Perspectives on High-Frequency Trading

Fokion Zervoudakis, Donald Lawrence,

Noemi Nava, George Gontikas, and Majd Al Merey
Department of Computer Science, University College London, London, UK
{f.zervoudakis,d.lawrence,n.nava,gontikas,malmerey}@cs.ucl.ac.uk

Abstract
High-frequency trading (HFT) is algorithmic trading that utilizes so-
phisticated low latency technologies to conduct automated market mak-
ing, spread trading, arbitrage and macrotrading. HFT currently generates
approximately 35% and 70% of UK and U.S. equity trades, respectively,
and thereby exerts significant influence over the global financial markets.
The ubiquitous diffusion of HFT systems proceeds in contrast to the dis-
cord and controversy engendered by those systems. Financial regulators,
market participants, academics and members of the media express diverg-
ing, and sometimes exaggerated, views with respect to HFT benefits and
risks. The May 6, 2010 Flash Crash, which temporarily erased $1 tril-
lion in market value, highlights the potential for amplification of financial
anomalies by HFT systems. We argue for an informed dialogue between
stakeholders to realize a common domain of discourse, and an appropri-
ate and resilient regulatory framework. To this end, we have conducted
a survey that collates the opinions of academics, market participants and
regulators, with respect to HFT benefits and risks. Our findings are pre-
sented in this report.

1 Introduction
The term algorithmic trading describes financial market transactions that are
based on instructions generated by quantitative algorithms [1]. These algorithms
can be encoded in a programming language, and executed automatically by
computers. High-frequency trading (HFT) is algorithmic trading that utilizes
sophisticated low latency technologies to conduct automated market making,
spread trading, arbitrage and macrotrading [2, 3]. We consider low latency
environments to be those environments that can support single digit millisecond
round-trip time, which is the time from data packet transmission to reception [4].
HFT currently generates approximately 35% and 70% of UK and U.S. eq-
uity trades, respectively, and thereby exerts significant influence over the global
financial markets [5, 6]. In European equity trading, HFT market shares range
from 13% to 40%, respectively, for Chi-X and Nasdaq OMX [7]. HFT adaptation
rates have been swift and profound. Figures 1 and 2 illustrate high-frequency

1
Figure 1: HFT activity for U.S. stock exchanges, as identified by the legend, on
March 21, 2007. The x and y axes represent market trading hours—from 9:30
to 16:00 Eastern Time—and number of packets, respectively.

trading activity for U.S. stock exchanges during March 2007 (top) and January
2012 (bottom), respectively [8].
The ubiquitous diffusion of HFT systems proceeds in contrast to the dis-
cord and controversy engendered by those systems. Financial regulators, mar-
ket participants, academics and members of the media express diverging, and
sometimes exaggerated, views with respect to HFT benefits and risks. One
point of contention is high-frequency quoting, which involves the rapid oscilla-
tion of bid and/or ask prices. HFT proponents consider this technique to be
a source of tight bid–ask spreads, which are associated with low volatility [9].
But high-frequency quoting has also been associated with the May 6, 2010 Flash
Crash [10], a period of extreme volatility [11].
The events of May 6, which temporarily erased $1 trillion in market value,
highlight the potential for amplification of financial anomalies by HFT systems.
Other comparable events include the Dow Jones Industrial Average flash crash,
which occurred September 29, 2008 [12]; the cocoa futures mini crash, which
occurred on March 1, 2011; and the dollar–yen sell off, which occurred on March
16, 2011 [13]. The parameters that characterize these events—including acute
price fluctuations that approach upper and lower equity price limits—are clearly
attributable to automated processes [14].
We argue for an informed dialogue between stakeholders to realize a common
domain of discourse, and an appropriate and resilient regulatory framework. To
this end, we have conducted a survey that collates the opinions of academics,

2
 Figure 2: HFT activity for U.S. stock exchanges on January 27, 2012.

market participants and regulators, with respect to HFT benefits and risks. Our
findings are presented in the remainder of this report, which is structured as
follows. Section 2 presents and evaluates various candidate definitions for HFT.
Section 3 describes the ecosystem that underpins HFT activity. Sections 4
and 5 describe the effects of HFT on liquidity and volatility, respectively. The
events of May 6 are described in Section 6. Finally, Sections 7 and 8 present
recommendations and conclusions, respectively.

