• No results found

Price discovery around monetary policy news : algorithmic trading in the foreign exchange market

N/A
N/A
Protected

Academic year: 2021

Share "Price discovery around monetary policy news : algorithmic trading in the foreign exchange market"

Copied!
62
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

Price discovery around monetary policy news:

Algorithmic trading in the foreign exchange market

An MSc Thesis, University of Amsterdam

Boudewijn Horikx

November 21, 2014

Abstract

In this thesis, I analyze how the presence of algorithmic traders affects price discovery in foreign exchange with regard to monetary policy news. Algorithmic news-traders can respond to monetary policy news within (milli)seconds upon release. However, they only process information that can be automatically quantified. Some formats of central bank communication are more apt for automatic quantification than others. Discretionary human traders incorporate the part of monetary policy news that can not be automatically quantified into currency prices, during which they interact with high-frequency traders. In this capacity, algorithmic traders play a secondary role in price discovery around monetary policy news. Algorithmic traders can increase price efficiency in the aftermath of monetary policy news by directly trading on the news faster than human traders, by anticipating and trading ahead of informed human order flow, or by engaging in inter-market and cross-market arbitrage.

I am grateful to my supervisor, prof. dr. Frank de Jong, for useful suggestions throughout the writing of this

thesis. I am also grateful to dr. Ward Romp and dr. Gabriele Galati for steering me in the right direction during the early stages of writing. Finally, my gratitude goes out to Roderik Schillemans, who provided valuable support with graphic illustrations.

(2)

2

Table of Contents

Introduction ... p. 3

1. Monetary policy news ... 6

1.1 Central bank communication formats ...6

1.2 Market impact ...9

1.3 Other sources of monetary policy news ... 12

2. Algorithmic news-trading ... 14

2.1 Low latency ... 15

2.2 Algorithmic processing ... 16

2.3 Media-driven algorithmic trading ... 21

2.4 Graphic presentation ... 21

2.5 Choices in central bank communication ... 24

3. High-frequency trading and information input ... 25

3.1 Defining HFT ... 25

3.2 The scope of HFT ... 26

3.3 Evidence from the equity market ... 29

3.4 Graphic presentation ... 37

4. HFT in the foreign exchange market ... 40

4.1 Market structure ... 40 4.2 Platform characteristics ... 42 4.3 Order flow ... 45 4.4 Triangular arbitrage ... 46 5. A model approach ... 49 5.1 Schematic overview ... 49

5.2 Fast and slow traders ... 53

Conclusion ... 59

(3)

3

Introduction

Speed has always been important with regard to trading in financial instruments. If a trader can respond to market-moving information faster than other traders, he is clearly in a better position to capitalize on the new information. In 1815, Nathan Rothschild was the first trader in London to receive the news that Napoleon had lost the battle of Waterloo (present-day Belgium). Rothschild received the news from one of his carrier pigeons twenty-four hours earlier than other traders, who were waiting on a horse-bound carrier. As he knew the news would increase the value of British government bonds, Rothschild was able to make a small fortune on his speed advantage (Steiner, 2012: 121).

Speed in financial trading is more formally defined as latency, which denotes how long it

takes for an investor to process new information and respond to it by placing the appropriate trades. Computer trading algorithms have redefined latency. These computer programs are designed by quantitative traders, or quants. Their input consists of quantified information that has some

predictive power for future asset price fluctuations. Their output consists of a trade response meant to capitalize on the predicted price fluctuations. As this process is automated, inputs are converted into outputs within a second, usually measured in milliseconds. Therefore, with the emergence of computer trading algorithms, or algorithmic traders, latency has decreased dramatically. Due to their speed advantage, algorithmic traders are currently a significant presence in financial markets.

The emergence of algorithmic trading has been accompanied by growing concerns from traditional traders, regulators and the media. They fear that some forms of algorithmic trading are detrimental to market quality. Their concerns were justified by a high-impact event that occurred on May 6, 2010, known as the Flash Crash. In a period of 20 minutes, the Dow Jones Industrial Average dropped and recovered by about 1000 points (9%), the biggest intra-day point decline in history. As of yet, there is no consensus on what caused this dramatic price fluctuation. During the Flash Crash, some stocks traded for a single penny, while they were trading for $30 to $40 in normal times. Although the crash may not have been caused by algorithmic traders, the speed of events that day would not have been possible without algorithmic trading. It is the uncertainty regarding the causes of the Flash Crash, rather than the Flash Crash itself, that fuel concerns about the effects of

algorithmic trading on financial markets. How is it possible that nearly a trillion dollars’ worth of equity disappears in minutes without an immediate cause? Algorithmic trading might have increased

(4)

4 systemic risk by creating fatter tails, i.e., a greater possibility of extreme market-wide price

fluctuations.1

Academic investigation of how algorithmic trading affects market quality is hindered by limited data availability. Algorithmic trading regularly involves hundreds of orders in a single second, most of which are cancelled. Academics are therefore in need of scarce high-frequency data,

preferably tick-by-tick. Furthermore, it must be identified in the data whether the trader that sends an order is an algorithmic trader or a discretionary human trader. This information is proprietary to the platforms where trading takes place. Notable empirical contributions include Brogaard,

Hendershott and Riordan (2013), who study price discovery in U.S. equities with respect to high-frequency trading (a form of algorithmic trading), and Chaboud, Chiquoine, Hjalmarsson and Vega (2013), who study algorithmic trading in the foreign exchange market. Theoretical academic

contributions usually emphasize the speed advantage of algorithmic traders. Foucault, Hombert and Roşu (2014) consider the case in which a speculator is better informed and faster than the dealer he is trading with.

My contribution to the existing literature is twofold. First, I consider algorithmic trading in the foreign exchange (FX) market. Relative to algorithmic equity trading, literature regarding

algorithmic FX trading is scarce. Notable exceptions include Chaboud et al. (2013) and a report by the Bank of International Settlements (2011). Existing literature regarding algorithmic FX trading does not explicitly distinguish between the effects and viability of different algorithmic trading strategies. I will determine to what extend algorithmic strategies found in equity trading are applicable to FX trading. Second, I consider monetary policy news as a source of information input for algorithmic traders. Existing literature provides evidence that algorithmic traders trade aggressively on the release of macroeconomic news announcements. By means of a literature review, I provide

suggestive evidence that algorithmic traders are also able to trade directly on the release of various forms of central bank communication.

I consider monetary policy news a relevant source of information to FX traders for the following reason. Interest rates, determined by monetary policy, influence the return of assets denominated in different currencies. A change in the domestic interest rate, changes the value of the domestic currency with respect to foreign currency. Therefore, monetary policy determines the value of exchange rates.2 When monetary policy news is released, it takes investors some time to

incorporate the new information into currency prices, i.e., the foreign exchange market is not

1

See the speech by Andrew Haldane, Executive Director, Financial Stability, of the Bank of England, at the International Economic Association Sixteenth World Congress, Beijing, July 8, 2011.

(5)

5 completely efficient in the immediate aftermath of monetary policy news. The analysis in this thesis revolves around the question of how low-latency algorithmic traders affect price discovery in the aftermath of monetary policy news in the foreign exchange market. Because algorithmic traders are able to process and trade on new information faster than human traders can, I conjecture that algorithmic traders increase price efficiency around monetary policy news.

