Is Best Really Better?Internalization in Xetra Best

Is Best Really Better?

Internalization in Xetra Best

CFR Working Paper No. 05-06

Joachim Grammig and Erik Theissen

Is Best Really Better?

Internalization in Xetra BEST

^*

Joachim Grammig and Erik Theissen^**

March 2005

Abstract: Xetra BEST, operated by Deutsche Börse AG as a part of the Xetra trading system, allows participating banks and brokers to internalize retail customer orders. This paper provides an empirical assessment of the market quality of Xetra BEST. For this purpose, we develop a trade indicator model of this dual market structure with an open order book and an internalization system that guarantees a mandatory price improvement over the inside spread of the book. We show that internalization is profitable, and that adverse selection costs of internalized trades are significantly lower than those of regular order book trades. However, the mandatory price improvement for internalized trades does not fully compensate for the difference in adverse selection costs. Our results imply that measures aimed at increasing competition among internalizers may yield improved execution quality for internalized orders.

JEL classification: G10

Keywords: Internalization, execution quality, adverse selection costs

* We thank seminar participants at the University of Frankfurt/Main, at KU Leuven, at CORE/Louvain-La-Neuve and at Humboldt University Berlin for their helpful comments and Oliver Wünsche for research assistance. We are grateful to Deutsche Börse AG for providing access to the data. Uwe Schweickert offered invaluable expertise in the trading systems and support for our work.

** Joachim Grammig, University of Tübingen, Department of Economics, Mohlstr. 36. 72074 Tübingen, Germany, email:

joachim.grammig@uni-tuebingen.de, phone: +49 7071 2976009, fax: +49 7071 295546; Erik Theissen, University of Bonn, BWL I, Adenauerallee 24-42, 53113 Bonn, Germany, email: theissen@uni-bonn.de phone: +49 228 739208, fax:

+49 228 735924.

1 Introduction

Internalization is a practice by which banks or brokers receiving (retail) customer orders do not route these orders to the exchange but rather execute them against their own book.

Internalization is a controversial issue in the European Union. The approaches to internalization in the member states differ widely. France and Italy have adapted a

"concentration rule" which requires that all transactions be processed on a regulated market.

Such a rule works to prohibit internalization. The UK is on the opposite end of the spectrum.

Hansch, Naik and Viswanathan (1999) report that 62% of the trades in their sample are internalized.

In 2002 the commission of the European Union published a draft paper for the Investment Services Directive (ISD) that took a very liberal approach to internalization. A hefty debate ensued. The final version of the ISD, published in the Official Journal of the European Union in April 2004, takes a somewhat more restrictive approach. It allows internalization but requires internalizers to "make public their quotes on a regular and continuous basis during normal trading hours" (Article 27).

EU member countries are obliged to change national legislation in order to comply with the ISD. Obviously, this will lead to considerable change. Exchanges in countries that currently prohibit or restrict internalization will have to deal with the competitive pressure imposed by the threat of internalization.

Deutsche Börse AG faced a similar threat at the beginning of the decade. In Germany, there was no internalization of retail order flow at a larger scale although there were no legal restrictions preventing brokers from internalizing. It became obvious, however, that some

large banks were planning to embark on internalization. Deutsche Börse AG reacted to that threat by implementing Xetra BEST. This system, operated as a part of Xetra, allows banks and brokers to internalize retail order flow on a platform operated by Deutsche Börse AG. It is designed such as to limit the potentially adverse effects of internalization on the quality of the main market. Such a system, if successful, may serve as a blueprint for other exchanges facing the threat of internalization.

The objective of the present paper is to provide an assessment of Xetra BEST. Besides providing a detailed description of the system we empirically analyze the execution quality in Xetra BEST and the components of the spread. Internalized orders are typically small orders submitted by retail customers. They are thus likely to be uninformed (Chakravarty, 2001 and Linnainmaa, 2003). For this reason it has been argued that internalizers "skim the cream" of the order flow, leaving only the less attractive orders for the main market (e.g., Easley, Kiefer and O'Hara, 1996). We test the cream skimming hypothesis by comparing the adverse selection costs of internalized orders to those of orders executed on the main market. To this end we develop a structural model of the specific dual market structure under scrutiny. The model builds on the trade indicator model of Glosten and Harris (1988).

Our results provide clear evidence of cream skimming. Realized spreads are significantly higher for orders executed in Xetra BEST than for orders executed in Xetra. The results of the trade indicator model indicate that adverse selection costs in Xetra BEST are significantly lower than those on the main market. It is for this reason that internalizers earn realized spreads that are larger than those earned by the suppliers of liquidity on the main market.

These results have important implications for regulation and market design. As internalizers appear to be earning rents, measures that enhance competition among internalizers are

warranted. The requirement in the ISD that internalizers have to publish their quotes on a regular and continuous basis is a step in that direction.

Our paper is related to previous papers analyzing internalization and the related phenomena of preferencing and payment for order flow.¹ On balance, these papers conclude that market quality is not negatively affected by internalization. There appears to be, however, some evidence of cream skimming, as documented in Easley, Kiefer and O'Hara (1996). Our paper contributes to this literature in several ways. Firstly, we analyze the effects of internalization in a fully electronic auction market. This contrasts with previous papers analyzing the floor- based specialist system of the NYSE or the dealer markets of Nasdaq and the London Stock Exchange. Secondly, we analyze internalization on a system that is operated by the exchange itself. This is potentially important because Deutsche Börse has implemented rules aiming at mitigating the potentially adverse effects of internalization on market quality. Whether these rules are effective is a question addressed in our paper.

