected because of these links with investors interested in un-
seasoned shares sales, must inform all of them of the IPO.10 In this context, the inter-
mediary observes u.
The intermediary will have many further interactions with professional investors in
the future. In contrast, his future interactions with the firm are more limited (although the
firm might tap the market again in the future in seasoned offerings). Hence it is likely that
the intermediary will be tempted to favour the professional investors and cooperate with
them. In our model, for simplicity, we consider an extreme case of cooperation, and
consider only one aggregated agent: the coalition of the professional investors and the
intermediary. This coalition endeavours to maximize the profits it earns on the occasion of
the IPO, and to do so, uses the information of the intermediary about u and of the
investors about .
2.2. A benchmark model: Rock (1986)

First, to put our analysis in perspective, a simple version of Rock (1986) is presented. The
IPO is a fixed price offering. The firm sets the price p so as to maximize the proceeds from
sale, under the constraint the issue will be entirely sold. In this simple market mechanism,
the intermediary plays only a passive role.
Suppose the professional investors can demand at most I shares, and each retail
investor can demand d shares, when he has the funds to bid in the IPO. Assume also that
the number of retail investors who can bid in the auction for shares is deterministic and
known by all agents in the model, and finally that d > u, so that retail investors alone can
purchase the whole issue. The price must be such that the expected profit of retail investors
is non-negative.
If the market valuation of the shares is larger than p, then the IPO is a good deal and
the professional investors demand I shares. In contrast, if the market valuation of the
share is below , the professional investors do not participate in the auction. Conse-
quently, the total demand is larger when the IPO is a good deal, or, equivalently, unin-
formed investors get a disproportionately large share of the bad deals, i.e. retail investors
suffer from a winner’s curse problem. This is reflected in their expected profit:
Eðqð pÞÞ; ð1Þ
where q is the quantity allocated to the retail investors, which is equal to d=ðI þ d=uÞ if
> p and u if < p. Note that the covariance between and q is negative. Consequently
non negative expected profits for uninformed investors require: p < EðÞ. Hence the
winner’s curse problem generates underpricing.
2.3. Random retail demand

Although quite intuitive, this interpretation of underpricing is not fully satisfactory.
Indeed, in the framework of this simple version of Rock (1986), it is easy to construct an
optimal mechanism, fully avoiding winner’s curse and underpricing: if the total demand is
9. In practice there is often more than one intermediary intervening in the IPO. For simplicity, in our
model, we consider only one intermediary.
10. This can be legally enforced, in the same spirit as the due dilligence duty.
strictly larger than d=u, set p ¼ ; else set p ¼ EðÞ. This will deter strategic informed
investors from bidding in the auction, ensure that only small retail investors bid, and thus
eliminate asymmetric information induced winner’s curse problems. Of course we are not
satisfied by such a cheap and unrealistic way out. The point we raise here is rather that a
convincing model of asymmetric information in IPOs must involve some additional
complexity, preventing the above described mechanism from solving the winner’s curse
problem. To do so, we assume that the mass of retail investors who can participate in the
IPO is a random variable. Denote ½ u; u the support of u. Further, to ensure that the issue
could be sold to the retail investors, assume d > u . Now, the intermediary collects all the
orders from the retail investors. Hence he observes the level of the uninformed demand,
i.e. the realization of u. In contrast, the seller is a priori uninformed about u. He only
knows its distribution. The retail investors are in an intermediate situation. They are less
informed than the intermediary, since they do not observe the realization of u. But they are
more informed about u than the seller, since they can update their prior on u; conditionally
on whether they have the funds to bid. To write this updating process, consider a parti-
cular retail investor and denote Y the indicator variable taking the value 1 if the retail
investor has the funds to bid, and 0 otherwise. Denote f ðu; ; Y Þ the joint density of these
3 random variables, f ðu; Þ the joint density of u and , and f ðu; jY ¼ 1Þ the joint density
of u and conditional on Y ¼ 1. Then:
f ðu; ; Y ¼ 1Þ PrðY ¼ 1ju; Þ ð1=uÞ f ðu; Þ

f ðu; jY ¼ 1Þ ¼ ¼ f ðu; Þ ¼ : ð2Þ
PrðY ¼ 1Þ Eð1=uÞ Eð1=uÞ
Hence, for all random variables X depending on u and only:
d EðX=uÞ
E ðX Þ ¼ EðX jY ¼ 1Þ ¼ : ð3Þ
Eð1=uÞ
To match the above notation denote f ðu; Þ the density f ðu; jY ¼ 1Þ associated with
expectation E .
2.4. Mechanism design

We formulate our problem as the design of a hierarchical direct mechanism whereby:11
. First, each retail investor reveals to the intermediary his private information, i.e.
whether he has the funds to bid or not.
. Then the intermediary reveals to the seller the number of retail investors who have
the funds to bid in the auction (1=u), and the market valuation () of the shares.
. Finally the seller determines the price pðu; Þ of the issue, the quantity qðu; Þ
allocated to each retail investor, and the commission tðu; Þ paid to the
intermediary. Since we normalize the total number of shares sold to 1,
pðu; Þ tðu; Þ can also be interpreted as the unit price paid by the intermediary.
Therefore when tðu; Þ ¼ 0 the intermediary pays the same price as the retail
investor, i.e. there is uniform pricing.
The incentive compatibility conditions are the following:
. For the retail investors the only requirement is that they prefer to participate in the
11. Focusing on a direct mechanism is without loss of generality since the revelation principle applies here.
auction when they have the funds:
E ðð pðu; ÞÞqðu; ÞÞ >

¼ 0: ð4Þ
. For the intermediary the incentive compatibility condition is that his profit be
larger if he reveals the truth ðu; Þ than if he announces any other pair ðu^ ; ^ Þ. His
profit is the sum of two terms: the unit margin p multiplied by his net trade
1 q=u, and the commission t. Consequently the incentive compatibility condition
of the intermediary is:
8ðu; Þ; ðu; Þ 2 Argmax ½ð pðu^ ; ^ ÞÞð1 qðu^ ; ^ Þ=uÞ þ tðu^ ; ^ Þ: ð5Þ
ðu^ ;^Þ
Denote Bðu; Þ the profit of the intermediary when he truthfully announces ðu; Þ:
Bðu; Þ ¼ ð pðu; ÞÞð1 qðu; Þ=uÞ þ tðu; Þ: ð6Þ
Note that we do not constrain u q to be non-negative. That is, we assume that the
number of shares sold by the intermediary (and the professional investors) to the
retail investors can be larger than the number of shares sold by the firm. One
institutional justification for this assumption is that in Europe or Japan financial
institutions quite often own shares of the companies going public.12 For example,
in the case of France, for all offres à prix minimum which took place between 1983
and 1994, banks owned on average 12 8% of the capital, while financial
institutions, very often linked to banks, owned on average 5 78% of the capital.13
We also obtained more anecdotal evidence consistent with our assumption in
private conversation with a Portuguese investment banker specializing in IPOs. She
told us that investment banks often allocate more than the offering size, creating
short positions of the order of magnitude of 15% which they can cover by buying
shares in the secondary market.
The (ex post) rationality condition of the intermediary is that after observing ðu; Þ he
is still willing to participate in the IPO. Since we neglect the administrative costs incurred
by the intermediary the condition is:
8ðu; Þ; Bðu; Þ >

¼ 0: ð7Þ
Our mechanism design problem (P) amounts to finding the direct mechanism:
ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ; ð8Þ
maximizing the (net) expected revenue of the seller.
Eð pðu; Þ tðu; ÞÞ; ð9Þ
under the incentive compatibility and individual rationality constraints (4), (5) and (7).
12. Note however that in the U.S., until the recent reform of the regulation of the securities industry, banks
were not authorized to own shares of the firm going public.
13. These numbers are quite large relative to the percentage of the equity capital usually sold in those
IPOs: between 10 and 15%.
3. THE MAIN THEOREMS