This thesis is organized as follows. In chapter one, I will analyze the nature of monetary policy news. This entails a discussion of empirical evidence regarding its impact on the foreign exchange market. In chapter two, I will discuss how algorithmic news-traders can automatically quantify monetary policy news and subsequently trade on it with minimal latency. In chapter three, I discuss high-frequency trading. In this capacity, algorithmic traders respond to monetary policy news

indirectly by trading on information that has already entered the market. I provide evidence from the equity market with respect to different high-frequency trading strategies. In chapter four, I extend these findings to the foreign exchange market. In chapter five, I discuss how the results from chapter one to four fit in a theoretical model of informed trading. The final segment contains concluding remarks.

(6)

6

1.

Monetary policy news

For simplicity’s sake, I assume that monetary policy for a given country is set by an independent central bank. For developed countries, this is probably a fair assumption. Central banks reveal their monetary policy decisions and intentions to the public using various communication formats, like written statements and press conferences. These communication formats constitute monetary policy news. As monetary policy significantly determines the value of currencies, they are an important source of information to agents in the foreign exchange market. The significance of a news shock can be measured by its market impact; the response of currency returns, volatility and volume. In this chapter, I assess the nature and the market impact of monetary policy news, paying special attention to different communication formats. Throughout this assessment, I will focus on monetary policy communication by the Federal Reserve System, the central banking system of the United States.

1.1

Central bank communication formats

Within the Federal Reserve System (Fed), the Federal Open Market Committee (FOMC) is responsible for monetary policy via open market operations. Therefore, FOMC communication is the main source of official U.S. monetary policy news. The Fed’s website states:

‘The FOMC holds eight regularly scheduled meetings per year. At these meetings, the Committee reviews economic and financial conditions, determines the appropriate stance of

monetary policy, and assesses the risks to its long-run goals of price stability and sustainable growth’3

Immediately following these meetings, the Committee communicates its decisions and intentions for future monetary policy to the general public in a written statement.

In accordance with a worldwide trend, the nature of FOMC communication has changed radically over the past two decades, from being quite secretive in its deliberations to being very transparent (Wynne, 2013). The February 1994 FOMC meeting was the first that was accompanied by a post-meeting statement that clarified the Committee’s decisions, albeit very concise. Before, market participants had to determine what decisions the FOMC had made on their own, by inducing the operations of the Open Market Desk based on asset price movements. In 1999, the Committee started to include forward guidance to post-meeting statements. This entails an assessment of risks to the Committee’s policy goals. This assessment basically provides an outlook for the future stance of monetary policy and it has become increasingly informative over the last 14 years. The statement

(7)

7 following the December 2012 meeting marked an important point in this regard. In this statement the Committee linked future changes to the target range for the federal funds rate to specific levels of future macroeconomic variables, such as the unemployment rate and inflation. The need for more transparency in FOMC communication has been fueled by the recent financial crisis. The

exceptionally low level of the federal funds rate has increased the importance of forward guidance as a monetary policy tool (Wynne, 2013). The notion that FOMC statements have become increasingly informative can be illustrated by an increase in statement word counts (Figure 1). The shaded area depicts the period of unconventional monetary policy with interest rates at the effective lower bound of near zero.

Figure 1. FOMC statement word counts (1994-2013)

Note:

Source: Wynne (2013)

Currently, the FOMC uses a range of other communication formats besides post-meeting statements. Three weeks after meetings, the FOMC releases the meeting’s minutes. These minutes provide additional insights into the thought process behind current policy decisions and intentions for future policy decisions. They include assessments of individual Committee members. Also, four out of eight meetings per year are followed by a press conference by the Committee’s Chair. In the aftermath of these four meetings, the FOMC provides extensive projections materials, which contain quantitative economic projections for key macroeconomic variables, like changes in real GDP, the

(8)

8 unemployment rate and inflation. They further explain to the public on which information the

Committee bases its decisions and intentions.45

Other central banks use variations of the communication formats used by the FOMC. For example, following monthly meetings of the Governing Council, the ECB issues a press release in which monetary policy decisions are revealed. This brief written statement is followed by a press conference by the ECB’s president and vice-president.6 This press conference is similar to the written FOMC statement in that is provides a rationale for current policy decisions and provides an outlook for future policy decisions. The ECB does not provide minutes for Governing Council meetings. The Bank of England’s Monetary Policy Committee (MPC) also meets every month after which it releases its decisions in written form. Minutes of MPC meetings are released before subsequent meetings. In addition, the Bank of England provides forward guidance statements and a quarterly Inflation Report, which is followed by a press conference by the governor.7

4

Future FOMC meeting dates and all discussed communication formats - statements, minutes, projections materials and video recordings of press conferences - are available on the Fed’s website;

http://www.federalreserve.gov/monetarypolicy/fomccalendars.htm. 5

The Fed communicates in several other ways with the public regarding the stance of U.S. monetary policy. The Federal Reserve Board submits a Monetary Policy Report to Congress semiannually, containing a discussion of the conduct of monetary policy and economic developments and prospects for the future. The report is accompanied by a testimony from the Federal Reserve Board’s Chair. The Fed also issues a Quarterly Report on

Federal Reserve Balance Sheet Developments and a Summary of Commentary on Current Economic Conditions

by Federal Reserve District, commonly known as the Beige Book. An analysis of these formats is beyond the scope of this thesis.

6

See http://www.ecb.europa.eu/mopo/decisions/html/index.en.html.

(9)

9

1.2

Market impact

Monetary policy is a significant determinant of exchange rates and this is illustrated by the market impact of monetary policy news. It is primarily the unanticipated part of central bank

announcements that have a market impact and constitute news. In this segment I will provide quantitative measures for the impact that central bank communication has on intra-day exchange rate returns, volatility and volume. I will differentiate between various communication formats.

FOMC decisions and statements

Carlo Rosa (2011) empirically analyzes the high-frequency response of exchange rates to U.S. monetary policy news. In measuring news, he differentiates between the unanticipated component of FOMC decisions and FOMC statements.8 To determine how exchange rates react to these two communication formats, Rosa uses high-frequency intra-day data with 5-minute intervals for five U.S. dollar currency pairs from 1999 to 2007. For the sake of statistical computation, two news indicators are constructed that capture the surprise component of both FOMC decisions and statements. Rosa then regresses exchange rate returns for the U.S. dollar against the five other currencies on these two indicators. He finds that the surprise components of both FOMC decisions and statements have large and highly significant effects on U.S. dollar currency pairs. He concludes from the regression results:

‘…an unanticipated 25-basis-point cut in the federal funds target rate is associated on average with a 0.5% depreciation of the exchange value of the dollar against foreign currencies. An unexpected downward revision in the tone of the statement from neutral to dovish is associated with about 0.3% decline in the US dollar. However, the news stemming from Fed’s statements matters more for the determination of exchange rates than news about actual monetary policy

decisions.’9(Rosa 2011: 488)

With regard to bond yields and stock prices, Gürkaynak, Sack and Swanson (2005) find a similar distinction between the market impact of FOMC decisions and FOMC statements. They interpret this result by noting:

8

To be clear, FOMC statements contain the policy decisions that the Committee decided on in the preceding meeting. In this context, FOMC statements refer to the rationale for policy decisions and the outlook for future policy.