The remainder of the paper is organized as follows. In section 2 we describe Xetra BEST in detail. In section 3 we describe our dataset and present descriptive statistics. The empirical

1 Battalio, Jennings and Selway (2001), Hansch, Naik and Viswanathan (1999) and Peterson and Sirri (2002) compare execution costs for preferenced and non-preferenced orders. Battalio (1997) and Battalio / Greene / Jennings (1997) analyze whether the execution quality of the main market was affected by the introduction of preferencing and payment for order flow arrangements. Easley, Kiefer and O'Hara (1996) compare the probability of informed trades on two exchanges known to execute a large fraction of purchased order flow to the probability of informed trading on the NYSE. Bloomfield and O'Hara (1998) analyze the effect of preferencing in an experimental setting. Battalio and Holden (2001), Kandel and Marx (1999) and Parlour and Rajan (2001) develop theoretical models of payment for order flow arrangements.

methodology and the main results are contained in section 4. Section 5 discusses the implications and concludes.

2 Market Structure

Xetra BEST was introduced in September 2002 and is a part of the trading system Xetra, an electronic open limit order book. Brokers (which have to be Xetra members to qualify) may become "best executors" in Xetra BEST.² This gives them the right to execute qualifying customer orders against their own account using Xetra BEST as a trading platform. At the same time, however, they are obliged to act as "liquidity managers" in the Xetra order book, i.e., they must contribute to the liquidity of the order book by submitting limit orders.³

Only market orders and marketable limit orders are eligible for execution in Xetra BEST.

Further, order execution in Xetra BEST requires consent of the customer. Orders executed in Xetra BEST receive price improvement. The execution price is at least one cent per share (the minimum price increment in Xetra) better than the reference price. The reference price, in turn, is the price at which the order would execute if it was submitted to Xetra instead. For orders not exceeding the depth at the best bid or ask, the reference price is the best bid or ask.

For orders exceeding the depth at the best quotes the reference price is calculated as the volume-weighted average of the relevant limit orders in the book.

The best executor specifies the maximum volume he is willing to trade on either side of the market and the amount of price improvement she grants and enters the corresponding values

2 See Gomber and Maurer (2004) for a detailed description of the Xetra BEST system.

3 Whether this constitutes a binding restriction is not clear, however, as there are no minimum spread or similar requirements.

into the system.⁴ When a Xetra BEST-eligible order is submitted, the system checks whether the size of the order exceeds the maximum volume specified by the best executor. It also checks whether the price improvement would result in a zero or negative spread. In both cases the order will not be executed via Xetra BEST but is automatically routed to the Xetra order book. The same is true when the order is submitted during a call auction in Xetra.⁵

Xetra BEST is fully post-trade transparent. Transactions are reported immediately. They are marked as "XB". This allows market participants to distinguish Xetra BEST transactions from regular order book trades. Xetra BEST allows for order flow provision agreements. Under such an agreement, a broker who is not himself a best executor may route eligible customer orders to a best executor. In turn, he may receive a payment.⁶ Obviously, such an arrangement shares many similarities with the practice of payment for order flow in the US.

Deutsche Börse AG claims that the design of the system enforces price and time priority across Xetra and Xetra BEST. This is, however, not exactly true. Considering the relation between the Xetra order book and Xetra BEST, price priority is enforced and time priority is irrelevant because of the mandatory price improvement. Considering the relation between several best executors, however, neither price nor time priority is enforced. Enforcement would require that an incoming order be routed to the best executor offering the largest price improvement or, in case of equal price improvements, would be routed to the best executor

4 In principle a best executor can choose to offer price improvements larger than 1 cent.

5 There is an opening call auction, an intradaily call auction around 1 pm (the exact time differs by categories of stocks) and a closing call auction.

6 Unfortunately, there is no information available as to whether such payments actually exist.

who was first to offer the specific amount of price improvement. This order routing across best executors does, however, not occur.

Internalization, if practiced on a larger scale, may adversely affect the quality of the main market (see Biais and Davydoff, 2002 and the literature addressing the cream skimming argument cited earlier). Deutsche Börse AG has deliberately designed Xetra BEST such as to reduce these adverse effects. Requiring the best executors to supply liquidity to the order book may alleviate the reduction in liquidity potentially caused by internalization and the fragmentation of markets it entails. The requirement may prove to be a tissue tiger, however, because the obligations of a liquidity manager are neither exactly specified nor enforced.

The mandatory price improvement obviously provides for better execution quality when compared to a system in which it is only required that the internalizer match the best quotes available in the main market. Consequently, a customer wishing to submit a market order will rationally choose execution in Xetra BEST over execution in Xetra. Furthermore, the price improvement rule serves to increase the relative attractiveness of market orders as compared to limit orders. The price improvement can also be interpreted as a compensation for the lower adverse selection risk associated with the subset of the order flow eligible for execution in Xetra BEST. Whether the price improvement fully accounts for the difference in adverse selection costs is a question addressed in our empirical analysis.