The solution of this multidimensional mechanism design problem is presented in this
section, where we also discuss its economic implications. The following section offers a
constructive proof of the theorems. The transfers and allocations arising in the optimal
mechanism are characterized in our first theorem:
Theorem 1. P has a unique solution. Let ð p ð Þ; q ð Þ; t ð ÞÞ denote the optimal
mechanism and p^ ¼ p ðu; Þ; q^ ¼ q ðu; Þ denote the resulting price and rationing rate. They
are characterized by the following properties:
. The IPO price is uniform, i.e. all investors purchase shares at the same price, or
equivalently there is no direct transfer to the seller:
t ðu; Þ 0:
. The IPO price is equal to the expectation of the market valuation of the asset,
conditional on the quantity allocated to retail investors, where the expectation is
taken under the modified distribution (i.e. the expectation operator is E ):
E ðjq^ Þ ¼ p^ :
Therefore, in the optimal mechanism, the winner’s curse is eliminated, in the sense that for
any realization of p^ and q^ , retail investors obtain non-negative expected profits at the interim
stage.
E ðq^ð p^ Þjq^ Þ ¼ 0:
It is remarkable that, although there are two variables of adverse selection, namely
the number of retail investors and the market valuation of the asset, only one of the two
possible instruments ( pð Þ and tð Þ) is used at the optimum. Below, we show that this leads
to bunching, i.e. different types conducting identical trades, and relate this phenomenon to
the general theory of mechanism design under multidimensional adverse selection.
One of the most striking and prevalent stylized facts in IPOs is underpricing. In our
analysis, underpricing equals:
Eð p^ Þ ¼ E ½ð1 q^ =uÞð p^ Þ þ q^ ð p^ Þ=u:
The second term in the right-hand side is the unconditional expected profit of the retail
investors, which, because of the above theorem and the law of iterated expectations, equals
0. Consequently, underpricing is equal to the expected profit or informational rent of the
intermediary:
Eð p^ Þ ¼ Eðð1 q^ =uÞð p^ ÞÞ;
which is positive. Hence, underpricing is not due to winner’s curse effects à la Rock (1986).
Rather it corresponds to the informational rent of the informed agent, as in Benveniste
and Spindt (1989) and Benveniste and Wilhelm (1990). Correspondingly, to minimize
underpricing in order to maximize proceeds, the optimal IPO mechanism minimizes the
informational rent of the intermediary (as established more formally in the next section).
Underpricing, i.e. Eð p^ Þ < EðÞ might seem at odds with the conditional expectation
condition stated in Theorem 1. In fact there is no contradiction. The conditional expec-
tation condition in the theorem is computed under the modified probability distribution
(E ), while underpricing is computed under the original distribution (E ).
Having characterized the optimal direct mechanism, we now turn to its imple-
mentation:
Theorem 2. The optimal mechanism can be implemented by the following simple rule:
. The seller announces a decreasing price schedule whereby he commits to set the price
as a function of the rationing rate: P ðqÞ.
. The intermediary observes ðu; Þ and chooses the rationing rate q maximizing his
profit.
. The price schedule is such that: P ðq ðu; ÞÞ ¼ p ðu; Þ, where p is as defined in
Theorem 1.
The intuitive reason why the optimal price schedule is decreasing in the rationing rate
is the following. When the intermediary observes that the future market valuation of the
stock () is large, he wants to buy a large fraction of the shares. To achieve this, the
intermediary must ration the retail investors a lot. This is similar to the winner’s curse
highlighted in Rock (1986). To adjust pricing to the information content of this rationing,
the price is shifted upward. Note that, in addition to reflecting the informational content
of demand, this adjustment of the price eliminates the winner’s curse, since when the retail
investors are allocated a large quantity the price is relatively low.
To discuss this price adjustment further, consider the iso-profit curves of the informed
agent, in the case where, as stated in Theorem 2, the mechanism boils down to a simple
price schedule: p ¼ PðqÞ. Setting the informed profit to a positive constant K, the iso-profit
curve is given, in the plane ðq; pÞ, by the following equation:
K
p¼ ;
1 q=u
graphically represented in Figure 1. The figure can be partitioned in two regions corre-
sponding to sales and purchases by the informed agent. When he buys (i.e. when u > q), it
must be profitable for him to do so, hence: > p. Symmetrically when he sells, (i.e. when
u < q), < p. The figure also shows that there is a trade-off for the informed agent
between a decrease in p, making purchases more profitable, and an increase in q, reducing
the volume of purchases. The first-order condition of the informed agent reflects this price-
quantity trade off. It can be rewritten:
PðqÞ
P 0 ðqÞ ¼ : ð10Þ
uq
The R.H.S. of equation (10) is positive. Consequently, the optimal price function is
decreasing in the quantity allocated to each retail investor.
The first-order condition of the informed agent is illustrated in Figure 2, which shows
that the intermediary who has observed ðu; Þ chooses the quantity q and price p such that
his indifference curve is tangent to the price function at this point. Now, all types ðu; Þ
who choose the same point ðq; pÞ satisfy the first-order condition. Put in geometric terms,
the first-order condition (10) states that in the plane ðu; Þ the set of types for which the
first-order condition holds for ðq; pÞ, is the line that goes through ðq; pÞ and is symmetric to
the tangent of the price function at this point. The equation of this line is:
¼ PðqÞ P 0 ðqÞðu qÞ:

FIGURE 1
The indifference curve of the intermediary
It contains the set q of the types ðu; Þ who choose q. In simple cases (when two such lines
do not intersect inside the domain of ðu; Þ), q is actually equal to the whole line. When
some types ðu; Þ belong to several of these lines, the second-order condition has to be used
and q is a subset (an interval) of the line. In any case, ‘‘bunching’’ arises.
It is natural that, in a multi-dimensional adverse selection set-up where only one of
the two possible instruments ( p and t) is used, bunching should arise. Rochet and Chone
(1998) show that this is the case in general in multi-dimensional screening problems à la
Mussa-Rosen. They derive a new technique (‘‘measure-sweeping’’) for characterizing the
optimal solution of such problems. This technique could be used here for expressing the
first order condition of the programme of the mechanism designer. Without going into
details, this first-order condition can be stated as follows: let be the Lagrange multiplier
associated to the participation constraint of the informed agent. It is a positive measure,
supported by the set where Bðu; Þ ¼ 0, a curve which can be parametrized by
fu ¼ q; ¼ p ðqÞ; q 2 ½ u; u g. Now, recall that f ðu; Þ is the distribution associated with
the conditional expectation E . The sweeping condition states that, for a given mechanism
to be optimal, it is necessary and sufficient that the corresponding bunches satisfy the
following property: For all q, the distribution of f ð ; Þ on the ‘‘bunch’’ q is obtained
from the restriction of to q by a mean preserving spread. Since q only intersects the
FIGURE 2
The first order condition of the intermediary
curve Bðu; Þ ¼ 0 at ðq; P ðqÞÞ, the restriction of to q only charges ðq; P ðqÞÞ. Therefore
the sweeping condition means exactly that E ½ujq ¼ q and E ½jq ¼ p ðqÞ. The latter of
these conditions is the conditional expectation condition stated in Theorem 1, while the
former condition stems from the latter via the first-order condition of the informed agent.
Using this ‘‘sweeping’’ technique one can easily determine distributions of ðu; Þ for which
a given price function PðqÞ is optimal.
To illustrate our results we now present a simple example that we will also use in the
econometric analysis. Consider the following hyperbolic price function:

PðqÞ ¼ þ : ð11Þ
q
Figure 3 plots graphically the hyperbola (11) in the plane ðq; pÞ.14 Its focus is ð ; Þ, and its
asymptots are: x ¼ and y ¼ . Assume > and u > , so that the support of ðu; Þ is
above the focus.15 Superimposing the plane ðu; Þ on the plane ðq; pÞ, the bunches q which
would prevail for the price function (11) can be graphically represented. Because of the
properties of hyperbolas, these lines intersect at the focus. This is depicted in Figure 3. In
this context, using the sweeping technique discussed above, it is easy to construct a density
f for ðu; Þ such that (11) is the optimal price schedule. First choose a density function g
of total mass 1 on the hyperbola. Second, spread the weight gðqÞ put on point q along q
so that ðq; PðqÞÞ is the centre of gravity of this line. For all distributions constructed in this
way, (11) satisfies the conditions stated in Proposition (7).
The simple uniform price schedule characterized in Theorem 2 can be interpreted as
an abstraction of the IPO methods used in the U.K. (offer by tender) and in France, (offre
à prix minimum). Since prices are increasing in demand while execution rates are
14. To ensure that all subsets of the support of ðu; Þ have positive mass we assume that the hyperbola
includes the points ðu ; Þ and ðu; Þ.
15. This implies that the objective function of the intermediary is concave. Hence, there is no need to check
the second order condition.
FIGURE 3
The hyperbolic case
decreasing with demand, higher prices correspond to more severe rationing, as in the price
schedule described in Theorem 2. Anticipating the pricing rule, the intermediary can
determine its demand to select the rationing rate and price maximizing its profit.
4. A CONSTRUCTIVE PROOF OF THEOREMS 1 AND 2

Solving Problem P is somewhat delicate. To prove Theorems 1 and 2 we will first intro-
duce two auxiliary problems: P 1 , the dual of a relaxed version of problem P, whereby the
rationality constraint of the retail investors is not imposed and P 2 , a simpler problem,
pertaining to the characterization of the optimal uniform price auction.16 We will then use
the solution of these auxiliary problems to solve the initial problem P.
4.1. The dual problem

As shown by Mirrlees (1971), it is often convenient, when analysing the design of optimal
mechanisms, to analyse the dual problem, whereby the rent of the informed agent is
characterized, rather than the transfer and allocations scheme. In this spirit, we now
introduce problem P 1 , which is the dual of P (except that the rationality constraint of the
16. Readers who want to skip the most technical part of our analysis may focus on subsections 4.2 and 4.4
and skip 4.1 and 4.3.
retail investors is not imposed). In this dual formulation, the objective of the mechanism
designer is to minimize the expectation of the rent of the informed agent: EðBÞ (where B is
defined in equation (6)), subject to the incentive compatibility and individual rationality
constraints of the informed agent.
d
Define: Uðu; Þ ¼ uBð1=u; Þ; which can be interpreted as a rescaled version of the rent
of the informed agent: his profit per retail investor. Since B ¼ U=u, minimizing EðBÞ is
equivalent to minimizing EðU=uÞ or equivalently E ðUÞ. Consequently, P 1 can be written
as:
8
>
> min E ðUÞ

: The incentive compatibility condition of the informed agent is satisfied; and :
>
U> ¼ 0:
The following lemma characterizes the implementable rent functions, i.e. the
informed agent’s utility function U satisfying his incentive compatibility condition.
Lemma 3. U is implementable by a direct mechanism if and only if the function:
ðu; Þ ! Uðu; Þ uv is convex.
Proof. Let U be implemented by a direct mechanism:
ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ:
Rewriting the incentive compatibility condition of the intermediary (5) in terms of ðu; Þ,
we obtain:
Uðu; Þ ¼ max ½ð pðu^ ; ^ ÞÞðu qðu^ ; ^ ÞÞ þ utðu^ ; ^ Þ ð12Þ

ðu^ ;^Þ
where the maximum is attained for ðu^ ; ^ Þ ¼ ðu; Þ. Subtracting u from both sides of (12)
and rearranging terms:
Uðu; Þ u ¼ max ½ pðu^ ; ^ Þqðu^ ; ^ Þ þ uðtðu^ ; ^ Þ pðu^ ; ^ ÞÞ qðu^ ; ^ Þ: ð13Þ

ðu^ ;^Þ
Hence Uðu; Þ u is a maximum of affine functions of ðu; Þ. Consequently it is convex.

Conversely, by Fenchel’s duality theorem (see Rockafellar (1970), p. 327) any convex
function can be obtained in such a way for an adequate choice of the mechanism17:
ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ: jj
Using Lemma 3, problem P 1 can be rewritten as the following convex optimization

problem:
8
>
> min E ðUÞ

: U u convex
>
U>¼ 0:
17. To be complete, we should incorporate the condition qðu; Þ > ¼ 0, which is equivalent to ð@U=@Þ <
¼ u.
However this constraint will turn out not to be binding at the optimum.
Before solving P 1 (and then P), we must introduce a second auxiliary problem: the
characterization of the optimal uniform price IPO auction.
4.2. A simpler problem: the optimal uniform price auction

Consider the simple mechanism whereby the seller chooses a price schedule p ¼ PðqÞ and
let the intermediary choose the rationing rate maximizing his rent. For this simple
mechanism, the rescaled rent of the informed agent (U ), simplifies to:
Uðu; Þ ¼ max ð PðqÞÞðu qÞ:

q;p2PðqÞ
Define the functional B as follows:

8Pð Þ 2 G; BðPð ÞÞ ¼ E max ð pÞðu qÞ ð14Þ
q;p2PðqÞ
where G denotes the set of compact graph correspondences from ½ u; u to Rþ , endowed

with the Hausdorf topology. In fact, BðPÞ is the expected profit of the intermediary when
the seller chooses the price schedule Pð Þ:
BðPÞ ¼ E ½Uðu; Þ=u ¼ E ðU ÞEð1=uÞ:
Determining the optimal price schedule entails solving the following programme, denoted
P2:
P 2 min BðPÞ:
P2G
The functional B is continuous and G is compact (see, e.g. Aubin, 1979). Consequently B
has a minimum P on G. Hence we can state the following proposition:
Proposition 4. P 2 has (at least) a solution P ð Þ:
To establish existence we considered price correspondences rather than price func-
tions, because the set of continuous price functions is not closed for the Hausdorf
topology. Thus, a priori, the solution P of problem (P 2 ) can be multivalued. The next
proposition, however, states that this is not the case.
Proposition 5. The optimal price correspondence P , solution of P 2 , is a differentiable
function.
Proof. See Appendix. jj
The programme of the seller is to choose P to minimize the expected profit of the
intermediary B. We prove in the appendix that the optimum for the informed agent is for
almost every ðu; Þ obtained for a unique q, denoted q ðu; Þ. This implies that for any H
(measurable function of q) the directional derivative of B exists:
ð
BðP þ tHÞ BðP Þ
limþ ¼ Hðq ðu; ÞÞðu q ðu; ÞÞ f ðu; Þdud: ð15Þ
t!0 t
The optimality of P (first-order condition) requires that this limit be 0 for all measurable
H, which is equivalent to:
E ðu q ðu; Þjq Þ ¼ 0: ð16Þ
Consequently we can state the following proposition:

Proposition 6. Let P be the optimal price schedule, and q ðu; Þ the corresponding
optimal choice of the intermediary, then:
E ðu q ðu; Þjq Þ ¼ 0: ð17Þ
The intuition of condition (17) is the following: If, conditionally on q, the firm
expected the intermediary to buy, it could reduce his profit by raising the price. Similarly,
if it expected the intermediary to sell, the firm would prefer to lower the price. Hence, at
the optimum, where the firm neither wants to raise nor to lower the price, it must expect
the intermediary trade to be zero on average.
Building on the conditional expectation property (17), and on the first-order condi-
tion of the intermediary, the following proposition obtains:
Proposition 7. In the optimal price function, the allocation and price ðPðqÞ; qÞ are the
weighted average of the types ðu; Þ bunched in q :
E ðujqÞ ¼ q; ð18Þ
and
E ðjqÞ ¼ PðqÞ: ð19Þ
To establish this proposition note that the former condition (E ðujqÞ ¼ q) has already
been established in the previous proposition, while the latter condition (E ðjqÞ ¼ PðqÞ) is
a direct consequence of the former and the first-order condition:
PðqÞ ¼ P0 ðqÞðu qÞ:
Taking conditional expectations on both sides, we get indeed:

E ½ PðqÞjq ¼ P0 ðqÞE ½u qjq ¼ 0;
which is the desired result.