9

This latter point can be illustrated by an increase of the adjusted 𝑅2 from 0,029 to 0,149 when the news indicator for FOMC decisions is augmented by the news indicator for FOMC statements, in a regression of dollar exchange rate returns against a constant and these two news indicators (Rosa, 2011: table 4).

(10)

10

‘[The] results do not indicate that policy actions are secondary so much as that their influence comes earlier – when investors build in expectations of those actions in response to FOMC statements (and perhaps other events, such as speeches and testimony by FOMC members).’ (Gürkaynak, Sack

and Swanson, 2005: 87)

Rosa (2011) further finds that the volatility of U.S. dollar exchange rate returns is higher for about 40 min to 1 hour in the aftermath of an FOMC statement release and that both news regarding FOMC decisions and statements is fully incorporated into currency prices in about 30 to 40 minutes. Finally, Rosa puts the market impact of monetary policy news in perspective by remarking that monetary policy news shocks can only explain about 15% to 22% of the variation in U.S. dollar exchange rate pairs in a narrow window surrounding the news event.

Fischer and Ranaldo (2011) study the volume response of the global foreign exchange market to FOMC post-meeting communication. Their results suggest that turnover in the spot and spot-next market increases on FOMC meeting days by about 5% between 2003 and 2007.10 Furthermore, their results suggest that global currency volume always increases on FOMC meeting days, irrespective of the size of unanticipated policy shocks.

Minutes of FOMC meetings

Rosa (2013b) examines to what extent minutes of FOMC meetings contain market-moving

information by looking at asset price volatility and trading volume in a narrow window around the release of the minutes. For a sample ranging from January 2005 to March 2011, he finds:

‘The release of the minutes induces significantly ‘’higher than normal’’ volatility on asset prices, especially at the time of the release, and up to roughly one hour after the announcement. […] Treasuries, especially at shorter maturities, are the most affected asset class, closely followed by U.S. dollar exchange rates, […]. This finding indicates that FOMC minutes provide market-relevant

information that is incorporated into asset prices.’ (Rosa, 2013b: 71)

To illustrate the significance of the minutes, Rosa (2013b) states that their average market impact is similar to the one following ISM manufacturing index releases, but smaller than the one following the release of the actual FOMC statement. Finally, he notes that the asset price response to FOMC minutes has in general declined since 2008, suggesting greater transparency of the FOMC.

(11)

11 The notion that minutes from FOMC meetings can contain significant market moving

information was demonstrated on October 8, 2014. On this date, minutes were released in which a number of committee members expressed particular concerns about a weak global economy. This sentiment was not expected by agents in financial markets and they induced from this news that interest rates are going to stay close to zero for longer than they had previously anticipated. In the aftermath of the minutes’ release, U.S. stocks rallied and the Dow Jones Industrial Average had its best day of the year to-date, closing nearly 2% higher.11 Other financial instruments like gold and foreign exchange, had similar significant responses. So although the actual statement was released three weeks prior, these minutes contained information that moved financial markets in spectacular fashion.

ECB decisions and press conferences

Instead of post-meeting statements, the ECB uses post-meeting press conferences to clarify

monetary policy decisions and provide an outlook on the future stance of monetary policy. Using an approach very parallel to his 2011 study, Carlo Rosa (2013a) studies the intra-day market impact of monetary policy news stemming from ECB decisions and press conferences on euro exchange rates. Based on a 1999 to 2009 sample, he concludes:

‘…the surprise component of communication (in the form of press conferences, ed.) has highly significant effects on exchange rates, whereas the response of the euro to unanticipated changes in the policy rate is more muted. For instance, a hypothetical positive news shock (based on

information in the press conference, ed.) of 100-basis-points is associated with an appreciation of the

euro against the US dollar of roughly 3.6%.’12(Rosa, 2013a: 168)

Regarding their market-impact, ECB press conferences are thus more important than actual ECB decisions, similar to how FOMC statements are more important than actual FOMC decisions. It takes about 1 hour before asset markets have fully incorporated the ECB news shocks.

11

See ‘Stocks recover losses, close up nearly 2% on fed minutes’ by Evelyn Cheng, 8 October 2014, on CNBC’s website; http://www.cnbc.com/id/102070077.

12 The relative market-moving importance of ECB press conferences compared to ECB decisions can be illustrated by an increase of the adjusted 𝑅2 from 0,007 to 0,180 when the news indicator for ECB decisions is augmented by the news indicator for ECB press conferences, in a regression of euro-dollar exchange rate returns against a constant and these two news indicators (Rosa, 2013a, table2).

(12)

12 1.3

Other sources of monetary policy news

In what is known as ‘the Bartiromo affair’ among traders, in April 2006 equity and bond markets reacted strongly following a report of CNBC anchor Maria Bartiromo on alleged remarks by then Fed chairman Ben Bernanke regarding the stance of monetary policy. Bernanke apparently told

Bartiromo at the White House Correspondents Dinner that financial markets ‘misunderstood’ his congressional testimony.13 In reaction to the referred testimony, bonds and stocks rallied. In contrast to official FOMC statements, his remark at the dinner was made in an informal setting. This didn’t withhold the S&P 500 index from dropping 0,8% and Treasury bond yields from jumping to four-year highs immediately following the report, a correction relative to the rally that followed the

congressional testimony. This anecdote illustrates that monetary policy news is not restricted to official central bank communication.

Furthermore, assume that central banks employ an implicit Taylor rule, where interest target rate decisions are a function of GDP/unemployment and inflation. Also assume that agents in

financial markets learn about the coefficients of this Taylor rule through central bank

communication. More transparent communication enables market agents to stricter define the implicit Taylor rule. For example, the December 2012 FOMC statement contained the phrasing:

‘[…] the Committee decided to keep the target range for the federal funds rate at 0 to 0,25

percent and currently anticipates that this exceptionally low range for the federal funds rate will be appropriate at least as long as the unemployment rate remains above 6,5 percent, […]’14

In this case, the Committee has thus linked future monetary policy rather strictly to the unemployment rate. Now consider the hypothetical situation that the unemployment rate unexpectedly drops below 6,5%. This drop would have immediate consequences for the expected stance of U.S. monetary policy. More realistically, a revision of future unemployment rate

expectations has implications for the expected future stance of monetary policy, which has a price impact today. Therefore, news about the unemployment rate indirectly constitutes monetary policy news. In general, if a central bank links policy to a set of macroeconomic variables, changes in these variables entail monetary policy news. As GDP growth, unemployment and inflation are all

determined by a wide variety of underlying factors, the set of variables that indirectly entail monetary policy news can grow quite large.

13

See Jennifer Hugh, ‘Bernanke remarks leave analysts bemused’, Financial Times, May 2, 2006; http://www.ft.com/intl/cms/s/0/034415c0-da1d-11da-b7de-0000779e2340.html#axzz3BssWdrTG. 14 See http://www.federalreserve.gov/newsevents/press/monetary/20121212a.htm.