3 Data

Our data set covers eight stocks and the period October 2002 through January 2003, a total of 84 trading days. The sample stocks are German blue chip stocks, and they are all among the

constituent stocks of the DAX index. For each of the sample stocks we have a complete record of all transactions in Xetra and Xetra BEST. If a market order walks up or down the book, the resulting transactions are recorded as one transaction at the volume-weighted average price.

Besides the transaction price the data includes the exact time of the trade, the volume of the trade, an indication whether the trade was buyer- or seller-initiated,⁷ and an indicator identifying trades in Xetra BEST. The data set does not contain best bid and ask prices.

Table I presents some descriptive statistics. The figures confirm that the sample stocks are indeed highly liquid. Even the least liquid stock has more than 70,000 transactions, or roughly 860 transactions each day. The market share of Xetra BEST is rather low. When considering the number of transactions, it amounts to 1.5% on average. Transactions in Xetra BEST are much smaller than regular order book trades. Therefore, when considering the Euro trading volume, the Xetra BEST market share is only about 0.25%.

Insert Table I about here

Given the small market share of Xetra BEST it is unlikely that the introduction of the system has measurable implications for the liquidity of the main market. That does by no means render our analysis irrelevant, however, as we compare transactions in Xetra and in Xetra BEST with respect to the quality of execution, the adverse selection cost and the realized spread earned by the supplyers of liquidity. We do not intend to analyze whether the quality of the main market has changed upon introduction of Xetra BEST.

7 This information was provided by the exchange. We therefore did not have to apply a trade classification algorithm.

4 Methodology and Results

A variety of procedures to estimate the spread and its components has been proposed in the literature. We employ two different approaches. We first estimate realized spreads to measure the revenue earned by the suppliers of liquidity. We then use a trade indicator model to estimate the spread components and compare them between the two trading systems.

Specifically, we use the Glosten and Harris (1988) approach and modify it to account for the specific dual market structure under scrutiny. We also estimate a restricted version of the Glosten-Harris model which is similar to the Huang and Stoll (1997) and Madhavan, Richardson and Roomans (1997) approaches.

4.1 Realized Spreads

The realized half spread is usually obtained by relating the price of a transaction to the quote midpoint a specified time (e.g., 5 minutes) after the trade (e.g., Huang and Stoll, 1996). The latter serves as an estimate of the asset value. Relating the transaction price to this estimate of the asset value results in an estimate of the gross revenue of the suppliers of liquidity.

As our data set does not contain quote data, we cannot use the quote midpoint to estimate the realized spread. We therefore proceed as follows. We match a buyer-initiated trade with the first seller-initiated trade after at least 5 minutes, and similarly we match a seller-initiated trade with the first buyer-initiated trade after at least 5 minutes. Let be the realized spread for transaction t,

R

st

{ }

, ,

i

P it ∈ a b the transaction price where a and b denote a buyer- and a seller- initiated trade, respectively, and let τ (chosen to be 5 minutes in our application) be the time interval between the matched trades. Then the expressions

R a

t t t

s =P −P^b_+τ (1)

b t

if the initial transaction was at the ask and

R a

t t

s =P_+τ −P (2)

if the initial transaction was at the bid yield an estimate of the realized spread (not half- spread). If the first trade occurring 5 minutes after the initial trade is equally likely to be buyer- or seller-initiated,⁸ our methods will yield the same result as the procedure proposed by Huang and Stoll (1996).

The results are presented in Panel A of Table II. The (unweighted) average realized spread in Xetra is 0.0076 Euro or 0.76 cents per share. The corresponding value in Xetra BEST is more than twice as high. Here, the average realized spread amounts to 1.82 cents. The relation that the realized spread in Xetra BEST is higher than the realized spread in Xetra holds for each of our sample stocks.

The averages considered thus far do not take into account that the average trade size in the two trading venues is different. We therefore repeat the analysis but now exclude all trades in the Xetra order book that are larger than the largest trade in Xetra BEST for the stock under investigation. The results are also presented in Table II. The last column presents the t-values of a test for equality of the means. The average realized spread in Xetra now amounts to 0.86 cents. It is larger than the average for the unrestricted sample for all eight stocks. This suggests that realized spreads are negatively related to trade size. Realized spreads are still

8 Although the order flow may be serially correlated, this correlation is unlikely to persist for five minutes.

Even the least liquid stock in our sample has a transaction frequency which corresponds to an average inter- trade duration of less than one minute. Our assumption that the first trade after five minutes is buyer or seller initiated with equal probability therefore appears to be innocuous.

larger in Xetra BEST for all sample stocks. The difference is statistically significant at the 5%

level for all but two stocks. Note that the realized spread estimates for Xetra BEST already incorporate the mandatory price improvement.

Insert Table II about here

Restricting the sample to trades that do not exceed the maximum trade size in Xetra BEST is a rather crude way to control for size dependence. We therefore run additional regressions that explicitly control for trading volume. Specifically, we estimate

( )

0 1 2ln

r B

st =γ +γ D +γ V_t +ε_t (3)

for each stock separately. s_t^r denotes the realized spread for transaction t, D^B is a dummy variable identifying transactions in Xetra BEST and Vt is the share volume of transaction t.