2 High-Frequency Trading: A Definition

A clear and consistent definition for HFT remains elusive [7]. While HFT can
be considered a subset of algorithmic trading, HFT and algorithmic trading can
also be considered equivalent. The latter viewpoint is justifiable when behavior
replaces speed as the primary HFT descriptor. The European Commission (EC)
defines algorithmic trading as the use of computer programs to enter trading
orders, where the algorithm decides on aspects of order execution such as timing,
quantity and price [15]. But HFT has yet to be defined by the EC, and by
regulators in general. High-frequency traders (HFTs) consider HFT a technique
that can be applied to various, low latency, strategies [16]. A technique is the
mechanism by which a strategy is executed; a strategy is an algorithm designed
to achieve a set of objectives.
On a practical level, HFTs utilize sophisticated technology to interact with
markets, analyze market data and execute multiple high-speed trades. Each
trade may yield fractional gains; HFT net profits for the fiscal year 2008 in the

3
U.S. were estimated by the TABB Group at $8–$20 billion. Most positions,
which may be held for as little as a few milliseconds, are closed by the end
of the trading day. While HFT is particularly applicable to market making
and arbitrage, any automated trading strategy can use HFT techniques to op-
timize execution and mask trading activities from competitors. In this context,
establishing a comprehensive definition for HFT becomes a challenging task.
In addition to practical concerns, HFT also engenders philosophical dis-
agreement. HFT can be considered the evolution of trading, which somewhat
parallels the relationship between race cars and horse-drawn carriages. Several
arguments support this point of view. HFT is perhaps best described as a race
to zero, where market participants compete and trade at near-light speeds. This
operational requirement imposes capital risk on HFT firms, which must assume
significant technological investment to remain competitive. As physical limits
are reached however, innovative technology becomes commoditized and speed
increases yield diminishing returns. Consequently, market and sentiment anal-
yses become the primary determinants of success [17, 18]. But this is also true
for algorithmic, and perhaps discretionary, trading in general. Moreover, unlike
arbitrage strategies that can be executed in milliseconds, HFTs that assume
a market making role are exposed to market risk. But again, this is true for
conventional market participants.
HFT can also be considered a significant extension of discretionary trading.
From this viewpoint, HFT is a contrived term that simplifies what is an in-
herently complex subject. In particular, HFT is performed by automated and
adaptive algorithms at near-light speeds. These algorithms operate in the con-
text of a global network that encompasses significant constituent components
and human–computer interactions [5]. It is arguable that a stochastic state
space of this magnitude cannot be delineated in a succinct manner.
Practical and philosophical concerns have implications for ongoing regula-
tory efforts. Instead of attempting to define HFT, or to delineate what is an
inherently complex state space, regulators should be focused on identifying and
restricting harmful behavior. Furthermore, if HFT constitutes the evolution of
trading, then any requests for responsibility—social or otherwise—must neces-
sarily be applied to conventional market participants as well as HFTs.

3 The Ecosystem
Cliff et al. argue that the global financial markets constitute a complex adap-
tive ultra-large-scale socio-technical system-of-systems [19]. The concepts that
characterize this type of system can be described as follows.

− The adaptive behavior of human and software agents, and the emergent1
behavior resulting from the multiplicity of interactions between those
agents, combine to form a complex adaptive system.
1 Emergence is a phenomenon that describes the appearance of complexity from a multi-

plicity of relatively simple interactions.

4
 − An ultra-large-scale system is unprecedented in size with respect to lines
of code; number of users; amount of data stored, accessed, manipulated
and refined; number of connections and interdependencies among software
components; and number of hardware elements [20].
− A socio-technical system comprises humans and their computational and
physical environments.
− A system-of-systems is a complex system that comprises independent,
significant and self-contained systems.

Fundamental differences between conventionally engineered systems and the

type of system described above limit scientific and engineering knowledge, there-
by constraining our ability to implement effective control structures for the
global financial markets [5]. It may therefore be impractical to demand indi-
vidual accountability in the context of failures that are effectively systemic [19].
A more deliberate strategy would be to develop techniques and tools that can
address underlying issues affecting the markets including emergent behavior;
heterogeneity via organic growth; normal failure; and scaling.