(13)

13 In conclusion, the primary source of monetary policy news is official central bank communication. There are different formats in which central banks reveal their decisions and intentions to the public and central banks generally use a combination. In the next chapter, I will argue that not every format is equally apt for processing by algorithmic traders. Also, some formats have a larger market impact than others. Statements that contain an outlook on the future stance of monetary policy are, on average, more important than actual policy decisions. As post-meeting statements have become more informative, minutes of FOMC meetings have generally become less important, but can still have a significant market impact. Finally, unofficial remarks by influential policy makers and unexpected changes in relevant macroeconomic variables are an additional source of monetary policy news.

(14)

14 2.

Algorithmic news-trading

In the previous chapter, I have analyzed how central banks communicate their monetary policy decisions and intentions to the public. I have also discussed how this communication has an impact on currency returns, volatility and volume. In an efficient market, new information is incorporated into asset prices immediately. However, there is compiling evidence that for intra-day data, this is not a fair assumption. Evans and Lyons (2007) describe how new information about macroeconomic fundamentals is not embedded into currency prices instantaneously. Market agents observe news asymmetrically and their heterogeneous beliefs are aggregated into exchange rates over time. A sharp currency rate revaluation in the immediate aftermath of the news release - say in the first 10 minutes - is followed by a period of increased volatility in which various market agents change their positions based on their view of the new information. As discussed in the first chapter, Rosa (2011, 2013a) finds that monetary policy news is fully incorporated into currency prices in roughly an hour.

Algorithmic traders have the ability to process new information, determine the appropriate trade response and send orders into the market, in less than a second. But does ‘new information’ include monetary policy news? Algorithmic traders are generally very secretive about their

strategies. Therefore, evidence on the nature of algorithmic news-trading is scarce. However, the notion that algorithmic traders trade directly on macroeconomic news releases in a sub-second time frame is by now generally accepted.15 In this chapter, I will not concern myself with the scope of algorithmic trading on monetary policy news. Instead, I will focus on the process by which

algorithmic traders extract information from central bank communication and use it to trade with low-latency. To use central bank communication as an algorithmic input, it has to be quantified. To trade with low-latency, the quantification process has to be automated. Automatic quantification entails a loss of information; it is impossible to capture the entire content of an FOMC statement using this method. But, If algorithmic news-traders incorporate some part of monetary policy news into asset prices sooner than would otherwise be the case, their presence increases market

efficiency.

15

See Graham Bowley, ‘Computers that trade on the news’, The New York Times, December 22, 2010, or Scott Patterson, 'Speed traders get an edge’, The Wall Street Journal, February 6, 2014.

(15)

15

2.1

Low latency

Post-meeting FOMC statements are released to journalists at the press room of the Treasury Department in Washington around 10 minutes before they are released to the general public at precisely 2 p.m., measured by an atomic clock. Prior to March 2013, journalist had to wait on the sound of a bell at 2 p.m. before they could release any information regarding the meeting they just attended. Following an FOMC meeting on September 18, 2013, within 1 millisecond after 2 p.m., trading intensified in a wide variety of financial instruments. Traders reacted to a decision by the FOMC that it would refrain from tapering its $85 billion in monthly bond purchases. According to Nanex, a provider of high-frequency market data, the 10 milliseconds after 2 p.m. were the most active in the history of the U.S. stock and futures markets.16

Market watchers and regulators alike were baffled by the speed at which traders responded to the FOMC decision. The actual trading in response to the decision took place predominantly on exchange’s matching engines in New York and Chicago where traders have placed co-located servers to reduce latency.17 If the information regarding the decision would indeed have been released at 2 p.m. from the lock-up press room in Washington, it would have taken 2 and 7 milliseconds for the information to travel to the exchanges in New York and Chicago respectively. This is not in accordance with the surge in trading activity 1 millisecond after 2 p.m. Apparently, latency has become so important that it induced news organizations to store a simple ‘no taper’ signal that day under-embargo at the exchange’s matching engines in New York and Chicago to be released there at precisely 2 p.m. This way, algorithmic news-traders wouldn’t have to wait a couple of milliseconds for the information to reach them from Washington.18

This episode led the Fed to review its agreements with news organization regarding the embargo on FOMC statements. Irrespective of whether the trading within the first millisecond after the statement’s release was legal, it illustrates to what extent news-trading on quantifiable easy-to-interpret information is bound to be dominated by low-latency quantitative traders. It is unlikely that human traders are able to react first to this kind of news ever again, as the reaction-time of

16

See http://www.nanex.net/aqck2/4436.html.

17 Latency is such a critical determinant of algorithmic trading profitability that algorithmic traders have placed their trading computers in the same building as the exchange’s matching engine, against a fee of course. Although signals travel back and forth between a trader’s computer and the matching engine with the speed of light, traders rather place their computers next to the matching engine than at their office, some miles away. To ensure that latency is equal for every co-located trader, exchanges usually provide equal cable length to al trading computers no matter where the trader’s computer is located in the building.

18

See http://www.bloomberg.com/news/2013-09-24/regulators-look-into-timing-of-trades-after-fed-statement.html.

(16)

16 algorithmic traders is currently measured in tens of milliseconds.19 Scholtus, van Dijk and Frijns (2014) find that for algorithmic traders that trade on U.S. macroeconomic news releases, a delay of 300 milliseconds significantly reduces returns.

However, the ‘no taper’ information in this example is relatively suitable for algorithmic processing, comparable to macroeconomic news releases like GDP growth forecasts or

unemployment rates. It is relatively easy to quantify FOMC decisions. The complexity of FOMC decisions ranges from the target range for the federal funds rate to details regarding open market operations. These details entail information about what sort of financial instruments are added to the Fed’s balance sheet, like the specific size and maturity of Treasury securities.20 The quantifiable nature of FOMC decisions makes them suitable for algorithmic news-trading. How about more nuanced information contained in FOMC statements or minutes for FOMC meetings? Do algorithmic news-traders play a role in incorporating this kind of information into asset prices?

2.2

Algorithmic processing

Gürkaynak et al. (2005) study the response of asset prices to FOMC decisions and statements over the period of July 1991 to December 2004. In the half-hour following the post-meeting FOMC statement on January 28, 2004, two- and five-year yields for Treasury bonds jumped up 20 and 25 basis points respectively, the largest market response to any FOMC announcement in their sample. This market response was not caused by an unanticipated FOMC decision. The Committee decided to keep the short-term interest rate target at 1%, which was completely anticipated by financial

markets (Gürkaynak et al., 2005: 56). The large response in financial markets was caused by a rather subtle change in the way the FOMC formulated the likely stance of future monetary policy in the post-meeting statement. In the previous four statements, the FOMC used the exact phrasing ‘policy

accommodation can be maintained for a considerable period’21 to indicate that a raise in the target interest rate in the near future would be highly unlikely. However, this statement did not contain this phrasing. Instead, it contained the phrasing ‘the Committee believes it can be patient in removing its

policy accommodation’.22 The next morning, The Wall Street Journal ran a front page article on the episode, titled: ‘Fed clears way for future rise in interest rates; Federal-funds target left at 1%, but

19

See http://experts.forexmagnates.com/will-human-traders-be-first-to-react-to-news-events-ever-again/. 20 The recent financial crisis has induced the Fed to engage in the purchase of rather unusual type of instruments, like agency mortgage-backed securities. This development makes it harder for algorithms to process FOMC decisions.