Order book trades with a size exceeding the maximum trade size in Xetra BEST are excluded from the analysis. The results are shown in Panel B of Table II. The conjectured negative relation between realized spreads and volume is confirmed. The corresponding coefficients are negative and significant for all sample stocks. Our previous finding that realized spreads are significantly higher in Xetra BEST is confirmed for six of the eight sample stocks. For the remaining two stocks the difference in realized spreads is insignificant with one coefficient being positive and the other one negative.

The analysis of the realized spreads thus provides evidence that internalization in Xetra BEST is profitable. Realized spreads are - even after taking the mandatory price improvement into account - higher than realized spreads for regular Xetra trades. In the next section we develop and test a structural model of the specific dual market structure in order to gain insights into the reason for the differences we uncovered.

4.2 Trade Indicator Models of the Xetra BEST Dual-Market Structure

Given that our data set identifies trades as either buyer-initiated or seller-initiated, it is natural to use a trade indicator model to compare the transaction costs and their components in Xetra and Xetra BEST. Trade indicator models have been proposed by Glosten and Harris (1988), Huang and Stoll (1997) and Madhavan, Richardson and Roomans (1997). For two reasons we choose to build on the Glosten and Harris (1988) approach. First, as documented in Table I, average trade sizes in Xetra and Xetra BEST are different. As the execution costs and their components may depend on trade size, a model including trade size is warranted. The Glosten- Harris model is the only model to include trade size as an explanatory variable. Both Huang and Stoll (1997) and Madhavan, Richardson and Roomans (1997) assume a constant trade size. Second, when the trade size in the Glosten-Harris model is assumed to be constant, the model reduces to the Huang-Stoll model and the Madhavan, Richardson and Roomans model (with the additional restriction of zero correlation in the order flow). We estimate such a restricted version.

We have to modify the Glosten and Harris (1988) model in order to account for the specific dual market structure under scrutiny.⁹ Let Qt be a trade indicator variable taking on the values 1 if transaction t is buyer-initiated and -1 if it is seller-initiated. Transaction price and share volume of transaction t are denoted by Pt and Vt, respectively. Let Mt denote the midpoint of

9 Note, however, that our derivation differs slightly from Glosten and Harris (1988). In their derivation, Mt is not the quote midpoint at time t but the conditional expected value of the asset at time t. The Mt in their definition thus already incorporates the information contained in the characteristics (i.e., direction and size) of the next transaction.

the best bid and ask quotation in the order book at the time the transaction t occurs. We assume that Mt evolves according to

( )

1 1 1 ^X1 1 1 ^B1 1

t t t t t t t t

M =M ₋ + −B₋ Z Q_{− −} +B Z Q_{− − −} +ε_t (4)

J

Zt is the adverse selection component of the spread. The index ^J∈

{

^{X B,}

}

refers to the Xetra order book (X) and Xetra BEST (B), respectively. Bt is a dummy variable taking on the value 1 if transaction t occurs in Xetra BEST and 0 otherwise. The intuition behind (4) is simple. The suppliers of liquidity incorporate the information revealed by transaction t-1, measured by the adverse selection component Z_t^J₋₁, into their bid and ask prices. The adverse selection component represents the information content of the trade. It is allowed to be different for transactions in Xetra and Xetra BEST, respectively. Public information releases, which will also affect the quotes, are incorporated into the error term ε_t. The adverse selection component is allowed to depend linearly on the size of the transaction:

0 1

J J J

t t

Z =z +z V (5)

From (4) and (5) we obtain the change in the quote midpoint

(

^Mt−^Mt−¹

)

=^{z Q0X t}−¹+^{z Q V1X t}−^{1 t}−¹+

(

^z0B−^z0X

)

^{B Qt}−^{1 t}−¹+

(

^z1B−^z1X

)

^{B Q Vt}−^{1 t}−^{1 t}−¹+^εt (6)

The bid and ask prices are based on the expected values of the asset conditional upon the characteristics (i.e., direction and size) of the next trade. They further incorporate the temporary component of the spread. For trades in the Xetra order book the transitory component, denoted C_t^X , is assumed to depend linearly on trade size:

(7)

0 1

X X X

Ct =c +c V_t

Consequently, the price for transaction t if it occurs in Xetra can be written as

(

0 1 0 1

)

X X X X

t t t t t

P =M + z +z V +c +c V Q +η_t (8)

where is simply the half spread and the error term captures random noise, rounding effects etc. This formulation implies that the transaction price is contingent upon the size of the trade. This is in line with models of limit order markets such as Glosten (1994). It is also broadly consistent with our data. Remember that transactions triggered by market orders walking up or down the book are recorded as one transaction at the volume- weighted average price in our data set. The prices of these transactions are clearly contingent upon the trade size. However, in order to check the robustness of our results, we will also estimate a version of the model in which we do not allow the transaction price to depend on contemporaneous trade size.

(

^{z0X +z V1X t +c0X +c V1X t}

)

Bid and ask prices in Xetra BEST are not derived independently, but are rather obtained by subtracting (for the ask price) or adding (for the bid price) the price improvement I from or to the Xetra ask and bid price, respectively. We assume I to be constant.¹⁰ For a transaction occurring in Xetra BEST the transaction price is thus

(

^{0X 1X 0X 1X}

)

t t t t t

P =M + z +z V +c +c V −I Q +η_t (9)

Combining (8) and (9) yields the following expression for the transaction price:

10 This is a realistic assumption because, as noted previously, the best executor enters the price improvement as a constant into the system and does not usually change it in the course of the trading day.

