Algorithmic Trading and Information

 

 

Algorithmic Trading and Information ∗

Terrence Hendershott
Haas School of Business
University of California at Berkeley

Ryan Riordan
Department of Economics and Business Engineering
Karlsruhe Institute of Technology

August 18, 2009

Abstract

We examine algorithmic trades (AT) and their role in the price discovery process in the 30
DAX stocks on the Deutsche Boerse. AT liquidity demand represents 52% of volume and AT
supplies liquidity on 50% of volume. AT act strategically by monitoring the market for liquidity
and deviations of price from fundamental value. AT consume liquidity when it is cheap and
supply liquidity when it is expensive. AT contribute more to the efficient price by placing more
efficient quotes and AT demanding liquidity to move the prices towards the efficient price.

∗We thank Bruno Biais and conference participants at the IDEI-R Conference on Investment Banking and Financial
Markets for helpful comments. Hendershott gratefully acknowledges support from the Net Institute, the Ewing Marion
Kauffman Foundation, and the Lester Center for Entrepreneurship and Innovation at the Haas School at UC Berkeley.
Riordan gratefully acknowledges support for the Deutsche Forschungs Gemeinschaft – graduate school Information
and Market Engineering at the Karlsruhe Institute of Technology.

1

1

Introduction

Technology has revolutionized the way financial markets function and the way financial assets

are traded. Two significant interrelated technological changes are investors using computers to

automate their trading processes and markets reorganizing themselves so virtually all markets are

now electronic limit order books (Jain (2005)). The speed and quality of access to such markets

encourages the use of algorithmic trading (AT; AT denotes algorithmic traders as well), commonly

defined as the use of computer algorithms to automatically make trading decisions, submit orders,

and manage those orders after submission. Because the trading process is central to efficient risk

sharing and price efficiency it is important to understand how AT is used and its role in the price

formation process. We examine these issues for DAX stocks (the 30 largest market capitalization

stocks) traded on the Deutsche Boerse (DB) with data identifying whether or not each trade’s

buyer and seller generated their order with an algorithm. Directly identifying AT is not possible

in most markets, so relatively little is known.1

Liquidity demanders use algorithms to try to identify when a security’s price deviates from the

efficient price by quickly processing information contained in order flow and price movements in

that security and other securities across markets. Liquidity suppliers must follow a similar strategy

to avoid being picked off. Institutional investors also utilize AT to trade large quantities gradually

over time, thereby minimizing market impact and implementation costs.

Most markets offer volume discounts to attract high-frequency traders. The development costs

of AT typically lead to it being adopted first by high-volume users who automatically qualify for the

quantity discounts. The German competition authority did not allow for generic volume discounts,

rather requires that such discounts have a cost sensitive component. The DB successfully asserted

that algorithm generated trading is lower cost and highly sensitive to fee reductions and therefore,

could receive quantity discounts. In December of 2007, the DB introduced its fee rebate program

for automated traders. The DB provided data on AT orders in the DAX stocks for the first three

weeks of January 2008.

1Biais and Weill (2008) theoretically examine the relation between AT, market monitoring, and liquidity dynamics.
Chaboud et al. (2009) study AT in the foreign exchange market. Hendershott, Jones, and Menkveld (2008) use a
proxy for AT to examine AT’s effect on liquidity in the equity market.

2

AT initiate 52% of trading volume via marketable orders. AT initiate smaller trades with AT

initiating 68% of volume for trades of less than 500 shares and 23% of volume for trades of greater

than 10,000 shares. AT initiate trades quickly when spreads are small and cluster their trades

together. AT are more sensitive to human trading activity than humans are to AT trading activity.

These are all consistent with AT closely monitoring the market for trading opportunities. If an

algorithmic trader is constantly monitoring the market, the trader can break up their order into

small pieces to disguise their intentions and quickly react to changes in market conditions. AT

could also be trying to exploit small deviations of price from fundamentals.

Moving beyond unconditional measures of AT activity we estimate probit models of AT using

market condition variables incorporating the state of the limit order book and past volatility and

trading volume. We find that AT are more likely to initiate trades when liquidity is high in terms

of narrow bid-ask spreads and higher depth. AT liquidity demanding trades are not related to

volatility in the prior 15 minutes, but AT initiated trading is negatively related to volume in the

prior 15 minutes.

Just as algorithms are used to monitor liquidity in the market, algorithms may also be used to

identify and capitalize on short-run price predictability. We use a standard vector auto-regression

framework (Hasbrouck (1991a) and Hasbrouck (1991b)) to examine the return-order flow dynamics

for both AT and human trades. AT liquidity demanding trades play a more significant role in

discovering the efficient price than human trades. AT initiated trades have a more than 20% larger

permanent price impact than human trades. In terms of the total contribution to price discovery—

decomposing the variance of the efficient price into its trade-correlated and non trade-correlated

components—AT liquidity demanding trades help impound 40% more information than human

trades. The larger percentage difference between AT and humans for the variance decomposition

as compared to the impulse response functions implies that the innovations in AT order flow are

greater than the innovations in human order flow. This is consistent with AT being able to better

disguise their trading intentions.