21

See http://www.federalreserve.gov/newsevents/press/monetary/2003monetary.htm. 22 See http://www.federalreserve.gov/boarddocs/press/monetary/2004/20040128/default.htm.

(17)

17

central bank changes phrase on timing’. 23 In its analysis, the writer of the article, Greg Ip, concluded that the market response was so large because investors interpreted the omission of ‘considerable

period’ as a signal that the Fed was closer to raising interest rates than many thought. Futures trading

indicated that traders moved up the month when they thought the Fed would raise interest rates from August to June 2004. Besides Treasury bonds, the dollar/euro exchange rate also responded to the news. Since higher future interest rates would improve the return on dollar-denominated loans, the dollar rallied against the euro.

This anecdote illustrates the impact that one sentence in an FOMC statement can have. As figure 1 illustrates, FOMC statements have only become lengthier over time. The type of information that FOMC statements contain, based on subtle narrative indicators of future policy, seems

inherently inapt for algorithmic news-trading. However, quantitative traders do not need to feed their algorithms with perfectly noiseless information inputs, as long as they compensate the noise with low-latency.

For an algorithmic trader to profitably trade on market-moving information in FOMC statements, he has to proceed in two steps. First, he has to successfully capture a significant part of the information by quantifying the statement into possibly multifaceted indicators regarding the expected future stance of policy. In its simplest form, this entails determining whether the statement contains a dovish, neutral or hawkish outlook. Second, the algorithm needs to compare this outlook to market expectations. The greater the quantified outlook indicator deviates from market

expectations, the greater will be the market impact. Finally, the algorithm needs to be swift in completing these two tasks. Otherwise, competing market participants will have already capitalized on the news. This swiftness does not pertain to analyzing market expectations. The low frequency of FOMC announcement and the fact that most are scheduled, gives quantitative traders ample time to identify market expectations in advance. Collective market expectations can be gleaned from futures and option prices.

Quantifying FOMC statements

To derive empirical conclusions regarding the market response to FOMC statements, Rosa (2011) had to convert the information in the statements into a format suitable for statistical computation. To quantify the tone of the statements, Rosa constructs what he calls a wording indicator variable. This variable can take on three values; -1, 0 and +1. A negative value indicates a dovish outlook for future

(18)

18 monetary policy whereas a positive value indicates a hawkish outlook, 0 stands for a neutral outlook. For example, the FOMC statement released on February 2, 2000, contained the phrase; ‘The

Committee believes the risks are weighted mainly toward conditions that may generate heightened inflation pressures in the foreseeable future’. It was therefore quantified by Rosa into +1. In contrast,

the FOMC statement released on December 19, 2000, contained the phrase; ‘The Committee

consequently believes the risks are weighted mainly toward conditions that may generate economic weakness in the foreseeable future’. It was therefore quantified as -1 (Rosa, 2011: table 1). This form

of quantification results in an easy-to-interpret, though single faceted, variable.24 Its drawback for quantitative trading is that it requires time-consuming human intervention.

In order for algorithmic traders to trade on news stemming from FOMC statements with low-latency, they have to circumvent the need for human intervention in quantifying the statement’s information content. Lucca and Trebbi (2009) satisfy this need by providing an automated approach to measure central bank communication with application to FOMC statements. Automation

drastically decreases the time it takes to quantify a statement, which makes it applicable for algorithmic trading.

Lucca and Trebbi aim to quantify FOMC statements into what Rosa (2011) called a wording

indicator variable; a variable ranging from -1 to +1 depending on the hawkishness of the monetary

policy outlook contained in the statement, where -1 indicates a dovish outlook and +1 indicates a hawkish outlook. A value is assigned to each sentence in the statement and the average value determines the variable. To automate this process, they rely on techniques from computational linguistics and statistical natural language processing. Their wording indicator variable is a function of two scores, based on two classes of automated scoring algorithms. These scoring algorithms use different ways in which they automatically, without human intervention, determine the semantic orientation, or meaning, of phrases in the statement on a dovish-hawkish metric.

The first score Lucca and Trebbi (2009) employ is called the Google semantic orientation

score. In its most basic form, the algorithm determines whether a sentence or string of words from

an FOMC statement features a larger co-occurrence with the word ‘hawkish’ or ‘dovish’ in a corpus of text, in this case the universe of webpages available through the Google search engine. For example, if a sentence in an FOMC statement like ‘Pressures on inflation have picked up.’, results in more search hits on Google when it is co-searched with ‘hawkish’ relative to when it is co-searched

24

In another academic endeavor to quantify the information contained in FOMC statements for statistical computation, Bernanke, Reinhart and Sack (2004) manually construct a dummy variable called Statement

(19)

19 with ‘dovish’, it is assigned a positive (hawkish) semantic score.25 The algorithm repeats this

procedure for the entire set of sentences and produces an average score. The algorithm does not actually visits websites, it relies on hit counts. It is the large scope of the search that makes up for this shortcoming.

The second score is called the Factiva semantic orientation score. This score is constructed in a similar fashion to the Google semantic orientation score. However, it uses a more relevant corpus of text for reference. The algorithm cross-searches sentences from the FOMC statement against discussions of FOMC announcements from newspapers, magazines, newswires and newsletters which are included in the Dow Jones Factiva database, a provider of business and financial news.

Lucca and Trebbi find that the wording indicator variable based on both the Google- and the Factiva semantic orientation score significantly predicts future policy rate actions with a lead of more than a year. The applicability of this automated approach for quantitative trading is apparent. FOMC statements can be quantified in a matter of seconds using this approach. However, automatic quantification – as does quantification in general - results in a relatively noisy description of FOMC statements. Algorithmic news-traders can decrease this noise by preparing their algorithms for specific announcements. As I have mentioned earlier, time is not a constraint preparing for

announcements. For example, an algorithmic trader that wants to trade on information in the next FOMC statement, might program his algorithm to search for specific sentences. In 2003, the

Committee used the sentence ‘policy accommodation can be maintained for a considerable period’ in four consecutive statements. The omission of this sentence in the subsequent statement caused a large response in financial markets. Consistency by the Committee in formulating the statements, greatly enhances the viability of this type of quantification. As another example, FOMC statements contain the voting record for the decisions that the Committee made during the preceding meeting. Dissident voters are not uncommon and savvy quantitative traders can program their algorithms to assess unexpected voting outcomes. By preparing algorithms for specific statements, quantitative news-traders can increase the predictive power of quantified statements beyond semantic scores. Profitable trading is realized when the noise created by quantification is sufficiently compensated for by low-latency. Simply put, a trader does not need to grasp 100% of the statement’s meaning if he is able to trade on the statement’s unanticipated information content within (milli)seconds.

25

I have verified that following this procedure would indeed result in a positive, hawkish, semantic score. Co-searching ‘hawkish’ and ‘pressures on inflation have picked up’ on Google yields about 662 results, whereas co-searching ‘dovish’ and ‘pressures on inflation have picked up’ yields only about 135 results, at time of writing.