( ) ( ) ( )

( )

0 1 0 1 0 1 0 1

0 1 0 1

1 ^{X X X X X X X X}

t t t t t t t t t t

X X X X

t t t t t t t

P M B z z V c c V Q B z z V c c V I Q M z z V c c V Q IB Q

ηt

η

= + − + + + + + + + − +

= + + + + − + (10)

First-differencing, inserting (6) and rearranging terms yields our basic model

( ) ( )

(

^{0 0 0}

)

^{11 1}

(

^{10 1}

)

^{11 1 11}

⁽

^{1 11 1}

⁾

X X X X

t t t t t t t t t t

B X B X

t t t t t t t t t

P z Q z Q V c Q Q c Q V Q V

z z B Q z z B Q V I B Q B Q υ_t

− − −

− − − − − − −

∆ = + + − + −

+ − + − − − + (11)

where υ ε η η_t = + −_{t t t−1} is the error term. Note that, for B_t =B_t−1 =0, (11) reduces to equation (2) in Glosten and Harris (1988).

Next we derive a version of the model in which we do not allow the price of transaction t to depend on contemporaneous trade size. When setting their bid and ask prices the suppliers of liquidity instead assume a standardized trade size which we denote by V . Equation (10) then becomes

( ) ( ) ( )

( )

*

0 1 0 1 0 1 0 1

*

0 1 0 1

1 ^{X X X X X X X X}

t t t t t t t t

X X X X

t t t t t t

P M B z z V c c V Q B z z V c c V I Q M z z V c c V Q IB Q

ηt

η

= + − + + + + + + + − +

= + + + + − + (12)

Note that we still assume that, when updating the quote midpoint after a transaction, the suppliers of liquidity take the size of the previous trade into account. The analogue to (11) then becomes

( ) ( )

( ) ( ) ^{( )}

0 1 0 1 1 1 1 1 0 1

*

0 0 1 1 1 1 1 1 1 1 1

X X X X X X X X

t t t t

B X B X

t t t t t t t t t t

P z z V c c V Q z Q V z V c c V Q

z z B Q z z B Q V I B Q B Q υ

t 1

− − −

− − − − − − −

∆ = + + + + − + +

+ − + − − − + (13)

Note that

(

0 1 0 1

X X X X

z +z V +c +c V

)

is the half spread. Some of the deep parameters of the model

(

^{c c V0C, 1X,}

)

are not identified. This is inconsequential, however, as we are mainly

)

interested in the difference in the adverse selection component between the two trading systems,

(

z_o^B −z₀^X

)

^and

(

1 1 , and the effective price improvement I.

B X

z −z

If we generally disregard trade size, we obtain¹¹

(

1

) ( )

1 1

(

1

)

^**

X X B X

t t t t t t t t t t t

P θ Q φ Q Q₋ θ θ B Q_{− −} I B Q B Q_{− 1}

∆ = + − + − − − ₋ +υ (14)

We change the notation here because the parameters do not have the same interpretation as in the previous models. The model implicitly assumes a constant trade size, and ^θJ,J∈

{

^{X B,}

}

is the adverse selection component in market J, adapted to this constant trade size V . Therefore, θ^J corresponds to

(

0 1

J J

z +z V

)

in the previous model. Similarly, φ^X denotes the transitory component of the spread in Xetra, adapted to the constant trade size, and thus corresponds to

(

^{c0X +c V1X}

)

in the previous model.

Note that for (14) corresponds to equation (5) in Huang and Stoll (1997), and it also corresponds to equation (4) in Madhavan, Richardson and Roomans (1997) with the additional restriction that

1 0

t t

B =B₋ =

ρ =0.

(11), (13) and (14) are our empirical models. We start with the unrestricted model (11). We estimate it for each stock using OLS with Newey-West standard errors to account for the serial correlation caused by the specific structure of the error term.¹² The results are presented in Table III. The explanatory variables apparently explain a significant portion of the price

11 Equation (14) can be derived by setting V_t−1 =V in (13), making the parameter substitutions explained below and rearranging terms. V is the trade size which is now assumed to be constant.

12 GMM estimation yields virtually identical results.

changes. This is evidenced by R²s ranging from 0.27 to 0.34. The adverse selection component for regular Xetra trades is significant for all sample stocks. It depends positively on trading volume (i.e., ) as models such as Glosten (1994) would suggest. The transitory component is characterized by a positive intercept and a negative relation with trade size (i.e., ). There thus appear to be economies of scale in the execution of trades.

These are, however, overcompensated by the increase in adverse selection costs, as is evidenced by the fact that

1^X 0

z >

1^X 0

c >

1 1

X X

z > c for all sample stocks.

The adverse selection component in Xetra BEST is significantly smaller than its counterpart in Xetra. The intercept, is significantly lower than for all sample stocks whereas there are (with one exception

0

zB z₀^X

13) no differences in the coefficient measuring the relation between the adverse selection costs and trade size. Comparing to reveals that the adverse selection component for small trades in Xetra BEST is close to zero. Thus, there is clear evidence of cream skimming.