The geometric interpretation of Proposition 7 is that in the plane ðu; Þ, the point ðq; pÞ
is the center of gravity of the set q of all the types bunched at this point.
There is a formal similarity between our conditional expectation result (equation (19))
that under the modified distribution E the price is the expectation of the market valuation,
while it is not under the original distribution, and the result, in asset pricing, that under the
risk-adjusted distribution prices are martingales.
4.3. The solution of the initial problem

Having characterized the solution of the simple price schedule problem, P 2 , we can now
turn to the solution of the dual problem P 1 .
Proposition 8. Let U ðu; Þ ¼ maxq ð P ðqÞÞðu qÞ be the indirect utility function
associated to the optimal price schedule P . U solves (P 1 ).
Proof. Proposition 7 states that for the optimal price schedule:

E ðujq Þ ¼ q; E ðjq Þ ¼ P ðqÞ:
Therefore:
E ðU ðu; Þ ujq Þ ¼ P ðqÞE ½u qjq qE ½jq ¼ qP ðqÞ:
Consider now a candidate solution U of (P 1 ). Since U u is convex, by Jensen’s
inequality, for all q:
EðU ujq Þ >
¼ Uðq; P ðqÞÞ qP ðqÞ:
Since U is non negative, this implies, for all q:
E ðU ðu; Þjq Þ <
¼ E ðUðu; Þjq Þ;
and the proof is completed by taking expectations with respect to q. jj
We can now proceed to solve the initial problem P.

Proposition 9. Let:
ðu; Þ ! ð p ðu; Þ; q ðu; Þ; t ðu; Þ 0Þ

be the mechanism associated to U . It is the solution of P.
Proof. Let ð p; q; tÞ be any admissible mechanism and U the associated indirect
utility function. By Proposition 8, we know that:
E ðUÞ >
¼ E ðU Þ;
Moreover, for all ðu; Þ, (12) implies that:
U ¼ u ð pÞq uð p tÞ:
Therefore:
Eð p tÞ
¼ E ðuð p tÞÞ ¼ E ðu U ð pÞqÞ;
Eð1=uÞ
whereas:
Eð p t Þ
¼ E ðu U ð p Þq Þ: ð20Þ
Eð1=uÞ
Now, by Proposition 7:
E ðð p Þq Þ ¼ 0 <
¼ E ðð pÞqÞ:
Therefore:
Eð p t Þ Eð p tÞ > >
¼ Eð1=uÞ½E ðU Þ E ðU Þ ¼ 0;
which establishes the desired result. jj
4.4. Properties of the optimal price schedule

Theorem 2 and Proposition 8 have established that the optimal IPO mechanism could be
decentralized by a price schedule P . This optimal schedule is obtained by minimizing the
convex functional BðPÞ over the set of compact graph correspondences from ½ u; u to the
positive part of the real line. The functional B being strictly convex this minimum P is
unique and is completely determined by the joint density f ðu; Þ. We already noticed that
P was decreasing and characterized by moment conditions. Another property is that the
range of P is equal to the support of .
Proposition 10. P ð½ u; u Þ ¼ ½ ; .
One may also wonder how the optimal price schedule P is affected by changes in the
dispersion of the random variables u and . To examine this, start with the situation where
u and have unit variance, and denote P1 the corresponding optimal price schedule. Now
stretch the distributions of u and to obtain variances u2 and 2 , while preserving a
constant mean. Then it is easy to see by identification that the new optimal price schedule
is obtained by homothetic transformation:

q EðuÞ
PðqÞ ¼ P1 þ EðÞ:
u
0
In particular, the ‘‘sensitivity’’, P ðEðuÞÞ, of the price schedule at the average point EðuÞ is
proportional to the ratio of standard deviations:18

P0 ðEðuÞÞ ¼ P 01 ðEðuÞÞ:
u
5. ECONOMETRIC ANALYSIS
As mentioned above, the offre à prix minimum, an IPO procedure commonly used in
France, exhibits similarities with the price schedule, characterized in Theorem 2, which
implements the optimal mechanism described in Theorem 1. Now, Theorem 1 char-
acterizes the relation between prices and allocations in the optimal mechanism. In this
section, we test the null hypothesis that the data generated by 88 offres à prix mimimum
between 1983 and 1996 conform to this relation. This amounts to testing the joint
hypotheses that (i) asymmetric information and agency problems prevailing in IPOs in
France are as in our theoretical model, and (ii) confronted with these problems the Bourse
designs the IPO mechanism in the best interest of the firms going public. The rationale for
the latter is that by doing so the Bourse enhances the incentives of French firms to list on
the exchange.
Table 1 provides some summary statistics on the data. As can be seen in the table,
underpricing is on average equal to 15 34% while the rationing rate is on average 15%.
The next step to confront this data to the theoretical model is to delineate the relationship
between observed and theoretical variables.
. First consider the price. The IPO prices we observe in the data are not immediately
comparable. Foremost, the scaling differs across issues.19 To obtain IPO prices that
are comparable in cross-section, we scale the observed IPO price by the reservation
price.
. Second, consider the market valuation of the stock about which the intermediary
and the professional investors have private information. We take the equilibrium
price observed in the secondary market just after the IPO to be the variable
corresponding to this value. Again, for cross-sectional comparability, we normalize
it by the reservation price.
18. This is reminiscent of Kyle (1985).

19. To illustrate this, compare (i) the IPO of Digital Design, which took place in November 1987, at a price
of 80 Francs and for 123,786 shares, and (ii) the IPO of SOMICA, which took place in December 1986, at a price
of 495 Francs, for 64,000 shares.
TABLE 1
Summary statistics on 88 ‘‘mises en vente/offres à prix minimum’’ between 1983 and 1996
IPO Reservation Market clearing % Execution

price price price Underpricing rate
Average 285 236 330 15 34% 15%

Standard
deviation 165 25 131 95 196 37 21 9% 19 78%
Min 76 70 76 0 07 2%
Max 905 750 911 1 14 100%
. Third, consider the quantity allocated to retail investors, q. While we do not

observe it directly, we observe the rationing rate: , and in our model: q ¼ d.
While is observable, q is not, it is a parameter of the model, to be taken into
account in the econometric analysis. Note that for rescaling is not needed, since
this is a ratio (of total supply to admissible demand) and is therefore dimensionless.
Figure 4 provides a scatter plot of the rescaled value, , against the rescaled IPO
prices, p. The correlation is clearly positive, consistent with the view that the IPO price
reveals information about the market valuation of the stock. In the theory, this rela-
tionship is balanced to reflect the optimality conditions of the informed agent and the
mechanism designer. Whether this is the case in the data remains to be tested.
Figure 5 plots the IPO price against the rationing rate, . Overall, consistent with
theory, the lower the execution rate, the higher the IPO price tends to be. Yet, due to an
institutional constraint which prevailed until 1994, there is a cluster of points at ¼ 6%,
while for three IPOs which took place in 1995–1996 the execution rate is below 6%. This
makes it difficult to fit a reasonably smooth and parsimonious specification of the model.
Hence, for the observations clustered at ¼ 6%, we replace the actual value of the
execution rate, by a random draw from the uniform distribution over f1%; 2%; . . . ; 6%g.
Figure 6 provides a scatter plot of the pairs ð p; Þ, where is obtained after the above-
discussed smoothing procedure.
The first restriction imposed by the theoretical analysis on the data is equation (18):
Eð1 q=ujqÞ ¼ 0: In this moment condition, u plays the role of an incidental parameter.
Moreover, this condition only reflects the optimality of the price schedule given the
intermediary’s strategy. The optimality of the intermediary’s strategy can, however, be
incorporated, by substituting the corresponding first-order condition (equation (10)) for u.
One thereby eliminates the incidental parameter and imposes optimality of the inter-
mediary’s strategy. The result is:

P0 q
E 1 q ¼ 0:
p þ P0 q
Our econometric analysis will focus on this moment condition.