(20)

20 Quantifying minutes for FOMC meetings

Minutes for FOMC meetings provide insights into the thought process behind policy decisions and into how the outlook for the future stance of monetary policy, conveyed in the statements, was determined by the Committee. The minutes include extensive individual assessments of Committee members. In chapter one, I have provided evidence that these minutes contain market-moving information. This result is in accordance with Boukus and Rosenberg (2006) who find that market participants can extract a complex, multifaceted information signal from the minutes. Because the Committee’s collective opinion is already expressed in the post-meeting statement, the precise composition of opinions held by the Committee’s members is perhaps valuable, but hard to evaluate. The nature of this information seems even more inapt for algorithmic news-trading than the

information in FOMC statements.

However, Apel and Grimaldi (2012) use an automated approach to quantify the tone in minutes of the Swedish Riksbank and find that the resulting measure helps to predict future policy after controlling for macroeconomic variables and market expectations. Their measure is created without the use of a reference text corpus like in Lucca and Trebbi (2009). Instead, an algorithm searches the minutes for a predetermined set of noun-adjective combinations, like ‘higher inflation’ and ‘lower growth’, to arrive at a hawkish-dovish metric similar to the one used by Rosa (2011) and Lucca and Trebbi (2009).

Boukus and Rosenberg (2006) find that ‘the complex, multifaceted signal’ that the minutes contain is rather unsuited to be quantified only into a hawkish-dovish metric. Therefore, they use another approach, called Latent Semantic Analysis (LSA). LSA is an automated statistical methodology that determines to what extend a text expresses different themes based on word frequencies26. Different themes are associated with specific macroeconomic and financial market indicators. The application of LSA for algorithmic news-trading is as follows. The use of a certain set of words in FOMC minutes is associated with a specific theme and this theme in turn is associated with for example a change in Treasury bond yields. Upon release of the minutes, an algorithm can quickly perform LSA to determine whether this theme is overrepresented in the text, relative to previous minutes or to an expected benchmark. If this is the case, a correction in Treasury bond yields can be expected and a profitable trading opportunity exists.

26

These themes are not really demarcated intuitive economic concepts. Rather, they are mathematical tools to differentiate between different subsets of words.

(21)

21 2.3

Media-driven algorithmic trading

An important feature of monetary policy news other than official central bank communication is that it often comes unannounced. Therefore, algorithmic traders cannot adequately prepare their

algorithms for specific information. They have to rely on a more generic approach. Algorithmic news-traders in the U.S. stock market are known to follow such an approach. A new industry is emerging that vendors machine-readable news to quantitative traders.27 These data feeds might have to be complemented with news-reading computer packages such as RavenPack before they can be utilized by algorithmic news-traders. RavenPack analyzes all articles on the Dow Jones Newswire and

converts them into a format of metrics usable for algorithmic trading. These metrics include which company is the articles’ subject, whether the article is relevant and whether the article presents a negative or positive image about the company. RavenPack is able to deliver these metrics within a second. Beschwitz, Keim and Massa (2013) studied the impact that RavenPack had on high-frequency stock market dynamics in the aftermath of news events, since its release in April 2009. They

determine that RavenPack is very skilled in identifying the relevance and sentiment of an article in that the program is able to predict subsequent stock returns. In this fashion, algorithmic traders might quickly respond to news reports that cover statements by influential monetary policy makers.

2.4

Graphic presentation

The analysis in the preceding part of this chapter demonstrates how algorithmic traders can process central bank communication and trade on it very quickly upon release. By embedding new

information into asset prices sooner than would otherwise be the case, the presence of algorithmic news-traders increases price efficiency. They overcome the noise created by automatic

quantification by being faster than other traders. I assume that once this speed advantage has disappeared, the role of the algorithmic news-trader is done. Several minutes after an FOMC statement is released, human traders have had time to read the statement and formulate the appropriate trade response. By this time, the quantified version of the statement is no match for human comprehension. In a theoretical framework, I therefore assume that algorithmic news-traders are only active immediately following the release of central bank communication, say in the first second. After that, they have already been beaten to the punch by faster algorithmic traders with the same noisy information or they are outgunned by more informed human news-traders.

Unfortunately, I have not found concrete measures regarding the extent to which algorithms can

27

Providers include Dow Jones Elementized News Feed, Thomson Reuters News Feed Direct, AlphaFlash, Bloomberg Event Driven Feeds and NASDAQ OMX Event-Driven Analytics.

(22)

22 extract information from central bank communication. I can only conclude from the analysis in this chapter that algorithms capture a larger part of information regarding FOMC decisions than

regarding FOMC statements and that they capture a larger part of information in statements than in minutes.28

Figure 2 graphically depicts how algorithmic news-traders can increase price efficiency in the aftermath of monetary policy news. This is explicitly not a depiction of empirical results. Rather, it is a graphical tool to clarify points I made earlier. The Y-axes depicts how much of the new information (‘I’) is incorporated in the asset price. The X-axes depicts the time (‘t’) since the news was released.29 Figure 2.a illustrates the situation in which human news-traders embed new information extracted from an FOMC statement into the asset price over time. I assume that 50% of the new information is embedded within 15 minutes. Price efficiency is measured by the surface beneath the curve.

Instantaneous embedding of all new information would result in complete price efficiency. In figure 2.b, I introduce a representative algorithmic news-trader. In the new situation, I assume that 30% of the news is embedded in the asset price immediately upon release. After one second, the role of the algorithmic news-trader is done and the remainder of the curve is unaffected. Figures 2.c and 2.d. depict price efficiency upon release of minutes for FOMC meetings. Because the information contained in these minutes is more nuanced, the representative algorithmic news-trader can only extract 20% of the minutes’ information. Also, it takes longer upon release before the fastest human news-trader can respond to the new information. For example, the minutes for the September 2014 FOMC meeting covered 26 pages. It is therefore a reasonable assumption that human news-traders need at least 5 minutes before they can trade on any new information contained in the minutes. Notably, the response time of the algorithmic news-trader is equal for statements and minutes; virtually non-existent. According to figure 2, the presence of algorithmic news-traders clearly increases price efficiency (the dashed line represents the old situation, whereas the solid line represents the new situation). This increase is bound by the extent to which algorithms can capture the information content of a certain communication format. I expect that technological advances in computational linguistics and statistical natural language processing push this bound upward in the future. For now, the inherent noise created by automatic quantification leaves room for news-trading by human traders, who take longer to evaluate news, but digest it more thoroughly.

28

I expect that oral statements like press conferences are even harder for algorithms to process. Because I assume that algorithmic news-trading on oral central bank communication is currently virtually non-existent, I have not addressed it.

29

The type of graphical presentation used here is copied from Beschwitz et al. (2013), figure 2. This figure depicts how the ‘share of stock price reaction’ increases from 0% to 100%, in the first 120 seconds after an online news message.