1

zB

0

zB z₀^X

The most surprising results are those for the effective price improvement I. We should expect a positive coefficient and, in fact, a coefficient close to 0.01, the minimum price improvement.

It turns out, however, that the effective price improvement is smaller than 0.01 for all sample stocks and is even significantly negative for three of the sample stocks.

13 As for this stock (DBK) the slope in Xetra BEST is larger, the adverse selection component will be larger in Xetra BEST than in Xetra for large trades. The trade size that makes the adverse selection component in both trading systems equal is 1,437 shares and is larger than the maximum trade size observed in Xetra BEST which is 1,200 shares. We can thus safely conclude that the adverse selection component is smaller in Xetra BEST for DBK, too.

Insert Table III about here

This result clearly requires explanation. The trade indicator models implicitly assume that the spread is constant. In reality, however, the spread fluctuates. Therefore, there are gains from timing a transaction. The information necessary to time a transaction in order to take advantage of a low spread is not freely available, however. Information on best bid and ask quotations is available on the internet with a 15 minute delay. Real time data must be purchased.

The benefit obtained from real time information depends on average trade size and trading frequency. As the retail customers whose orders are internalized tend to be small traders, it is quite likely that most of them do not purchase real time access to quote data. Consequently, there are two opposing effects that have an impact on the difference between the spread these traders pay and the spreads on regular order book trades. The informational disadvantage due to missing real time data prevents the timing of trades. This, in turn, will increase the average effective spread paid by these customers. The mandatory price improvement in Xetra BEST, on the other hand, serves to decrease the effective spread.

The relative magnitude of these opposing effects depends on the relative magnitude of the 1 cent price improvement as compared to the magnitude of intradaily fluctuations in the quoted spread. These fluctuations as well as the level of the spreads are likely to be larger for higher priced stocks.¹⁴ Thus, for higher priced stocks a one cent price improvement is of relatively lower value than the same price improvement for lower-priced stocks. This argument implies that the effective price improvement we are measuring (and which captures the mandatory

14 Remember that we measure spreads in Euro, not in percentage terms.

price improvement as well as timing effects) should be negatively related to the price level of the stock. We find supporting evidence for this conjecture. The three stocks with negative effective price improvement are also the stocks with the highest average prices. The rank correlation between the effective price improvement and the average price is -0.71.¹⁵

We turn to the restricted model (13) next. The R²s are almost as high as those from the unrestricted model. As noted previously, some of the deep parameters of the model are not identified. The parameters of interest here - the difference in the magnitude of the adverse selection costs and the effective price improvement - can readily be estimated, however. The results are presented in Table I. The effective half spread is, as is to be expected, positive and significant. More importantly, the coefficients are slightly higher than the sum of the coefficients and in Table III. This is evidence that the results of the restricted model are consistent with those of the unrestricted model. This is borne out by the other results. The adverse selection component of the spread is lower in Xetra BEST for all sample stocks. There are no differences in the size sensitivity of the adverse selection component, again with the exception of DBK. The effective price improvement is smaller than 0.01 for all stocks and we find negative values for the same three stocks as before. Thus, the restricted model corroborates the evidence that there is cream skimming, and that the mandatory price improvement does not generally succeed in lowering the effective spreads paid by the retail customers whose orders are internalized.

0

zX c₀^X

Insert Table IV about here

15 This figure should obviously be interpreted with care as it is based on eight observations only.

Our final model (14) assumes a constant trade size. The spread and its components thus relate to a trade of standard size. The results are presented in Table V. They are perfectly consistent with those of the other models. The adverse selection component in Xetra, θ^X , is positive and significant for all stocks. In all cases, it is slightly larger than the coefficient estimate for in Table III. This is to be expected given that is the intercept of the adverse selection component (i.e., is the adverse selection component for a trade of size zero) and that the adverse selection component was shown to increase with trade size. The transitory component,

0

zX

0

zX

0

zX

φX , is also significant and positive, and is slightly smaller in magnitude than in Table III.

This is again as expected since is the intercept of the transitory component (i.e., is the transitory component for a trade of size zero), and the transitory component was shown to decrease with trade size.

0

cX

0

cX c₀^X

The adverse selection component in Xetra BEST, θ^B, is significantly smaller than θ^X . In fact, it is close to zero for most of the sample stocks. This is also consistent with the results of the unrestricted model. The estimates of the effective price improvement are slightly larger than those in Table III but show the same pattern. They are all smaller than 0.01, and we obtain significant negative values for the same three stocks as before.

Insert Table V about here

The results of the trade indicator models can be summarized as follows. All three models yield remarkably similar results. There is clear evidence of cream skimming. The adverse selection component is smaller in Xetra BEST for all sample stocks. The mandatory price improvement does not fully compensate for the difference in adverse selection costs. In fact, the effective price improvement is much smaller than the required minimum price improvement of 1 cent

and is even negative for three of the sample stocks. A potential explanation for this surprising result rests on the supposition that retail customers often do not have access to real time quote information and therefore are not able to time their trades.