We also examine when AT supply liquidity via non-marketable orders. The nature of our data

makes it possible to build an AT-only limit order book, but makes it difficult to perfectly identify

3

when AT supply liquidity in transactions (see Section 3 for details). Therefore, we focus our analysis

on quoted prices associated with AT and humans. While AT supply liquidity for exactly 50% of

trading volume, AT are at the best price (inside quote) more often than humans. This AT-human

difference is more pronounced when liquidity is lower, demonstrating that AT supply liquidity more

when liquidity is expensive.

We also examine the role of AT quotes in the price formation process. We calculate the in-

formation shares (Hasbrouck (1995)) for AT and human quotes. AT quotes play a larger role in

the price formation process than their 50% of trading volume. The information shares decompose

the changes in the efficient price into components that occur first in AT quotes, human quotes,

and appear contemporaneously in AT and human quotes with the corresponding breakdown being

roughly 50%, 40%, and 10%, respectively. The ability of AT to update quotes quickly based on

changing market conditions could allow AT to better provide liquidity during challenging market

conditions.

The results on AT liquidity supply and demand suggest that AT monitor liquidity and informa-

tion in the market. AT consume liquidity when it is cheap and supply liquidity when it is expensive,

smoothing out liquidity over time. AT also contribute more to the efficient price by having more

efficient quotes and AT demanding liquidity so as to move the prices towards the efficient price.

Casual observers often blame the recent increase in market volatility on AT.2 AT demanding liq-

uidity during times when liquidity is low could result in AT exacerbating volatility, but we find no

evidence of this. AT could also exacerbate volatility by not supplying liquidity when the liquidity

dries up. However, we find the opposite.

Section 2 relates our work to existing literature. Section 3 describes the algorithmic trading on

the Deutsche Boerse. Section 4 describes our data. Section 5 analyzes when and how AT demands

liquidity. Section 6 examines how AT demand liquidity relates to discovering the efficient price.

Section 7 studies when AT supply liquidity and its relation to discovering the efficient price. Section

8 concludes.

2For example, see “Algorithmic trades produce snowball effects on volatility,” Financial Times, December 5, 2008.

4

2 Related Literature

Due to the difficulty in identifying AT, most existing research directly addressing AT has used data

from brokers who sell AT products to institutional clients. Engle, Russell, and Ferstenberg (2007)

use execution data from Morgan Stanley algorithms to study the tradeoffs between algorithm

aggressiveness and the mean and dispersion of execution cost. Domowitz and Yegerman (2005)

study execution costs of ITG buy-side clients, comparing results from different algorithm providers.

Several recent studies use comprehensive data on AT. Chaboud et al. (2009) study the devel-

opment of AT in the foreign exchange market on the electronic broking system (EBS) in three

currency pairs euro-dollar, dollar-yen, and euro-yen. They find little relation between AT and

volatility, as do we. In contrast to our results, Chaboud et al. (2009) find that non-algorithmic

order flow accounts for most of the variance in FX returns. There are several possible explanations

for this surprising result: (i) EBS’ origins as an interdealer market where algorithms were closely

monitored; (ii) humans in an interdealer market being more sophisticated than humans in equity

markets; or (iii) there is relatively little private information in FX. Chaboud et al. (2009) find

that AT seem to follow correlated strategies, which is consistent with our results of AT clustering

together in time. Hendershott, Jones, and Menkveld (2008) use a proxy for AT, message traffic,

which is the sum of order submissions, order cancelations, and trades. Unfortunately, such a proxy

makes it difficult to closely examine when and how AT behave and their precise role in the price

formation process. Hendershott, Jones, and Menkveld (2008) are able to use an instrumental vari-

able to show that AT improves liquidity and makes quotes more informative. Our results on AT

liquidity supply and demand being more informed are the natural mechanism by which AT would

lead to more informationally efficient prices.

Any analysis of AT relates to models of liquidity supply and demand.3 Liquidity supply involves

posting firm commitments to trade. These standing orders provide free-trading options to other

traders. Using standard option pricing techniques, Copeland and Galai (1983) value the cost of the

option granted by liquidity suppliers. The arrival of public information can make existing orders

stale and can move the trading option into the money. Foucault, Ro¨ell, and Sandas (2003) study

3Parlour and Seppi (2008) for a general survey on limit order markets.

5

the equilibrium level of effort that liquidity suppliers should expend in monitoring the market to

avoid this risk. AT enables this kind of monitoring and adjustment of limit orders in response to

public information,4 but AT can also be used by liquidity demanders to pick off liquidity suppliers

who are not fast enough in adjusting their limit orders with public information. The monitoring

of the state of liquidity in the market and taking it when cheap and making it when expensive is

consistent with AT playing an important role in the make/take liquidity cycle modeled by Foucault,

Kadan, and Kandel (2008).

Algorithms are also used by traders who are trying to passively accumulate or liquidate a large

position. Bertsimas and Lo (1998) find that the optimal dynamic execution strategies for such

traders involves optimally braking orders into pieces so as to minimize cost.5 While such execution

strategies pre-dated wide-spread adoption of AT (cf. Keim and Madhavan (1995)), brokers now

automate the process with AT products.