(23)

23 Figure 2. Price discovery in the aftermath of monetary policy news

(24)

24 2.5

Choices in central bank communication

Amidst an emergence of algorithmic news-trading in financial markets, the type of format in which central banks communicate with the public, has implications for the high-frequency response of asset prices. Increased verbal communication in the form of more press conferences might remove noise in the expectations of market participants in general. However, because verbal communication is a type of information that is inherently difficult to digest for algorithmic traders, it will take relatively long for asset prices to respond to news in press conferences. Other decisions about communication might be strictly beneficial. To make their news articles more machine-readable, some news

providers started attaching links to for example company names. This makes it easier for algorithms to discern the company name ‘Apple®’ from the fruit ‘apple’. The FOMC might use similar tactics in the future to court algorithmic traders and decrease the noise in their communication for algorithms, all the way up to releasing two related FOMC statements of which one is completely

machine-readable and makes no sense to humans. The effect that algorithmic traders have on ultra-short-term asset price reactions to monetary policy news, requires further investigation.

(25)

25 3.

High-frequency trading and information input

In chapter two, I have illustrated how algorithmic traders can trade directly on monetary policy news. I have also indicated that algorithmic news-trading takes place only in the immediate aftermath of the news release. Once the speed advantage of the algorithmic news-trader is gone, so is its role in price discovery. In roughly an hour following the release of central bank announcements, human traders are imbedding their heterogeneous beliefs about the new information into asset prices. During this period – with exception of the first few seconds - human comprehension is superior to algorithmic news processing. That doesn’t mean that algorithmic traders are inactive during this period. Continuously throughout the day, algorithmic traders engage in high-frequency trading (HFT).30 In this capacity, algorithmic traders play a secondary role in price discovery around monetary policy news. By trading on the information of other market participants, high-frequency traders might embed new information into asset prices sooner than would otherwise be the case. Therefore, the presence of algorithmic traders that engage in high-frequency trading in the aftermath of

monetary policy news, might increase price efficiency.

In this chapter, I will discuss what kinds of information high-frequency traders process. How do they interact with each other and with human traders? The first paragraph provides a definition of high-frequency traders. Paragraph two illustrates the scope of HFT over time, geography and asset classes. Paragraph three provides evidence from the U.S. equity market regarding HFT strategies and information input. Finally, paragraph four provides a graphic presentation of how HFT affects price efficiency around monetary policy news. In the next chapter, I extend these findings to the foreign exchange market.

3.1

Defining HFT

High-frequency trading is a subset of algorithmic, automated or quantitative trading. According to the U.S. Securities and Exchange Commission (2010: 45), high-frequency traders can be defined as:

‘professional traders acting in a proprietary capacity that engage in strategies that generate a large number of trades on a daily basis.’ […] ‘Other characteristics often attributed to proprietary firms engaged in HFT are: (1) the use of extraordinarily high-speed and sophisticated computer programs for generating, routing, and executing orders; (2) use of co-location services and individual data feeds offered by exchanges and others to minimize network and other types of latencies; (3) very

30

To be clear, a trading firm can simultaneously engage in algorithmic news-trading and in high-frequency-trading. For simplicity’s sake, I treat algorithmic news-traders and high-frequency traders as separate types of algorithmic traders.

(26)

26

short time-frames for establishing and liquidating positions; (4) the submission of numerous orders that are cancelled shortly after submission; and (5) ending the trading day in as close to a flat position as possible (that is, not carrying significant, unhedged positions over-night).’ 31

HFTs generate a large number of trades by trading in relatively small sizes and by employing very short risk holding periods, usually measured in seconds. The generic HFT business model is based on capturing a large number of small per-trade profits. In order to establish and liquidate positions within seconds with minimal market impact, HFTs gravitate towards the most liquid instruments. In the subsequent part of this chapter I will gradually elaborate on the five HFT characteristics stated by the SEC as they become relevant to specific HFT strategies

3.2

The scope of HFT

High-frequency trading has emerged at a staggering pace in the previous decade. In 2010, HFT accounted for more than half of trade volume in U.S. equities. Equity trading in Europe, Japan, Australia and Canada has also been increasingly driven by HFT. Figure 3 depicts the share of HFT activity in equities by geography for 2010.

Figure 3. Share of HFT activity in equities by geography (2010)32

Source: Celent.

31

An extensive review of HFT definitions can be found in Gomber, Arndt, Lutat and Uhle (2011). 32 From Grant and Demos, FT, March 5, 2012

(27)

27 In recent years, HFT market share and revenue have fallen in both the U.S. and the European equity market. Figure 4 illustrates that HFT market share in U.S. equities has dropped to about half of total trade volume and HFT market share in European equities has dropped to about a quarter of total trade volume. Massoudi and Stafford (Financial Times, 2014) link this development to worsened market conditions. The combination of high volatility and high overall volume in equities made HFT very profitable in the previous decade. Higher volatility and volume create more arbitrage

opportunities; small price discrepancies regarding a single instrument that is traded on multiple platforms or regarding an Exchange-Traded Fund (ETF) and its underlying stocks. Because of their superior speed, HFTs are particularly good at capitalizing on these arbitrage opportunities. In the aftermath of the 2007-08 financial crisis, overall volume and volatility dropped, along with HFT market share and profit. Aggregate HFT profit further declined because new HFT firms, drawn by the high profits earned by pioneering HFTs, increased market competition. It is estimated that there are currently around 400 HFT firms active globally, of which about 10 firms account for the bulk of volume.33

Figure 4. Share of HFT activity in U.S. and European equities (2005-2014)34

Worsened market conditions and increased competition in equities forced HFT firms to explore new geographies and asset classes. Figure 5 illustrates that despite a declining market share in U.S. equity trading, HFT was still growing globally in futures, foreign exchange and fixed income trading in recent years. According to this estimate, HFT accounted for about 40% of volume in global foreign exchange

33

According to Celent, A research and consulting company, as voiced in an article by Anthony Malakian published on www.waterstechnology.com, May 5, 2014

(28)

28 trading in 2012. While HFT firms expand their activities across markets, they usually adapt strategies and technologies they are familiar with from equity trading.

Figure 5. Share of HFT activity by asset class (2008-2012)35

Based on figure 5, there is a clear convergence between HFT market share in U.S. equities and global foreign exchange trading. However, the vast majority of academic literature regarding

high-frequency trading relates to equity trading. The notion that HFT is becoming an increasingly relevant factor in the FX market should drive more academic research on HFT in FX. This thesis aims to contribute towards that objective. Apart from whether HFT is increasing or decreasing in different asset classes and geographies, it is clear from figures 3, 4 and 5 that HFT has become a significant presence in financial markets around the world.

(29)

29

3.3

Evidence from the equity market

To assess how HFTs interact with other market participants and what their role is in the price discovery process, it is critical to acknowledge that there is heterogeneity in HFT strategies. I follow Hagströmer and Nordén (2013) who distinguish between market making HFTs and opportunistic HFTs. They analyze HFT in 30 Swedish large-cap stocks traded on the NASDAQ-OMX Stockholm exchange for August 2011 and February 2012, and find that the lion’s share of HFT activity, 65%-71%, results from market making. The residue of HFT activity results from opportunistic HFT. Opportunistic HFT involves strategies such as arbitrage and directional trading based on order flow anticipation.

HFT market making

In 2007, trading platform Chi-X was launched. It competed with market incumbent NYSE-Euronext for trading in, among other equities, Dutch index stocks. Menkveld (2013) studies the strategy of a large high-frequency trader that started trading on both the entrant and the incumbent market.36 He finds that this HFTr uses mainly passive limit orders to continuously post on both sides of the limit order book. This strategy earns the HFTr the spread, while providing liquidity to the market. Therefore, Menkveld characterizes this HFTr as a modern market maker.