4 Liquidity
Arbitrage enables HFTs to profit from pricing inefficiencies—expressed as wide
bid–offer spreads—that are inherent in a fragmented market structure. In the
process, HFT systems reduce those spreads by allocating liquidity—the ability
to buy or sell an asset without significantly impacting its price—efficiently across
multiple trading venues [5, 13]. As spreads narrow transaction costs are reduced,
and prices become more consistent, efficient and transparent.
The issue of liquidity is complicated by HFT activity and strategies; market
fragmentation; and the concept of phantom, or transient, liquidity [5]. Phan-
tom liquidity is a consequence of high-frequency quoting, which increases order
to trade ratios (i.e., the fraction of submitted orders resulting in a trade) [21].
Due to the inherent speed of high-frequency quoting, phantom liquidity is un-
available to average investors. Liquidity complications force investors to access
liquidity either via expensive technology or brokers that will assume the costs
of trading [5]. This requirement disadvantages small institutional and private
investors who may not be able to afford the technological arms race.
Despite the fact that computer trading has improved liquidity in general,
HFTs are accused of withdrawing liquidity during volatile markets [5, 13, 21].
This type of investor behavior can create feedback loops with potential to am-
plify undesired interactions and outcomes [5, 13]. According to Brogaard, there
is no evidence to suggest that HFTs withdraw from markets during periods of
volatility [5]. But because HFTs constitute primary liquidity providers, their
modus operandi—including limited capital and ultra-fast execution—can en-
gender periodic illiquidity.

5
4.1 High-Frequency Quoting
HFTs achieve narrow bid–offer spreads by means of high-frequency quoting, a
technique that strains pre- and post-trade IT infrastructure—including trading,
clearing and settlement systems—with traffic volumes that significantly exceed
trade message traffic [8]. Figure 3 illustrates growth in quote message traffic
for the Consolidated Quote System (CQS), the official electronic data stream
for real-time quote information generated by U.S. stock exchanges. Conversely,
Figure 4 illustrates lack of growth in trade message traffic for the Consolidated
Tape System (CTS), counterpart to the CQS for real-time trade information.
CQS capacity increases are promptly equalized by growth in quote mes-
sage traffic [10]; this phenomenon is illustrated by Figure 3 with respect to
the CQS, where gaps between groups of lines indicate quote message traffic
increases. When capacity limits are tested in this manner, latencies resulting
from strained infrastructure could interfere with the efficient operation of the
markets; for example, pricing and trade settlement delays could obstruct clear-
ing and settlement systems [13]. This type of systemic risk, which may have
contributed to the events of May 6 [22], is discussed further in Section 6.2.

5 Volatility
Economic research provides no direct and unambiguous evidence of causality be-
tween HFT and increased volatility, which is asset price variability over time [5].
If HFT contributed to volatility, then HFT diffusion should have increased the
intraday to overnight volatility ratio; but this correlation is not evident. On
the contrary, increased HFT activity has been associated with decreased short-
term volatility. According to Chaboud et al., algorithmic trading lowers realized
volatility in FX markets; and Brogaard concludes that HFT may have reduced
volatility on the Nasdaq [13]. While HFT has also been associated with the
events of May 6, which constitute a period of extreme volatility, causality be-
tween HFT and the events of May 6 has not been established.

6 May 6, 2010
During May 6, major equity indices in futures and securities markets, and al-
most 8,000 individual equity securities and exchange traded funds, experienced
severe price fluctuations [23]. The Dow Jones Industrial Average declined 998.50
points—the biggest intraday point drop in its history—before recovering losses
to close down 347.80 points, or 3.2%, at 10520.32 points. Approximately 20,000
trades—across more than 300 securities—were executed at devaluations of over
60%, with many of those trades approaching upper (as high as $100,000) and
lower (less than a penny) equity price limits.

6
Figure 3: An illustration of CQS two millisecond peak quote rates for the period
between February 2008 and February 2012, as identified by the legend. The x
and y axes represent market trading hours and quotes per second, respectively.

7
Figure 4: An illustration of CTS two millisecond peak quote rates for the period
between February 2008 and February 2012.