5 Summary and Conclusion

The practice of internalizing order flow (and the related practices of preferencing and payment for order flow) has been subject to severe criticism. It is said to impair the quality of the main market by fragmenting the order flow, by weakening price competition between suppliers of liquidity, and by diverting easy-to-fill orders (i.e., those least likely to be affected by adverse selection problems) away from the main market. Although the existing empirical evidence hardly supports this criticism, the same arguments have been brought forward again when the commission of the European Union published its proposal for an Investment Services Directive (ISD). Although the revised ISD that was finally accepted is less liberal than the initial draft, many exchanges in the EU will have to face the threat of internalization once national legislation has implemented the ISD rules.

In Germany internalization, albeit not practiced on a larger scale, was and is legal. Early this decade it became apparent that some major banks were planning to embark on internalization.

Deutsche Börse AG responded to that threat by launching Xetra BEST in September 2002.

This system allows participating brokers to internalize customer orders on a platform owned and operated by the exchange. The exchange has designed the system in a way that is intended to reduce the potentially adverse impacts on the main market. Most importantly, there is a mandatory price improvement rule guaranteeing that internalized orders receive execution at a price that is at least 1 cent better than the price the order would receive were it executed in Xetra. Further, internalizing brokers must act as liquidity managers in the Xetra order book.

Xetra BEST, if successful, may be a blueprint for other EU exchanges facing the threat of internalization. Against this background, the objective of the present paper is to provide a thorough empirical assessment of Xetra BEST. We find that the market share of Xetra BEST is small. It amounts to 1.5% when considering the number of transactions and to 0.25% when considering the trading volume.

The main part of our analysis focuses on execution quality and the components of the spread.

We first estimate the realized spread using a method similar to the two-way decomposition proposed by Huang and Stoll (1996). We find that realized spread are unanimously higher in Xetra BEST after taking into account the mandatory price improvement. This indicates that internalization is profitable.

We then proceed by developing a structural model of the specific dual market structure under scrutiny. The model is based on the trade indicator model suggested in Glosten and Harris (1988). When estimating this model (and two restricted versions thereof) we find that adverse selection costs are clearly lower in Xetra BEST than in Xetra. This is interpreted as evidence in favor of cream skimming.

Cream skimming is not in itself disadvantageous. If the execution costs of the internalized orders reflect the lower adverse selection risk, then cream skimming is a way to achieve price discrimination. Our findings suggest, however, that the execution costs do not adequately compensate for the lower adverse selection risk. We find that the effective price improvement is clearly smaller than the mandatory minimum price improvement of 1 cent. It is even negative for three of the sample stocks. Our explanation for this surprising result rests on the supposition that retail customers are less likely to have access to real time quote information than institutional investors. Therefore, they are unable to time their transactions in order to

take advantage of low spreads. This effect counterbalances the price improvement and may even overcompensate it. It should be noted, however, that even a negative effective price improvement does not imply that the customers are put at a disadvantage by having their trades executed in Xetra BEST. On the contrary - if their trades were executed in Xetra they would still be at a disadvantage with respect to real time quote information, but they would forego the price improvement.

What are the implications of our results? Apparently, internalization is profitable (as is evidenced by higher realized spreads as compared to transactions in Xetra) because the price improvement does not fully compensate for the lower adverse selection risk. Increasing competition among internalizers may lead to increased price improvement and may thus reduce the execution costs of internalized orders. The transparency of the internalizers' quotes mandated by the ISD may well serve to increase competition among internalizers and therefore appears to be a step in the right direction.

References

Battalio, R.,1997. Third Market Broker-Dealers: Cost Competitors or Cream Skimmers?

Journal of Finance 52, 341--352.

Battalio, R., Holden, C.,2001. A Simple Model of Payment for Order Flow, Internalization, and Total Trading Cost. Journal of Financial Markets 4, 33--71.

Battalio, R., Greene J., Jennings,R., 1997. Do Competing Specialists and Preferencing Dealers Affect Market Quality? Review of Financial Studies 10, 969--993.

Battalio, R., Jennings, R., Selway, J., 2001. The Relationship Among Market-Making Revenue, Payment for Order Flow, and Trading Costs for Market Orders. Journal of Financial Services Research 19, 39--56.

Biais, B., Davydoff, D., 2002. Internalization, Investor Protection and Market Quality.

Working Paper, July.

Bloomfield, R., O’Hara, M., 1998. Does Order Preferencing Matter? Journal of Financial Economics 50, 3--37.

Chakravarty, S., 2001. Stealth-Trading: Which Traders' Trades Move Stock Prices? Journal of Financial Economics 61, 289--307.

Easley, D., Kiefer, N., O’Hara, M., 1996. Cream-Skimming or Profit Sharing? The Curious Role of Purchased Order Flow. Journal of Finance 51, 811--833.

Gomber, P., Maurer, K., 2004. Integration of Market Access Intermediaries' Requirements Into Market Design. Electronic Markets 14, 214--222.

Glosten, L., 1994. Is the Electronic Open Limit Order Book Inevitable? Journal of Finance 49, 1127--1161.

Glosten, L., Harris, L., 1988. Estimating the Components of the Bid-Ask Spread. Journal of Financial Economics 21, 123--142.

Hansch, O., Naik, N., Viswanathan, S., 1999. Preferencing, Internalization, Best Execution, and Dealer Profits. Journal of Finance 54, 1799--1828.