There is an inherent downside to market making in the form of adverse selection. The market maker’s quote is hit by a counterparty that might be better informed because it was able to process information, regarding the traded instrument, that entered the market after the market maker posted its limit order. For example, the market maker accumulates a net long position in the stock when its passive limit buy order is hit by an aggressive market order to sell. If this order is motivated by new information, the price of the stock subsequently declines. The market maker has to wait for another counterparty to trade against its limit sell order before he can offload his position. In the end, the market maker has earned the half-spread twice but incurred a positioning loss due to the adverse selection. In decomposing the HFTr’s profit, Menkveld (2013) finds that the HFT market maker makes money on the spread but indeed loses money on its positions.

Adverse selection explains why HFTs gravitate towards the most liquid instruments. Liquidity enables HFT market makers to revert to neutral positions more easily. It also explains the

comparative advantage of HFT market makers vs. traditional human market makers. If a market maker is better able to mean-revert positions because he can trade in shorter time frames, he

36

Menkveld identifies this trader as a high-frequency trader based on similar HFT characteristics discussed in paragraph 2.1. The single high-frequency trader is otherwise anonymous.

(30)

30 reduces potential positioning losses caused by adverse selection. This intuition is confirmed by Menkveld (2013) as he finds that HFT positions that are hold for less than five seconds are generally profitable whereas positions that are held for longer than five seconds generally cost money. To increase the probability of returning to a flat position, HFT market makers skew quotes relative to fundamental value (Menkveld, 2013). This means that when a risk-averse market maker is long relative to his optimal position - presumably a neutral position - he adjusts his quotes downwards to trade out of his position. This mean-reverting tactic might be augmented by the use of ‘costly’ marketable orders if a non-flat position grows too large.37

A related reason why high-frequency traders replaced traditional market makers is because they can update, or refresh, quotes more frequently. By cancelling and resubmitting their limit orders more frequently, HFTs reduce the time in which new information can cause quotes to become stale. This reduces adverse selection costs and allows HFT market makers to compress spreads, pushing traditional market makers out of the market in the process. In the late 1980’s, when the first algorithmic-like market makers entered the U.S. equity market, human market makers had to read new prices on a computer screen, process that information, make a decision regarding its own orders and then type in those orders on the keyboard of a NASDAQ terminal (Steiner, 2012). This would take seconds at best. Nowadays, computer algorithms can perform this task in milliseconds.

Market making and rebate income

Back in 1998, Joshua Levine, the creator of the Island ECN, acknowledged that market makers face adverse selection costs. Island was an upcoming electronic equity trading platform that suffered from a lack of liquidity. Levine understood that to boost liquidity, Island had to attract implicit market makers. To that avail, Island started to provide a revolutionary fee structure. Previously, Island - similar to other market centers like NYSE and NASDAQ - charged traders $1.00 per order executed. Starting June 1, 1988, Island paid customers who made a trade – because their limit order was hit – $0.01 for every hundred shares. Traders that took the trade – by using a market order or an

aggressive limit order – paid $0.025 for every hundred shares. In effect, the Island fee structure provided additional compensation, beyond the spread, for market makers to bare the adverse selection cost. The new fee structure, now known as the maker-taker model, indeed boosted

liquidity on Island. Market makers, incentivized by the rebates, flocked to the entrant ECN and this in turn attracted liquidity demanders (Patterson, 2012).

37

Marketable orders are costly in this case because they entail crossing the spread, instead of earning the spread.

(31)

31 Today, the maker-taker model has become the standard pricing model for equity exchanges. Because the rebate is typically less than the fee, exchanges extract a guaranteed income from the pricing model. On platforms where rebates are relatively high, HFT market makers will post bids and offers closer to the fundamental value of the stock. By lowering the spread, they are more

susceptible to adverse selection costs, but this is compensated for by higher rebate income. Furthermore, HFT market makers lower spreads on platforms that provide volume incentivized rebates. Angel, Harris and Spatt (2013) believe that maker-taker pricing has not changed net spreads, but has only decreased quoted spreads. HFT market makers compress spreads on a platform that provides a higher rebate, but this compression doesn’t benefit liquidity takers as fees are likely also higher. In conclusion, the HFT market maker’s profitability is the aggregate of a profit earned on the spread and rebates, and a positioning loss caused by adverse selection.

HFT market making and hard information

HFTs’ ability to update quotes relatively frequently is not very valuable if HFT market makers gain insufficient information to direct updates. Chapter two already indicated that algorithmic news-traders can process a limited portion of macroeconomic news. Similarly, to update quotes, an HFT market maker can only process the limited portion of information, relevant for ultra-short-term asset price fluctuations, that is apt for real-time algorithmic computation. Jovanovic and Menkveld (2011) term this portion hard information, as opposed to soft information. They construct a model where hard information enables an HFT market maker to update quotes relatively frequently, thereby reducing adverse selection costs. Consequently, they test the model's implications using a sample of trade and quote data on all Dutch nonfinancial index stocks traded on both Chi-X and Euronext, for January 10 to April 23, 2008.

Jovanovic and Menkveld (2011) do not provide an in-depth analysis of the nature of hard information. They construct a proxy for the relative proportion of hard information, relative to all information that can move the price of a stock in the immediate future. This proxy is based on the idea that a stock's return is comprised of a market wide return and a stock-specific return. They further assume that changes in the market wide return can be gleaned from the highly active index-futures market. HFT market makers predict the price change of a stock in the immediate future based on futures-prices – hard information – and adjust their quotes accordingly. The actual price may turn out other than expected due to unanticipated stock specific news, like a management change – soft information. Hard information enables HFT market makers to adjust quotes on individual stocks relatively frequently, as index-futures prices are continuously available.

Referenties

GERELATEERDE DOCUMENTEN

Here, the returns of Stellar and the lagged HE sentiment scores (for a number of different variants) produce relatively high correlations (as high as 0.09). The set of variables

Door er geen aandacht aan te besteden vallen zij echter toch onder de nieuwkomers binnen het fantasyveld die fantasyboeken goed vinden op basis van inherente

En las montañas de Holanda, es un producto híbrido, una obra que él mismo ha creado, entretejido como una trenza; el autor crea una región geográfica ‘sur’ de Los Países Bajos

On the other hand, the overlap between the concepts can be used to answer the main research question of how network governance can accelerate the transition process in Smart

De positieve toon waarop dagblad De West verslag doet over de V7 en de V7 partijen samen, is ook niet langer significant wanneer de opiniestukken eruit worden gefilterd en

Furthermore, by probing the structure of the dark matter density spike around the black hole, eLISA can be used to determine the initial dark matter halo profile of the

van die reeks besluite aan om die Sap-pers, die Sap-liberaliste en die Torch Commando daaraan te hcrinner dat bulle geen ongunstige kommentaar op die konfcrensie se

Maar het antwoord dat het meeste voor komt, is dat spiritualiteit iets is waar ze altijd mee bezig zijn en niet iets is ‘wat je er extra bij doet’.. Evelien zegt bijvoorbeeld dat