8
6.1 The CFTC–SEC Report
The U.S. Commodity Futures Trading Commission (CFTC) and the U.S. Se-
curities and Exchange Commission (SEC) issued a joint report regarding the
events of May 6. These events can be summarized as follows.
Even before the flash crash, May 6 was an unusually turbulent trading day
characterized by decreasing liquidity, high volatility and unsettling news re-
garding the European sovereign debt crisis. At 2:32 p.m., a large fundamental
trader —a market participant trading to accumulate or reduce a net long/short
position—initiated an automated algorithm to sell 75,000 E-Mini S&P 500 (E-
Mini) contracts valued at approximately $4.1 billion. The algorithm’s execution
rate—the rate at which the algorithm submits orders to the market—was set
at 9% of the trading volume calculated over the previous minute, disregarding
price and time.
The algorithm submitted an initial batch of orders, which was likely absorbed
by HFT systems and intermediaries. Between 2:41 p.m. and 2:44 p.m., increased
HFT activity—including aggressive selling—prompted the algorithm to increase
its own rate of execution. Selling pressure from the algorithm, cross-market
arbitrageurs, HFT systems and other traders caused the E-Mini and the S&P
500 SPDR (SPY) to devaluate. Against a backdrop of vanishing liquidity and
declining indices, HFT systems generated a hot potato volume effect, and acute
price fluctuations, by rapidly trading contracts with each other. Following a five
second trading pause, which was triggered by the Chicago Mercantile Exchange
Stop Logic Functionality in order to prevent further price declines, buy-side
interest increased and the E-Mini and SPY began to recover.
The CFTC–SEC report attributes May 6 to a conjuncture of factors includ-
ing the automated execution of a large sell order during stressed market condi-
tions; the unreliability of trading volume as an exclusive liquidity indicator; the
coupling of derivatives and securities markets; the failure of cross-market circuit
breakers; and liquidity erosion resulting from 1) interactions between automated
execution programs and algorithmic trading strategies, and 2) the simultaneous
withdrawal of multiple market participants.

6.2 The Nanex Theory

Nanex—a financial data vendor and research firm—presents a markedly different
point of view [22].
Nine U.S. exchanges—including BATS, CBOE, CHX, ISE, NASDAQ, NAS-
DAQ OMX BX, NSX, NYSE and NYSE Arca—submit bids and/or offers for
NYSE listed stocks. The highest bid and lowest offer prices become, respec-
tively, the best available bid and best available offer, collectively known as the
National Best Bid and Offer (NBBO) for U.S. equities. Each exchange avoids
crossing competing exchanges (i.e., bidding higher than other exchanges are
offering, or offering lower than other exchanges are bidding) to limit arbitrage
opportunities.
At 2:42:46 p.m., NYSE quotes began to queue, and thereby fall behind quotes

9
from competing exchanges. This delay caused NYSE bids to cross above the best
available offer prices for approximately 250 NYSE listed stocks. HFT systems
attempted to exploit this arbitrage opportunity by directing sell order flow to
NYSE. While delayed lower priced quotes had yet to be disseminated, sell orders
arriving at NYSE were executed against lower priced buy orders from NYSE’s
designated market maker order book. Because 1) trades and quotes are reported
to different data streams (CTS and CQS, respectively), and 2) quote message
traffic significantly exceeds trade message traffic (as discussed in Section 4.1),
trade data was disseminated with little or no delays. HFT systems detected the
sudden price drop and automatically went short, thereby creating an amplifying
feedback loop.
Nanex asserts that the events of May 6 were caused by delays in the dissem-
ination of NYSE quotes [22]. However, this causation has not been proven, nor
is it accepted by the CFTC–SEC Report [23].

6.3 Market Resilience

Reliance on HFT systems to function as primary liquidity providers has posi-
tive and negative implications for market resilience. During May 6 and other
comparable events, equity prices experienced rapid declines and unprecedented
volatility; but losses were reversed by equally rapid and unprecedented recov-
eries. This type of resilience—the speed with which prices returned to their
original levels—was not evident when HFTs were absent from stressed markets;
for example, Monday October 19, 1987 [13]. It is also important to note that
HFT strategies were largely unaffected by credit constraints resulting from the
financial crisis of 2007–2008. Consequently, HFTs remained relatively active,
and thereby continued to provide liquidity, during the extended market dislo-
cation of that period.