Huang, R., Stoll, H., 1996. Dealer versus Auction Markets: A Paired Comparison of Execution Costs on NASDAQ and the NYSE. Journal of Financial Economics 41, 313--357.

Huang, R., Stoll, H., 1997. The Components of the Bid-Ask Spread: A General Approach.

Review of Financial Studies 10, 995--1034.

Kandel, E., Marx, L., 1999. Payments for Order Flow on Nasdaq. Journal of Finance 54, 35-- 66.

Linnainmaa, J., 2003. Who Makes the Limit Order Book? Implications for Contrarian Strategies, Attention-Grabbing Hypothesis, and the Disposition Effect. Working Paper, University of California Los Angeles, October.

Madhavan, A., Richardson, M., Roomans, M., 1997. Why Do Securities Prices Change? A Transaction-Level Analysis of NYSE Stocks. Review of Financial Studies 10, 1035--1064.

Parlour, Ch., Rajan, U., 2001. Payment for Order Flow. Working Paper, Carnegie Mellon University, May.

Peterson, M., Sirri, E., 2002. Order Preferencing and Market Quality on U.S. Equity Exchanges. Forthcoming in: Review of Financial Studies.

Table I: Descriptive Statistics

The table presents descriptive statistics for our sample. The first column identifies the stock. Stocks are sorted by total € trading volume in the sample period. Columns 2 - 4 show the number of transactions, the average trade size in shares and the average trade size in € for regular order book trades. Columns 5 - 7 provide the same information for Xetra BEST trades.

Xetra Xetra BEST

Stock id # of transactions

av. trade size, shares

av. trade size, €

# of transactions

av. trade size, shares

av. trade size, €

DBK 289,425 1,450.03 65,022.85 3,009 173.89 7,770.01

DCX 324,365 1,466.66 47,597.64 8,687 160.00 5,501.09

EOS 221,584 1,285.71 56,346.44 1,741 155.67 6,810.13

BAY 195,192 1,694.96 34,189.54 2,946 306.44 6,128.24

RWE 155,967 1,204.76 34,040.08 1,370 219.35 6,143.28

SCH 110,351 809.52 36,038.15 1,920 156.63 6,979.33

TKA 91,177 2,079.62 22,614.20 1,214 440.11 4,768.81

DPW 70,789 1,629.19 16,523.77 1,484 404.62 4,054.37

Table II: Realized Spreads

Panel A shows average realized spreads for the sample stocks. Realized spreads are calculated as follows: We match a buyer-initiated trade with the first seller-initiated trade after at least 5 minutes, and similarly we match a seller-initiated trade with the first buyer-initiated trade after at least 5 minutes. Let st^r be the realized spread for transaction t, P j_t^j; ∈{ }a b, the transaction price where a and b denote a buyer- and a seller-initiated trade, respectively, and let τ (chosen to be 5 minutes in our application) be the time interval between the matched trades. Then the expressions s_t^R =P_t^a−P_t^b_+τ and s_t^R =P_t^a_+τ−P_t^b if the initial transaction was at the ask and the bid, respectively, yield an estimate of the realized spread (not half-spread). Column 2 shows average realized spreads for regular order book trades. Column 3 excludes trades with a size larger than the maximum Xetra BEST trade size for the stock in question. Column 4 contains average realized spreads for trades in Xetra BEST. The last column presents the t-statistics for a test of the null hypothesis that the averages given in columns 3 and 4 are equal.

Panel B presents the results of the regression s_t^r =γ0+γ1D^B+γ2ln( )V_t +ε_t. Order book trades with a size exceeding the maximum trade size in Xetra BEST are excluded from the estimation. t-values are given in parentheses.

Panel A: Univariate Results

Realized spread Xetra

Realized spread Xetra; Trade size

< maximum size in Xetra BEST

Realized spread Xetra BEST

t-value (3) vs. (4)

DBK 0.0081 0.0105 0.0242 4.29

DCX 0.0107 0.0119 0.0198 5.78

EOS 0.0078 0.0091 0.0242 4.26

BAY 0.0056 0.0061 0.0105 2.81

RWE 0.0051 0.0052 0.0153 3.61

SCH 0.0128 0.0146 0.0378 6.28

TKA 0.0059 0.0062 0.0074 0.90

DPW 0.0049 0.0052 0.0065 1.15

average (unweighted) 0.0076 0.0086 0.0182

Panel B: Regression Results

constant Dummy Xetra BEST ln(volume)

DBK 0.0349

(16.72)

0.0086 (2.77)

-0.0042 (11.87)

DCX 0.0293

(23.42)

0.0039 (2.95)

-0.0029 (14.11)

EOS 0.0206

(10.15)

0.0127 (3.56)

-0.0020 (5.80)

BAY 0.0137

(12.89)

0.0030 (2.05)

-0.0012 (7.24)

RWE 0.0098

(6.85)

0.0091 (3.63)

-0.0007 (3.23)

SCH 0.0402

(15.42)

0.0181 (4.84)

-0.0045 (10.02)

TKA 0.0128

(16.63)

0.00001 (0.01)

-0.0010 (8.67)

DPW 0.0143

(19.36)

-0.0001 (0.11)

-0.0014 (12.52)