For simplicity, we carry the analysis for a simple parametrization of the price func-
tion, the hyperbolic function (11), which inspection of Figure 6 suggests would fit the data
rather well. This implies that our econometric analysis is not fully non-parametric. We
restrict the analysis to the class of joint distributions of u and such that the optimal price
shedule is hyperbolic. This class was characterized at the end of Section 3.
BIAIS, BOSSAERTS & ROCHET
OPTIMAL IPO MECHANISM
FIGURE 4
Rescaled value ðÞ plotted against rescaled IPO price ð pÞ, 88 mises en vente between 1983 and 1996
135
136
REVIEW OF ECONOMIC STUDIES
FIGURE 5
Rescaled IPO price ð pÞ plotted against execution rate 88 mises en vente between 1983 and 1996
FIGURE 6
Rescaled IPO price against smoothed allocation rate
Yet, Figure 6 also shows there is randomness in the relation between the execution rate
and the price. This reflects heterogeneity in the data, remaining after scaling the IPO price
by the reservation price. To model this phenomenon statistically, we allow the intercept
in the hyperbolic price function to vary across issues. That is:
¼ þ "; ð21Þ
where is a constant and Eð"Þ ¼ 0. For identification purposes we assume that the het-
erogeneity component is independent from the variable in the information set in the
moment condition: q.
By allowing heterogeneity, we effectively extend the probability space beyond that
generating the outcomes and u in the theoretical sections of this paper. Let E þ denote the
expectation taken with respect to the corresponding extension of the probabilities in the
original model. The moment condition imposed by the theory remains valid under E þ as
well. With the functional form in (11), this leads to the following moment condition:

þ 1
E q ¼ 0: ð22Þ
1 þ ½ q=ð pÞðq Þ2
Substituting d for q, this can be rewritten:

1
Eþ ¼ 0; ð23Þ
1 þ ½B =ð pÞð GÞ2
where B ¼ =d and G ¼ =d:
In addition, since " in equation (21) is independent of q, and using (11):

þ
E p q ¼ 0; ð24Þ
q
or:

þ B
E p ¼ 0; ð25Þ
G
Equations (22) and (24) together generate a set of moment conditions with which to
test the theory. They incorporate optimality of the pricing functions and of the strategy of
the intermediary, while simultaneously allowing for heterogeneity through the random
coefficient . GMM allows us to estimate the parameters by minimizing a quadratic form
in the sample versions of all the moment conditions and test the theory comprehensively
using a 2 statistic whose number of degrees of freedom equals the number of moment
conditions imposed minus the number of parameters being estimated. To operationalize
the conditional expectation conditions we chose as instruments the constant and . We
used them to build the following moment restrictions, whose sample versions enter the
GMM quadratic criterion:

þ 1
E ¼ 0;

1
Eþ ¼ 0;

þ B
E p ¼ 0;
G

B
Eþ p ¼ 0:
G
As always with instrumental variables estimation, the choice of instruments is rather
arbitrary and may very well drive the results. Poor instruments may lead to low power,
i.e. low probability of rejecting a false theory. Worse even, if enough poor instruments are
present, the GMM parameter estimator may not even be consistent. We were comforted in
our choice of instruments by the finding that the model was rejected for most parameter
choices. Only the ones given below (and values in their neighbourhood) lead to values of
the criterion function which remained inside the acceptance region.
With the above moment conditions, we estimated the three parameters B; G and .
Note that while we can identify B, which is equal to =d; and G, which is equal to =d , we
cannot identify separately ; ; and d. The results of the estimation are in Table 2. The
TABLE 2
GMM estimates
Parameter Estimate Standard error

B 0 66 2 08
G 0 02 6 15
1 04 0 12
2 1 45
p level 0 23
p value corresponding to the 2 statistic at the optimum is 0 23. Hence, the hypothesis that
the data conform to the restrictions imposed by the theory is not rejected at the 10% level.
Figure 7 plots the error " against the instrument . As requested by the moment
condition, there is little correlation between the two. In Figure 8 we draw the price
function corresponding to our parameter estimates. The standard errors for B and G
(reported in Table 2) are high. This is to be expected from the hyperbolic form of the
pricing function. The difficulty in estimating parameters of a hyperbolic function was also
reflected in the strong correlation between the estimates (not reported). Still note that our
structural approach to fitting the pricing relation outperforms the purely descriptive
approach based on non-linear least squares. Fitting prices against using non-linear least
squares leads to a 2 of 27 74, with a p-level of 0 001.
FIGURE 7
Fit of heterogeneity moment condition
6. ASSESSING THE ROBUSTNESS OF OUR THEORETICAL ANALYSIS

TO THE ASSUMPTION THAT u q CAN BE NEGATIVE
In our theoretical analysis we do not constrain u q to be non-negative, assuming that the
number of shares sold by the intermediary (and the professional investors) to the retail
investors can be larger than the number of shares sold by the firm. If we impose the non-
negativity constraint:
u q>
¼ 0;
FIGURE 8
Rescaled IPO prices and fitted price function
the optimal selling mechanism becomes much more complex to characterize, and we were
not able to solve this difficult mathematical problem in full generality. In this section, we
analyse a discrete example in which the non-negativity constraint is imposed. While in this
example the optimal mechanism is still characterized by a decreasing price schedule, it
must be complemented by transfers to the intermediary, in contrast with the case where
the non-negativity constraint is not imposed.
The simplest case which does not lead to a degenerate solution is that of a five point
distribution, as illustrated in Figure 9. The pair ðu; Þ can take 5 different values (denoted
A; B; C; D and E ) symmetrically located on the corners (ðu; Þ; ðu ; Þ; ðu ; Þ; ðu; Þ) and
centre (ðum ; m Þ) of the square ½ u; u ½ ; . Without loss of generality the square is nor-
malized to ½1; 22 . Denote the probability distribution adjusted for conditioning on
Y ¼ 1 (which was denoted f in the continuous case analysed above). We assume it
satisfies the symmetry condition: A ¼ C : In this case, if the non-negativity constraint is
not imposed and u q can be negative (as in the above sections), then the optimal IPO
mechanism is simply the decreasing price schedule corresponding to the second diagonal
of the square (see Figure 9). In this context, the informed coalition does not trade shares
for its own account in states B; D and E, while it sells 12 shares at price 32 in state A and buys
1 3
2 shares at price 2 in state D. Consequently the expected rent of the informed is:

1 A
ðA þ C Þ ¼ :
4 2
Now, consider the case where the non-negativity constraint is imposed, while keeping
the same price schedule. Constraining the informed coalition to u q >
¼ 0 prevents it from
FIGURE 9
Price schedule with and without the non-negativity constraint u q >
¼0
selling in state A. For that price the informed coalition must choose q ¼ 1 and p ¼ 2 in
state A. In this context, however, the participation constraint of the retail agents is no
longer satisfied, as they must buy the shares at an excessively high price in state A. Hence,
the IPO mechanism must be modified.
For simplicity we first consider the case where only price schedules can be used. The
main features of the optimal schedule are stated in the next proposition:
Proposition 11. If 34 E > A , when u is constrained to be greater than q; the optimal

uniform price mechanism yields rent:
3
4 E A
K¼ ;
2D þ 6E
to the informed in state A, and the price schedule is as depicted in Figure 9.
In this case the expected rent of the informed agent, A K, is positive (since
3
4 E> A ). Note also that this expected rent is lower than its counterpart arising when the
non-negativity constraint is imposed: A =2. Consistent with intuition, imposing this
constraint reduces the ability of the informed agent to retain rents, and correspondingly
reduces underpricing.
Now consider the case where, in addition, direct transfers to the intermediary can be
used. To reduce further the informed rent in state C, the mechanism designer can offer a
transfer t in that state. The profit of the informed agent if he truthully announces C,
thereby accepting to buy only q at price , and receiving the transfer t is:
q
ð Þ 1 þ t:
u
The transfer must be such that this is at least as large as the profit obtained when selecting
any other point on the schedule. Taking into account the incentive compatibility and
participation constraints, we characterize in the Appendix the optimal mechanism arising
in this context. Its main features are stated in the next proposition:
Proposition 12. In our simple example, when u is constrained to be greater than q; the
optimal mechanism yields rent:
3
1 4 E A
K ¼ ;
2 A þ D þ 3E
to the informed in state A.
Note that this rent is lower than the rent earned by the informed when the mechanism
does not involve transfers. To conclude, the optimal IPO mechanism arising when the
non-negativity constraint is imposed has the following properties:
. In contrast with the case where the non-negativity constraint is not imposed,
transfers are used and lead to lower informational rents.
. In line with the case where the non-negativity constraint is not imposed, the
optimal IPO mechanism involves a downward sloping demand schedule, as
illustrated in Figure 9.
In light of this examination of the robustness of our theoretical analysis, two remarks
can be made regarding our econometric investigation of data generated in the context of
offre à prix minimum: First, in practice, in offres à prix minimum, the intermediary
receives only quite limited fees, and these are fixed rather than contingent on the outcome
of the IPO. This fits our analysis of the case where u q >
¼ 0 is not imposed. Second, as the
qualitative features of the price schedule are similar irrespective of whether u q > ¼ 0 is
imposed or not, we cannot rely on visual inspection of the pricing functions to empirically
discriminate between the two cases. Rather we need to rely on quantitative econometric
work. From this perspective, our result in the previous section that the moment conditions
derived from the theoretical model (see Theorems 1 and 2) are not rejected by the data
does not suggest it is necessary to impose u q > ¼ 0:
7. CONCLUSION
We analyse an IPO setting whereby the coalition of the professional investors and the
intermediary has private information about the market valuation of the shares and the
demand of the retail investors, while the firm has no private information. We study the
optimal IPO mechanism by which the seller can extract private information to maximize
the expected net IPO proceeds. We find that:
. The optimal mechanism is equivalent to a price function, mapping the quantity
allocated to each retail investor into the IPO price.
. Uniform pricing is optimal, i.e. the coalition of the intermediary and the
professional investors pay the same price for the shares as the retail investors.
. The optimal price function is decreasing in the quantity allocated to each retail
investor. This serves two goals. First it generates information revelation by the
intermediary, because it implies that when the market valuation of the shares is
high the intermediary must admit it and pay a relatively large price, otherwise he
would have to leave too many shares to the uninformed. Second, it mitigates the
winner’s curse, since it implies that when the value is low and the intermediary
allocates many shares to the retail investors the price is low.
. Underpricing arises, but it is not driven by the Rock (1986) winner’s curse effect,
rather it corresponds to the necessity to leave an informational rent of the
intermediary.
. The optimal mechanism we characterize is similar to auction-like IPO procedures
used in the U.K. (offer by tender) and in France (offre à prix minimum).
Empirically, prices arising in the offre à prix minimum are indeed decreasing in the
rationing rate, as implied by our theoretical analysis. Using GMM, we test the
hypothesis that the data generated by the offre à prix minimum conforms to the
restrictions imposed by the optimality of the mechanism and of the strategy of the
informed agent. The theoretical restrictions are not rejected by the data.
APPENDIX: PROOF OF PROPOSITION 5

For clarity, the proof of the proposition is divided into the following four lemmas. The technique used here is very
similar to Rochet and Vila (1994).
Lemma 13. Let P be any price correspondence (with closed graph GrðPÞ) and
ðP; u; Þ ¼ max ð pÞðu qÞ:
ðq;pÞ2GrðPÞ
Then fðu; Þ=ðP; u; Þ ¼ 0g has a zero measure.

Proof. For a given u, ðP; u; Þ can only be equal to zero if, for all q > u and p 2 PðqÞ, p <
¼ : otherwise the
intermediary could make a profit by selling short some quantity ðq uÞ. Similarly it must be that for all q 
¼ . Therefore:
[ [
sup ¼<
PðqÞ < ¼ inf PðqÞ: ð26Þ
q>u q u2 g fq > u1 g ) PðqÞ PðqÞ:
q>u2 q>u1
þ
Hence p is decreasing. The proof for p is similar. Moreover if u1 < u2 , then for any q0 2u1 ; u2 ½:
[ [
sup PðqÞ >
¼ inf PðqÞ:
q>u1 q<u2
Therefore,
8u1 < u2 ; pþ ðu1 Þ >

¼ p ðu2 Þ:
Suppose now that pþ ðuÞ 
¼ limu2 !uþ p ðu2 Þ. Hence
pþ and p are discontinuous at u. Since decreasing functions are a.e. continuous, we have that, for a.e. u, the set
½ pþ ðuÞ; p ðuÞ is a singleton. But, by equation (26):
f=ðP; u; Þ ¼ 0g ½ pþ ðuÞ; p ðuÞ:
Therefore, by Fubini’s theorem: fðu; Þ; s:t ððP; u; Þ ¼ 0g has zero measure. jj
Lemma 14. Let p be a closed graph correspondence. Then:
Mðu; Þ ¼ argmax ð pÞðu qÞ

q;p2GrðPÞ
is a singleton for almost every ðu; Þ:
Proof.
ðP; u; Þ ¼ max fu up q þ pqg:
ðq;pÞ2GrðPÞ
Then one can write:

ðP; u; Þ ¼ u þ Gðu; Þ
where G is a convex function of ðu; Þ, as a supremum of affine functions.