7 Recommendations
The consensus among regulators and market participants is that HFT activity
must not be interrupted. HFTs should be permitted to operate in the markets
as long as they do not engage in abusive or parasitic activities that undermine
market quality. We advocate that as stakeholders and beneficiaries of a powerful
technology, HFTs must assume the responsibility of developing and deploying
that technology in an ethical, risk-averse and transparent manner. To mitigate
the potential for systemic failures described in Section 3, HFTs should adopt
techniques that are specific to the development of high-integrity safety-critical
systems. HFTs should also implement transparent systems that can support
the regulatory process.
Because regulation compounds the risks that HFTs need to manage, reg-
ulators should consider that even minor regulatory adjustments can give rise
to significant software development costs and capital risk. Regulation should
therefore proceed at a consistent and predictable pace that affords HFTs suffi-

10
cient time for compliance. To this effect, regulators should engage with market
participants, while resisting politically motivated and media induced pressure.
Regulators should also strive to understand electronic trading and HFT by
means of empirical data, and not presume all HFT activity to be malicious.
While unregulated and untested algorithms are cause for concern, rogue activ-
ities are not limited to HFT. The solution to persistent rogue behavior is not
the elimination of HFT altogether, but rather the implementation of appropri-
ate monitoring procedures and safeguards. From a technological perspective,
market abuse detection systems need not be prohibitively expensive nor dis-
ruptive; for example, data tagging—as required by the Markets in Financial
Instruments Directive II (MiFID II)—can realize greater transparency by sup-
porting the sophisticated analysis of large and complex data sets.
The importance of data in an HFT environment cannot be overstated [23].
A fair and orderly price discovery process requires robust and timely access to
market data for all market participants, whether human or machine. To this
end, exchanges must ensure the integrity and reliability of information systems
that publish trades and quotes to consolidated market data feeds.

8 Conclusions
The complex technologies and continuous technological advancements that un-
derpin HFT impose significant demands on legislators, regulators and market
participants. Regulators must consider specialized, non-financial knowledge in
order to establish effective financial regulation. HFTs must consider the severe
consequences of irresponsible and harmful behavior on the financial industry,
and society in general. But at present, the focus is on impending regulation.
If regulators arrive at erroneous conclusions regarding HFT, then regulatory
action could damage a technology that has been generally beneficial for the
markets.
Our aim has been to review the opinions of academics, market participants
and regulators, with respect to HFT benefits and risks. By understanding the
risk factors associated with HFT, and the impact of HFT activity on financial
stability, stakeholders can realize risk management practices that promote HFT
benefits while reducing social costs.

Acknowledgements
This report is sponsored by the Association for Financial Markets in Europe
(AFME); the Professional Risk Managers’ International Association (PRMIA);
and the UCL Financial Industry Series (UCL FIndS). The authors would like
to thank PRMIA’s Board of Directors for their input and support. The authors
would also like to thank—in alphabetical order—the following individuals for
their contribution.

11
 − Anthony Kirby (Director of Regulatory Reform and Risk Management at
Ernst & Young)
− Clive Furness (Managing Director at Contango Markets)
− Evangelos Benos (Research Economist at Bank of England)
− Jasper Jorritsma (Policy Officer at the European Commission)
− John Mueller (Managing Director at Getco Europe)
− Jonathan Howitt (PRMIA Publications Committee)
− Kee-Meng Tan (Managing Director at Knight Capital Europe)
− Mike O’Hara (Co-Founder & CEO at HFT Review)
− Paul Hinton (Partner at Kemp Little LLP)
− Sam Tyfield (Partner at Katten Muchin Rosenman UK LLP)
− Tim Rowe (Manager of Trading Platforms and Settlement Policy at the
Financial Services Authority)
− Walter Hendriks (Principal Advisor at Financial Markets Advisory)
− William Rhode (Principal, Director of Fixed Income at TABB Group)

Any conclusions, findings, opinions and recommendations expressed in this

report are those of the author(s) and do not necessarily reflect the views of
AFME, PRMIA, UCL FIndS or the contributors listed above. Additional HFT
resources can be found in an online repository.