For each component of the larger transaction, a trader (or algorithm) must choose the type

and aggressiveness of the order. Cohen et al. (1981) and Harris (1998) focus on the simplest

static choice: market order versus limit order. If a trader chooses a non-marketable limit order,

the aggressiveness of the order is determined by its limit price (Griffiths et al. (2000) and Ranaldo

(2004)). Lo, MacKinlay, and Zhang (2002) find that execution times are very sensitive to the choice

of limit price. If limit orders do not execute, traders can cancel them and resubmit them with more

aggressive prices. A short time between submission and cancelation suggests the presence of AT,

and in fact Hasbrouck and Saar (2008) find that a large number of limit orders are canceled within

two seconds on the INET trading platform (which is now Nasdaq’s trading mechanism).

3 Deutsche Boerse’s Automated Trading Program

The Deutsche Boerse’s order-driven electronic limit order book system is called Xetra (see Hau

(2001) for details).6 Orders are matched using price-time-display priority. Quantities available at

4Rosu (2006) implicitly incorporates AT by assuming limit orders can be constantly adjusted.
5Almgren and Chriss (2000) extend this by considering the risk that arises from breaking up orders and slowly

executing them.

6Iceberg orders are allowed as on the Paris Bourse (cf. Venkataraman (2001)).

6

the 10 best bid and ask prices as well as numbers of participants at each level are disseminated

continuously. See the Appendix for further details on Xetra.

During our sample period Xetra had a 97% market share of German equities trading. With

such a dominant position the competition authorities (Bundeskartellamt) required approval of all fee

changes prior to implementation. Fee changes must meet the following criteria: (i) all participants

are treated equally; (ii) changes must have a cost-related justification; and (iii) fee changes are

transparent and accessible to all participants. Criterion (i) and (iii) ensure a level playing field for

all members and is comparable to regulation in the rest of Europe and North America. The second

criteria is the most important for this paper. AT are viewed as satisfying the cost justification for

the change, so DB could offer lower trading fees for AT.

In December of 2007 the DB introduced its Automated Trading Program (ATP) to increase the

volume of automated trading on Xetra. To qualify for the ATP an electronic system must deter-

mine the price, quantity, and submission time for orders. In addition, the Deutsche Boerse ATP

agreement requires that: (i) the electronic system must generate buy and sell orders independently

using a specific program and data; (ii) the generated orders must be channeled directly into the

Xetra system; and (iii) the exchange fees or the fees charged by the ATP member to its clients

must be directly considered by the electronic system when determining the order parameters.

Before being admitted to the ATP, participants must submit an high-level overview of the

electronic trading strategies they plan to employ. The level of disclosure required here is intended

to be low enough to not require ATP participants to reveal important details of their trading

strategies. Following admission to the ATP, the orders generated by each participant are audited

monthly for plausibility.

If the order patterns generated do not match those suggested by the

strategy plan submitted by a participant or are considered likely to have been generated manually,

the participant will be terminated from the ATP and may also be suspended from trading on

Xetra. Conversations with the DB revealed that a small portion of AT orders may not be included

in the data set. The suspicion on the part of the DB is due to the uncommonly high number of

orders (message traffic) to executions of certain participants which is typical of AT. However, these

participants make up less than 1% of trades in total and are, therefore, unlikely to affect our results.

7

The ATP agreement and the auditing process ensure that most, if not all, of the orders submitted by

an ATP participant are electronically generated and that most, if not all, electronically generated

orders are included in our data.

The DB only charges fees for executed trades and not for submitted orders. The rebate for ATP

participants can be significant. The rebates are designed to increase with the total trade volume

per month. Rebates are up to a maximum of 60% for euro monthly volume above 30 billion. The

first Euro volume rebate level begins at a 250 million Euro volume and is 7.5%. Table 1 provides

an overview of the rebate per volume level.

[Insert Table 1 Here]

For an ATP participant with 1.9 billion euros in eligible volume, the percentage rebate would

be (volumes are in millions of euros):

(250 ∗ 0% + 250 ∗ 7.5% + 500 ∗ 15.0% + 900 ∗ 22.5%)/1, 900 = 15.6%

(1)

In the example above, an ATP participant would receive a rebate of 15.6%. This translates

into roughly 14,000 euros in trading cost savings on 91,200 in total, and an additional 5,323 euros

savings on 61,500 in total in clearing and settlement costs. This rebate (14,000 + 5,323) translates

into a 0.1 basis point saving on the 1.9 billion in turnover. For high-frequency trading firms, whose

turnover is much higher than the amount of capital invested, the savings are significant.

The fee rebate for ATP participants is the sole difference in how orders are treated. AT orders

are displayed equivalently in the publicly disseminated Xetra limit order book. The Xetra matching

engine does not distinguish between AT and human orders. Therefore, there are no drawbacks for

an AT firm to become an ATP participant. Thus, we expect all AT to take advantage of the

lower fees by becoming ATP participants. From this point on we equate ATP participants with

algorithmic traders and use AT for both. We will refer to non-ATP trades and orders as human or

human-generated.

8