By definition of G; for all ðu; Þ and ðq1 ; p1 Þ in GrðPÞ:
Gðu; Þ >
¼ p1 q1 up1 q1 :
Suppose that ðq1 ; p1 Þ 2 Mðu1 ; 1 Þ. Then the above inequality can be rewritten as:
Gðu; Þ >
¼ Gðu1 ; 1 Þ ðu u1 Þp1 ð 1 Þq1 ;
which means exactly (by definition of the subdifferential @Gðu1 ; 1 Þ (see Rockafellar (1972)):
!
p1
2 @Gðu1 ; 1 Þ:
q1
Since a convex function on a finite dimensional space is a.e. differentiable, @Gðu; Þ is a singleton for a.e.
ðu; Þ and Lemma 14 is established. jj
Lemma 15. Almost every seller is pooled with (at least) a buyer and vice-versa.
Proof. Equation (17) states that the average position of the types who choose any given quantity q is 0.
Except for those q which are chosen by a unique type (from Lemma 13 the set of such types is of 0 measure), there
is necessarily at least one buyer and one seller in the ‘‘pool’’. jj
Lemma 16. For almost every ðu; Þ the optimal price correspondence p is single valued, continuous and
differentiable at qðu; Þ.
Proof. Suppose w.l.o.g that ðu; Þ is a seller and take a buyer ðu1 ; 1 Þ who is pooled with ðu; Þ. Then:
(
qðu; Þ ¼ qðu1 ; 1 Þ ¼ q
with ð p ; u; Þg
and
2 ¼ fðq; pÞ=u1 q > 0 and ð1 pÞðu1 qÞ > ð p ; u1 ; 1 Þg:
It follows that p ð Þ has to be single valued, continuous and differentiable in q. jj
Proof of Proposition 10. The second moment condition, Eð j qÞ ¼ p, implies that the range of p is
included in ½ ; . Since p is decreasing, we have to establish conversely that p ðuÞ ¼ . (The proof that p ðu Þ ¼
is completely symmetric). Suppose by contradiction that p ðuÞ < , then, for all ðu; Þ in the triangle:
T ¼ fu >
¼ u; <
¼ ; pðuÞ >
0
¼ p ðuÞðu uÞg;
the profit of the intermediary of type ðu; Þ is maximum for q ¼ u (a corner solution). Moreover, no type ðu; Þ
below the graph of p chooses q ¼ u. This means that the set u of types that are bunched in q ¼ u coincides with
the triangle T (instead of being a line, like for other values of q). Since T lies entirely above the graph of p , this
contradicts the moment condition: Eð j qÞ ¼ p. jj
Proof of Proposition 11. For the same reasons as in our analysis in the above section, the optimal price
schedule is decreasing, since this facilitates information elicitation and at the same time mitigates the winner’s
curse. When the non-negativity constraint is imposed, a candidate transfer, eliminating all the rents captured by
the informed agent, would be the rectangular schedule: ðDCBÞ. In this case the price would be equal to its
maximum possible value: ¼ 2 in all states, except in B and C where the price would be set at its lowest possible
value: ¼ 1. Note also that, with such a schedule, the informed coalition would not be able to buy shares. Yet, if
3
4 E > A , this schedule violates the participation constraint of the retail investors. In this case, on which we will
focus hereafter, somewhat lower prices must be set in states A, D; and E. Recall that, as shown above, the iso-
profit curve of the informed investor in state ðu; Þ is:
u
pðqÞ ¼ K ;
uq
for the profit level K.

For incentive compatibility set the IPO schedule equal to:
u
pðqÞ ¼ K ;
uq
and to eliminate informed rents in all states but C, select K such that this iso-profit curve is above point E. To
minimize the fraction of the underpricing benefitting the informed agent in state C (and consequently minimize
the informed rent), select the pair ðq; pÞ corresponding to the largest possible q on the schedule
pðqÞ ¼ Kðu=ðu qÞÞ. This leads to q ¼ 2ð1 KÞ: In this context, saturating the individual rationality condition
of the retail investor,
E ðqð pÞÞ ¼ A ð1 pðuÞÞ þ C q þ E ðm pðum ÞÞ 32 þ D ð2 pðuÞÞ ¼ 0;
or:
A ð1 2KÞ þ E ð 12 4KÞ 32 ¼ D 2K þ C ð2 2KÞ;
which after simple manipulations pins down the rent of the informed coalition:
3
4 E A
K¼ ;
2D þ 6E
and correspondingly the price schedule, depicted in Figure 9. jj
Proof of Proposition 12. Saturating this incentive compatibility condition yields:

q
1 þ t ¼ K:
2
In this context, the participation constraint of the retail investor is:
A ð1 pðuÞÞ þ C q þ E ðm pðum ÞÞ 32 þ D ð2 pðuÞÞ >

¼ 0:
Saturating it and substituting in the price schedule yields:
A ð1 2KÞ þ E ð12 4KÞ 32 ¼ D 2K þ C q:
The mechanism design problem is to maximize the expected proceeds:
A pðuÞ þ B þ C ð tÞ þ D pðuÞ;
subject to the above characterized incentive compatibility condition of the informed agent and participation
constraint of the retail investor. After simple manipulations this leads to the optimal value of q:
q ¼ u ¼ 2;
and the corresponding minimized informed rent:
1 34 E A
K ¼ ; jj
2 A þ D þ 3E
REFERENCES
ALLEN, F. and FAULHABER, G. (1989), ‘‘Signaling by Underpricing in the IPO Market’’, Journal of Financial
Economics, 23, 303–323.
ASSMAN, H.-D. (1998), ‘‘Anleger tappen im Dunkeln’’, Wirtschaftwoche, NR 49/November 26, 171.
AUBIN, J. P. (1979), ‘‘Mathematical Methods of Games and Economic Theory’’ (Amsterdam: North-Holland).
BARON, X. X. (1982), ‘‘A Model of the Demand for Investment Banking Advising and Distribution Services for
New Issues’’, Journal of Finance, 37, 955–976.
BENVENISTE, L. M. and SPINDT, P. A. (1989), ‘‘How Investment Bankers Determine the Offer Price and
Allocation of New Issues’’, Journal of Financial Economics, 24, 343–362.
BENVENISTE, L. and WILHELM, W. (1990), ‘‘A Comparative Analysis of IPO Proceeds under Alternative
Regulatory Environments’’, Journal of Financial Economics, 28, 173–207.
BIAIS, B. and FAUGERON-CROUZET, A. M. (2002), ‘‘IPO Auctions: English, Dutch, . . . French and
Internet’’, Journal of Financial Intermediation (forthcoming).
BRENNAN, M. and FRANKS, J. (1995), ‘‘Underpricing, Ownership and Control in IPO of Equity Securities in
the U.K.’’ (Mimeo: London Business School).
CHE, Y. K. and GALE, I. (1999), ‘‘Mechanism Design with a Liquidity Constrained Buyer: The 2 by 2 Case’’,
European Economic Review, 43, 947–957.
CHEN, H. S. and RITTER, J. (1999), ‘‘The 7% Solution’’, Journal of Finance (forthcoming).
CORNELLI, F. and GOLDREICH, D. (1999), ‘‘Bookbuilding and Strategic Allocation’’ (Working paper:
London Business School).
DERRIEN, F. and WOMACK, K. (1998), ‘‘IPO Selling Procedures and the Control of IPO Underpricing in
Various Market Conditions: Evidence from the French IPO Market’’ (Working paper: Dartmouth College).
ELLIS, K., MICHAELY, R. and O’HARA, M. (1998), ‘‘When the Underwriter is the Market Maker: An
Examination of Trading in the IPO Market’’ (Working paper: Cornell University).
HANLEY, K. and WILHELM, W. (1995), ‘‘Evidence on the Strategic Allocation of IPO’’, Journal of Financial
Economics, 37, 239–257
HANSEN, L. P. (1982), ‘‘Large Sample Properties of General Method of Moments Estimates’’, Econometrica,
50, 1029–1054.
JACQUILLAT, B., MACDONALD, J. G. and ROLFO, J. (1978), ‘‘French Auctions of Common Stock New
Issues’’, Journal of Business Finance, 2, 305–322.
KANDEL, S., SARIG, O. and WOHL, A. (1999), ‘‘The Demand for Stocks: an Analysis of IPO Auctions’’,
Review of Financial Studies, 12, 227–248.
KYLE, A. S. (1985), ‘‘Continuous Auction and Insider Trading’’, Econometrica, 53, 1315–1336.
LELEUX, B. and PALIARD, R. (1993), ‘‘The Posted Price Paradox: Evidence on the Flotation Mechanism
Selection Process in France’’ (Working paper: Insead).
MIRRLEES, J. A. (1971), ‘‘An Exploration of the Theory of Optimum Income Taxation’’, Review of Economic
Studies, 38, 175–208.
REIMER, H. (1998), Das Kartell der Absahner, Wirtschaftwoche, 26(26), 160–173.
ROCHET, J. C. and CHONÉ, P. (1998), ‘‘Ironing, Sweeping and Multidimensional Screening’’, Econometrica,
(forthcoming).
ROCHET, J. C. and VILA, J. L. (1994), ‘‘Insider Trading without Normality’’, Review of Economic Studies, 61,
131–152.
ROCK, K. (1986), ‘‘Why New Issues are Underpriced’’, Journal of Financial Economics, 15, 187–212.
ROCKAFELLAR, T. (1972) Convex Analysis (Princeton: Princeton University Press).
SICONOLFI, M. and MCGHEELAN, P. (1998), ‘‘Wall Street Brokers Press Small Investors to Hold IPO
Shares’’, Wall Street Journal, 124, June 26.
SPATT C. and SRIVASTAVA, S. (1991), ‘‘Preplay Communication, Participation Restrictions, and Efficiency in
Initial Public Offerings’’, The Review of Financial Studies, 4, 709–726.
WELCH, I. (1989), ‘‘Seasoned Offerings, Imitation Costs and the Uderpricing of IPOs’’, Journal of Finance, 44,
412–449.