References
[1] I. Aldridge, High-Frequency Trading: A Practical Guide to Algorithmic
Strategies and Trading Systems. John Wiley & Sons, Inc., 2010.
[2] London Stock Exchange Group, “London Stock Exchange Group Response
to CESR Call for Evidence on Micro-Structural Issues of the European
Equity Markets (CESR/10-142).” [online] Available at: http://bit.ly/
14AAHXq, April 2010.
[3] G. Nuti, M. Mirghaemi, P. Treleaven, and C. Yingsaeree, “Algorithmic
Trading,” Computing Now, 2011.
[4] V. (Willy), “The Value of a Millisecond: Finding the Optimal Speed of a
Trading Infrastructure.” TABB Group, April 2008.
[5] The Government Office for Science, London, “Foresight: The Future of
Computer Trading in Financial Markets.” [online] Available at: http://
bit.ly/UvGE4Q, 2012.

12
 [6] S. Lee, “New World Order: The High Frequency Trading Community and
Its Impact on Market Structure.” Aite Group, LLC, February 2009.
[7] P. Gomber, B. Arndt, M. Lutat, and T. Uhle, “High-Frequency Trading.”
[online] Available at: http://ssrn.com/abstract=1858626, 2011.
[8] Nanex, “The Rise of the HFT Machines.” [online] Available at: http://
www.nanex.net/aqck/2804.html, 2012.
[9] K.-H. Bae, H. Jang, and K. S. Park, “Traders’ Choice Between Limit and
Market Orders: Evidence From NYSE Stocks,” Journal of Financial Mar-
kets, vol. 6, pp. 517–538, August 2003.
[10] Nanex, “HFT Is Insatiable.” [online] Available at: http://www.nanex.net/
research/cqscap/cqscap.html, 2011.
[11] A. A. Kirilenko, A. S. Kyle, M. Samadi, and T. Tuzun, “The Flash Crash:
The Impact of High Frequency Trading on an Electronic Market.” [online]
Available at: http://ssrn.com/abstract=1686004, May 2011.
[12] Nanex, “The Other Flash Crash.” [online] Available at: http://www.
nanex.net/research/29sep2008/29sep2008.html, 2012.
[13] W. Barker and A. Pomeranets, “The Growth of High-Frequency Trading:
Implications for Financial Stability.” [online] Available at: http://bit.
ly/VqDBax, June 2011.
[14] D. Easley, M. M. L. de Prado, and M. O’Hara, “The Microstructure of
the ‘Flash Crash’: Flow Toxicity, Liquidity Crashes and the Probability of
Informed Trading,” The Journal of Portfolio Management, vol. 37, pp. 118–
128, November 2010.
[15] D. G. I. Market and E. C. Services, “Review of the Markets in Financial
Instruments Directive (MiFID).” [online] Available at: http://bit.ly/
gAj3gl, December 2010.
[16] P. Kumar, M. Goldstein, F. Graves, and L. Borucki, “Trading at the
Speed of Light: The impact of High-Frequency Trading on Market Perfor-
mance, Regulatory Oversight, and Securities Litigation.” [online] Available
at: http://www.brattle.com/_documents/uploadlibrary/upload932.
pdf, 2011.
[17] J. Bollen, H. Mao, and X. Zeng, “Twitter Mood Predicts the Stock Market,”
Journal of Computational Science, vol. 2, pp. 1–8, March 2011.
[18] P. A. Hafez and J. Xie, “Factoring Sentiment Risk Into Quant Models.”
[online] Available at: http://ssrn.com/abstract=2071142, 2012.
[19] D. Cliff and L. Northrop, “The Global Financial Markets: An Ultra-Large-
Scale Systems Perspective.” [online] Available at: http://bit.ly/GDhfuV,
2010.

13
[20] P. Feiler, R. P. Gabriel, J. Goodenough, R. Linger, T. Longstaff, R. Kaz-
man, M. Klein, L. Northrop, D. Schmidt, K. Sullivan, and K. Wallnau,
Ultra-Large-Scale Systems: The Software Challenge of the Future. Carnegie
Mellon Software Engineering Institute, 2006.
[21] T. Rijper, W. Sprenkeler, and S. Kip, “High-Frequency Trading.” [online]
Available at: http://www.optiver.com/corporate/hft.pdf, 2010.
[22] Nanex, “Analysis of the “Flash Crash”.” [online] Available at: http://www.
nanex.net/20100506/flashcrashanalysis_intro.html, 2010.
[23] U.S. Commodity Futures Trading Commission and U.S. Securities and Ex-
change Commission, “Findings Regarding the market Events of May 6,
2010.” [online] Available at: http://www.sec.gov/news/studies/2010/
marketevents-report.pdf, 2010.

14