Video 3

Uploaded by

Copyright:

Available Formats

Video 3

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Video 3

Uploaded by

Copyright:

Available Formats

Review of Economic Studies (2002) 69, 117–146 0034-6527/02/00050117$02.

An Optimal IPO Mechanism

We analyse the optimal Initial Public Oﬀering (IPO) mechanism in a multidimensional

more successful (i.e. oversubscribed) issues’’.1 Biased allocations, favouring institutional

2.2. A benchmark model: Rock (1986)

Eðqð pÞÞ; ð1Þ

2.3. Random retail demand

f ðu; ; Y ¼ 1Þ PrðY ¼ 1ju; Þ ð1=uÞ f ðu; Þ

Hence, for all random variables X depending on u and  only:

2.4. Mechanism design

auction when they have the funds:

E ðð pðu; ÞÞqðu; ÞÞ >

Bðu; Þ ¼ ð pðu; ÞÞð1 qðu; Þ=uÞ þ tðu; Þ: ð6Þ

8ðu; Þ; Bðu; Þ >

ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ; ð8Þ

maximizing the (net) expected revenue of the seller.

Eð pðu; Þ tðu; ÞÞ; ð9Þ

3. THE MAIN THEOREMS

Eð p^ Þ ¼ E ½ð1 q^ =uÞð p^ Þ þ q^ ð p^ Þ=u:

Eð p^ Þ ¼ Eðð1 q^ =uÞð p^ ÞÞ;

 ¼ PðqÞ P 0 ðqÞðu qÞ:

4. A CONSTRUCTIVE PROOF OF THEOREMS 1 AND 2

4.1. The dual problem

Proof. Let U be implemented by a direct mechanism:

ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ:

Uðu; Þ ¼ max ½ð pðu^ ; ^ ÞÞðu qðu^ ; ^ ÞÞ þ utðu^ ; ^ Þ ð12Þ

Uðu; Þ u ¼ max ½ pðu^ ; ^ Þqðu^ ; ^ Þ þ uðtðu^ ; ^ Þ pðu^ ; ^ ÞÞ qðu^ ; ^ Þ: ð13Þ

Hence Uðu; Þ u is a maximum of aﬃne functions of ðu; Þ. Consequently it is convex.

ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ: jj

Using Lemma 3, problem P 1 can be rewritten as the following convex optimization

4.2. A simpler problem: the optimal uniform price auction

Uðu; Þ ¼ max ð PðqÞÞðu qÞ:

Deﬁne the functional B as follows:

where G denotes the set of compact graph correspondences from ½ u; u to Rþ , endowed

Proof. See Appendix. jj

Consequently we can state the following proposition:

E ðu q ðu; Þjq Þ ¼ 0: ð17Þ

Taking conditional expectations on both sides, we get indeed:

which is the desired result.

4.3. The solution of the initial problem

Proof. Proposition 7 states that for the optimal price schedule:

We can now proceed to solve the initial problem P.

ðu; Þ ! ð p ðu; Þ; q ðu; Þ; t ðu; Þ  0Þ

which establishes the desired result. jj

4.4. Properties of the optimal price schedule

18. This is reminiscent of Kyle (1985).

IPO Reservation Market clearing % Execution

Average 285 236 330 15 34% 15%

. Third, consider the quantity allocated to retail investors, q. While we do not

Our econometric analysis will focus on this moment condition.

In addition, since " in equation (21) is independent of q, and using (11):

Parameter Estimate Standard error

6. ASSESSING THE ROBUSTNESS OF OUR THEORETICAL ANALYSIS

Proposition 11. If 34 E > A , when u is constrained to be greater than q; the optimal

to the informed in state A, and the price schedule is as depicted in Figure 9.

APPENDIX: PROOF OF PROPOSITION 5

Then fðu; Þ=ðP; u; Þ ¼ 0g has a zero measure.

8u1 < u2 ; pþ ðu1 Þ >

f=ðP; u; Þ ¼ 0g  ½ pþ ðuÞ; p ðuÞ:

Eðqð pÞÞ; ð1Þ

f ðu; ; Y ¼ 1Þ PrðY ¼ 1ju; Þ ð1=uÞ f ðu; Þ

Hence, for all random variables X depending on u and only:

E ðð pðu; ÞÞqðu; ÞÞ >

Bðu; Þ ¼ ð pðu; ÞÞð1 qðu; Þ=uÞ þ tðu; Þ: ð6Þ

8ðu; Þ; Bðu; Þ >

ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ; ð8Þ

Eð pðu; Þ tðu; ÞÞ; ð9Þ

Eð p^ Þ ¼ E ½ð1 q^ =uÞð p^ Þ þ q^ ð p^ Þ=u:

Eð p^ Þ ¼ Eðð1 q^ =uÞð p^ ÞÞ;

¼ PðqÞ P 0 ðqÞðu qÞ:

ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ:

Uðu; Þ ¼ max ½ð pðu^ ; ^ ÞÞðu qðu^ ; ^ ÞÞ þ utðu^ ; ^ Þ ð12Þ

Uðu; Þ u ¼ max ½ pðu^ ; ^ Þqðu^ ; ^ Þ þ uðtðu^ ; ^ Þ pðu^ ; ^ ÞÞ qðu^ ; ^ Þ: ð13Þ

Hence Uðu; Þ u is a maximum of aﬃne functions of ðu; Þ. Consequently it is convex.

ðu; Þ ! ð pðu; Þ; qðu; Þ; tðu; ÞÞ: jj

Uðu; Þ ¼ max ð PðqÞÞðu qÞ:

E ðu q ðu; Þjq Þ ¼ 0: ð17Þ

ðu; Þ ! ð p ðu; Þ; q ðu; Þ; t ðu; Þ 0Þ

Proposition 11. If 34 E > A , when u is constrained to be greater than q; the optimal

Then fðu; Þ=ðP; u; Þ ¼ 0g has a zero measure.

f=ðP; u; Þ ¼ 0g ½ pþ ðuÞ; p ðuÞ:

Mðu; Þ ¼ argmax ð pÞðu qÞ

is a singleton for almost every ðu; Þ:

where G is a convex function of ðu; Þ, as a supremum of aﬃne functions.

E ðqð pÞÞ ¼ A ð1 pðuÞÞ þ C q þ E ðm pðum ÞÞ 32 þ D ð2 pðuÞÞ ¼ 0;

A ð1 2KÞ þ E ð 12 4KÞ 32 ¼ D 2K þ C ð2 2KÞ;

A ð1 pðuÞÞ þ C q þ E ðm pðum ÞÞ 32 þ D ð2 pðuÞÞ >

A ð1 2KÞ þ E ð12 4KÞ 32 ¼ D 2K þ C q: