An Empirical detection of HFT strategies

[preliminary ]

October 1, 2015
Abstract

This paper detects empirically the presence of High Frequency Trading strate-
gies from public data and examines their impact on financial markets. The objec-
tive is to provide a structured and strategic approach to isolate signal from noise
in a high frequency setting. In order to prove the suitability of the proposed ap-
proach, several HFT strategies are evaluated on the basis of their market impact,
performance and main characteristics.
1 Introduction
High-frequency trading (HFT) refers to a subset of algorithmic trading that involves the
usage of specialized digital infrastructure and ultra-low latency computing to execute
a wide range of strategies on financial markets (OHara 2015). While HFT has been
present since the digitization of markets, it has only become wide spread over the
past decade. Evidence suggests that a very large percentage of the trading activity
in many markets is now carried out via HFT algorithms across a variety of assets.
Projections indicate that this presence is likely here to stay, and perhaps even to increase
over the coming years (Kumar, Goldstein, Graves, 2011). HFT has been the recent
focus of researchers, particularly those focusing on market microstructure, as well as
regulators and even the popular media, mainly owing to increasingly frequent incidents
of abnormal market behavior supposedly linked to HFT algorithms. Examples of such
events include the Flash Crash of 2010, which has since attracted investigations by
multiple researchers, media reports and has even inspired entire books (Lewis 2014).
Similar, but less high profile, accidents have also occurred in other markets, namely the
treasury markets in October 2015 and on the fixed income instruments. The impact has
not been homogenous, and in some instances, prices have actually increased abnormally.
This was the case with a recent episode with US Treasuries, where the market saw
a flash crash of yields indicating a rapid increase in price as discussed in the U.S.
Department of Treasury, 2015. Whether these accidents are entirely to blame on HFT
activity is not yet certain, but it has captured the interests of regulators, investors and
market participants. Hence, a deeper understanding of their working and impact on
the wider markets and asset prices is crucial. The HFT paradigm shifts for markets
makes necessary the usage of new tools to detect and analyze the activities of high speed
traders. Such tools may be of interest to regulators, investors and academics alike. The
methods presented in the current paper provide a detailed strategic approach which
aims at isolating signal from noise. Appropriate benchmarks are selected to evaluate
the performance and the impact of two different HFT strategies.
Weighing up the costs and benefits to financial markets arising from HFT activities
is an exceptionally complex task. On the one hand, lightning fast trading may mean
that markets are becoming increasingly informationally efficient, as news and data are
rapidly incorporated into prices. Efficient capital markets are of crucial importance
to the optimal allocation of scarce resources in an economy. By utilizing the pricing
mechanism, efficient and a well functioning capital markets bridge the gap between
economic agents with a deficit of funds and those with a surplus and ensure that wealth
is employed where it produces maximal returns. On the other hand, critics of HFT

2
often argue that they contribute to the noise in the market, and that their activities
are disconnected from economic or financial fundamentals, creating volatility which
discourages long term investors and issuers of securities. Others yet argue that HFTs
are simply deliberately set up to game the system and exploit opportunities arising
from market microstructure as organized by exchanges in order to profit at the expense
of other participants. Many of these arguments appear to have some merit (Turbeville
2013).
The debate between supporters and opponents of HFT is heated and often confusing.
The powerful interests of lobbyists and vested stakeholders on both sides (HFT firms on
one end, buy-side investors and traditional traders on the other) are in constant battle
to persuade regulators, exchanges and the public to their views (Rijper, Sprenkeler,
and Kip 2010). Joining this debate is not the objective of the present paper. Instead,
it deliberately focuses on providing a positive and empirically driven point of view.

2 HFT strategies and their impact

One of the main challenges for researchers on the topic of HFT is the wide diversity of
strategies employed by practitioners these may range from traditional market making
through statistical arbitrage to more speculative and aggressive strategies.
Traditional market making is well researched and involves a market participant taking
both sides in the market for a security simultaneously, by providing pairs of Bids and
Offers. A market maker aims to make a profit based on the spread between the Bid
and Ask prices, and in the process needs to manage inventory and execution so as
to minimize undesired exposure to price risk. Market makers have been present on
markets for a long time, and in fact in some cases have been institutionalized such as
historically the case of the NYSE specialists, the Designated Market Makers (DMMs)
of Euronext and the Kursmakler of Frankfurt Stock Exchange. In such arrangements,
the market makers are allowed to profit from an exclusive proprietary knowledge of
the security and the order flow relating to it. In exchange, they are often expected to
maintain an orderly market in the security a duty to step in and stabilize the market in
cases of extreme price fluctuations, perhaps even at the costs of incurring a loss in the
process. Traditional market microstructure models have been utilized to understand the
behavior of market makers and the factors influencing their decisions see for instance,
(Kyle 1985).
HFT practitioners which chose to act as market makers in securities may follow the ex-
act same strategies as their traditional counterparts with several significant differences.
The main distinction of HFT market makers is their mandate for opportunistic behav-

3
ior and their lack of long term obligations to market participants. While traditional
market makers are well known and have certain responsibilities for the markets they
make, HFT firms are often small and relatively unknown entities. They may withdraw
from the market at any point if they feel that conditions are unfavorable. In fact, such
decisions may be taken automatically by the algorithms which can detect changes in
market conditions much more rapidly than their human counterparts. Indeed, this is
what many argue may have occurred during the Flash Crash where a deterioration in
market conditions may have triggered algorithms to withdraw from the market, causing
further deterioration in liquidity, in a vicious cycle. On the other hand, prices quoted
by HFT market makers may be quicker to adjust to news regarding the fundamentals
of the underlying security, or the information conveyed by rapid changes in order flow.
Similarly, from a theoretical point of view HFT may be better at managing invento-
ries in real time, thus reducing the inventory risk component priced in quoted spreads.
Indeed, it is common for HFT firms to carry flat positions overnight regardless of the
strategy they follow. Another benefit for market efficiency should be the instanta-
neous incorporation of cross asset arbitrage into pricing in real time which may only be
achievable through the usage of low latency algorithms.
Statistical arbitrage, pattern recognition and directional trading are also well known
strategies which have been employed widely before the rise to prominence of HFT. These
are mainly aimed at exploiting long lasting patterns and relationships in asset prices
in order to profit. Such strategies are often employed by aggressive buy and sell-side
institutions such as hedge funds or the quantitative strategy desks of investment banks.
While many of these techniques may be carried out through traditional algorithmic
trading, it is clear that HFT may provide some incremental benefits, due to the speed of
execution and detection of the patterns, which is always valuable in highly competitive
financial markets.
The last group of strategies employed by HFTs are exclusive to them and thus form a
very distinct group which is separate from the previously examined ones. This group
of strategies have been called predatory, malicious and manipulative by researchers
and market participants and may be seen as exploiting market microstructure at the
expense of other traders, see (Foucault, Hombert, and Roşu 2015) and (Biais, Foucault,
and Moinas 2015) and (Biais and Foucault 2014).
Some of the most commonly referenced predatory strategies include Quote Stuffing,
Momentum Ignition, Order Fade and Pinging. Quote Stuffing is a strategy characterized
by HF traders rapidly submitting a torrent of orders (bid, ask or both simultaneously)
in the order book over an extremely short period of time. Reports by Credit Suisse
estimate that the majority of Quote Stuffing episodes last up to 2 seconds. However the

4
reported distribution is uneven, with some episodes lasting significantly longer. These
events are frequently observed in various markets and assets. An obvious objective is
to flood the market systems with (useless) information in the form of rapidly changing
quotes, so as to slow down and confuse the responses of competing algorithms and other
traders. This is very similar to a DDoS attack on a computer system. However, there are
other possible objectives. One measure frequently used by algorithmic traders as well as
traditional investors in benchmarking trade execution and market prices is the midpoint
this is derived as the average of the Bid and Ask prices and is then used to determine the
relative spread or the ex-post quality of trade execution. By practicing Quote Stuffing,
HFT may induce significant quote volatility which may cause the midpoint to also
fluctuate. These dynamics may induce certain algorithmic or human traders to trade
based on the rapidly moving midpoint values. However, in the instance that the trader
carelessly utilizes market orders instead of limit orders, they may receive execution at
a different price than anticipated. During Quote Stuffing episodes for instance, the bid
and/or ask quotes rapidly fluctuate over the space of microseconds. In the case where
the underlying asset is also traded in Dark Pools which use the midpoint in lit markets
as part of their derivative pricing mechanism, HFT may attempt to manipulate the
midpoint via Quote Stuffing in order to gain favorable execution in the Dark Pool.
Quote stuffing is easily observed via several characteristic patterns of quote volatility
which may occur on the Ask, the Bid or both simultaneously. Empirically, Quote
Stuffing is observed as having some influence on the direction of price moves immediately
following an episode with prices seen as more likely to move in the direction of the
stuffing. Therefore, when the stuffing occurs on the Bid side, prices are observed as
they are more likely to fall; whereas the opposite holds if stuffing occurs on the Ask
side. Quote Stuffing is a very frequently observed phenomenon, up to 53 occurrences
per day in some market segments as reported by (Tse, Lin, and Vincent 2012).
Momentum Ignition is a strategy which is used by HF traders to cause and exploit
sharp and rapid movements in prices occurring over a limited amount of time. (Tse,
Lin, and Vincent 2012) observe that the average Momentum Ignition events occur over
the space of 1.5 minutes. The aim of this strategy is to instigate other traders to join the
speculative movement and cause a rapid snowball like price cascade. This is achieved
through trading in high volume in specific patterns aimed at triggering other market
participants to follow. If successful, the strategy results in a sharp price move which is
usually not of very large size; although it is significant when the short time frame of
the episode is taken into account. Once prices have moved, the HFT will cover their
position and the move will reverse itself albeit on much lower volume than the original.
A fully executed episode of Momentum Ignition leaves a distinct impact on price and

5
volume data. This is characterized by a spike in volume accompanied by flat prices.
This is then followed by a sharp price jointly with an even greater increase in volume.
Once the high or low point of the price move is achieved, volume declines and price
gradually reverts back to its original level. These events are relatively less frequently
observed with an average of at least 1.6 event per stock per day as reported by (Tse,
Lin, and Vincent 2012). This pattern is of particular interest for the current paper
as it carries a direct and readily measurable economic result for practitioners, while
assuming seemingly very low risk as the move is caused and exploited at the same time
by HFT.
Order Fade and Layering refers to two very similar strategies used by HF traders to
front run large order volume or create a false illusion of liquidity within the order book.
Practitioners will cancel partially (for some of the volume) or completely their orders on
different levels of the order book, following the execution of a trade. The aim is to make
the traders demanding liquidity and thus pay a higher price by having to run up or down
the book. This is also, however, an example of fleeting liquidity caused by HFT. The
orders submitted in the order book disappear before they can be executed, therefore
creating an illusion of liquidity which is not really there (since it cant be transacted
on). Similarly, when Layering, the HF traders will submit a large number of orders
on the bid or ask side of the order book in order to create a false impression of strong
buying or selling pressure and manipulate other traders. The instigators however have
no real intention of executing these orders which will be cancelled by the low latency
algorithms as soon as execution is attempted. Both of these practices are extremely
frequently observed and have become widespread in many markets. Observing and
analyzing these strategies fully requires the usage of more than one level of the order
book and is beyond the scope of the current research. Further factors which complicate
the analysis are data synchronicity issues and cross venue fade, where events on one
trading venue cause the algorithms to adjust their position in the order book on a
different venue simultaneously.
Pinging refers to a practice where the high speed traders place a series of small volume
orders with the objective to survey the market and hunt for large institutional traders
who are seeking to trade for liquidity reasons. This is done by sending orders within
the frames of the prevailing bid-ask spread at the time in order to entice the large
uninformed traders to react and execute. Once this is detected, the algorithms deplete
all liquidity on the side of the order book the large trader is interested in transacting
with. If the large trader is looking to sell for instance, the algorithm will look to dry
up all the bid liquidity in the market over a very short period of time and replace it
with its own orders at a much higher spread than before. This is a tactic used to force

6
large traders to pay a higher spread, and once executed, the algorithms will unload
their position at the much more favorable market prevailing spreads. This tactic of
front running is one of the most frequently cited by critics of HFT and is referred to as
whale hunting, as in (Turbeville 2013).
Clearly, in a complex and fragmented market place there can be no real end to the
strategies employed by market participants, and even more so for highly sophisticated
ones such as algorithmic and HF traders. For the sake of maintaining the focus of the
present research paper, only two of these strategies, Quote Stuffing and Momentum
Ignition are analyzed in detail. Hopefully, the techniques and analytical tools provided
in this paper may be adapted and used by future researchers to analyze a further array
of strategies encountered in the market place.

3 Selection of assets and data sample

3.1 Selection of assets
The main contribution of this paper lies on its strong empirical focus to offer an array of
ready tools aimed at detecting and examining the presence of HFT strategies from the
public market data. The main immediate prerequisite for examining HFT strategies
is the access to high quality data. As there is ample evidence of the presence of high
frequency traders in the market across various markets and assets, we select a limited
but diverge range of assets for testing the new measures. Upon their successful use, the
data sample may be readily expanded for future research to include additional assets.
The assets selected are: Apple Inc. (ticker AAPL), Bund Futures contracts maturing
in September 2015, and the US Oil ETF (ticker USO). This is a diverse range of assets,
encompassing an equity instrument, a futures contract on German 10 year government
bonds (fixed income), and an ETF closely related to a major commodity. Furthermore,
each of these assets is often anecdotally hypothesized by market participants to be
favorite for HF traders. For instance, it is known that during the events of the 2010
Flash Crash, Apple stock was briefly driven up in value to as high as $ 100 000 within a
few instantaneous trades by malfunctioning algorithms, while the majority of the other
assets were collapsing (Patterson 2013). Apple is also the most actively traded share in
the world, with an average daily volume of over 63 million shares in the last 50 days.
It means that Apple is likely to attract high levels of activity from a large amount
of diverse market participants, including the HF traders. Furthermore, the stock is
included as a component of several other actively products with a heavy weighting,
such as QQQ ETF focused on the technology sector, and the Nasdaq index. This

7
suggests that cross asset arbitrage which is the favorite HFT strategy is also available.
Additionally, being a tech-stock, the stock exhibits sufficient volatility to make short
term price fluctuations plausible and thus making it potentially more attractive to HFT
activities such as Momentum Ignition. The period of study selected for Apple is the
week spanning from January 26th till 30th of 2015 around the earning report release. A
week marked by a scheduled significant catalyst event is likely to be a week of increased
trading activity by investors. The large volume may then draw the attention of HF
traders interested in employing predatory strategies.
Similarly, traditional market makers and proprietary trading firms often lament the
increasing activity of HFT in the market for Bund contracts, and express their diffi-
culties of competing against their high speed algorithms. The Bund futures contract
which is traded on Eurex is a contract based on Fixed Income Securities issued by the
German Government. This can be for the delivery of a notional debt obligation of the
Federal Republic of Germany with a time to maturity between 8.5 and 10.5 years, and
a notional coupon rate of 6%. The contracts are issued in March, June, September
and December. This asset is extremely popular with traditional proprietary trading
firms and market makers, and it is considered one of the most accurate indicators of
the prevailing interest rates in the Eurozone. As such, it is heavily influenced by the
monetary policy decision of the European Central Bank, as well as very sensitive to
data about the variety of macroeconomic indicators such as inflation, economic growth,
unemployment, sovereign debt levels. The week selected for this study spans from June
1st till June 5th of 2015, as this week has been marked by significantly high volatil-
ity in European Fixed Income markets, referred to ”bloodath”. During this week, the
monthly monetary policy decisions and press conference were hosted by the ECB on the
3rd of June. This particular day saws a continuation of the sharp moves in the Fixed
Income markets raising the question whether sovereign yields volatility being raised
during the press conference. The ECB president highlighed technical factors, including
the volatility clustering and the poor liquidity conditions as some of the possible causes
of the moves. Nevertheless, the abnormal price reaction of the week raises the question
of possible HFT involvement, perhaps being drawn by the extreme volatility or even
instigating it. Therefore, a week of heightened trading activity and the presence of
a strong catalyst event vis-à-vis the monetary policy announcements may be a good
starting place to search for the footprint of HFT strategies.
Finally, the use of US Oil ETF (ticker USO) is particularly relevant, as it is reported as
one of the top holdings of major HFT firms 1 , such as the Virtu Financial. The US Oil
1
http://www.bloomberg.com/news/articles/2015-02-19/berkshire-hathaway-exotic-etfs-among-
flash-boy-holdings.

8
ETF is an ETF designed to track the daily movements of WTI light, sweet crude oil.
The objective of the ETF is for the daily changes in its net asset value (NAV) to reflect
the daily relative changes in the price of WTI as measured by a Benchmark Futures
Contract traded on NYMEX. The product trades in $0.01 increments, and incurs a
0.45% management fee by its administration company, Brown Brothers Harriman Co.
Much more significantly, however, USO is listed as the second largest holding for promi-
nent HFT firm Virtu Financial as of the beginning of 2015. This has several important
implications in the present context. Traditionally, the aim of Market Making Firms
like Virtu is to manage and minimize inventory risk, particularly arising from overnight
exposure. In order to achieve this, it is unusual for them to hold overnight positions in
securities. This only makes sense in a case where a particular security is very heavily
focused on, and therefore warrants incurring some overnight exposure in order to facili-
tate smoothness market making operations. This is done to ensure that an appropriate
level of inventory is maintained at all times. Furthermore, anecdotal evidence from
the industry suggests that ETF and Delta One product arbitrage is one of the most
popular HFT strategies for many firms, including the likes of Flow Traders, Optiver
and Virtu. The basic premise of this family of strategies is to exploit transient discrep-
ancies between the NAV of an ETF product, and the value of its underlying basket of
benchmark assets. This type of cross-asset arbitrage is a possible strategy with assets
such as USO. Finally, as with the other assets examined in the present research, there
are significant fundamental catalyst events occurring for the security during the time
frame used. While oil prices have experienced elevated volatility since the beginning of
2014, July 2015 marks an important geopolitical development for Oil markets. After
several days of intense high-profile negotiations between the Iran Government and the
international community, a deal was reached to lift sanctions on Iranian Oil exports, in
exchange for restrictions in the countrys nuclear programme. Fundamentally, this de-
velopment has the potential to cause significant volatility in oil markets due to changes
in supply to western markets. It can be argued that events of this magnitude may have
enticed strategic portfolio re-balancing by commodity traders. This could also mean
an increase in the activity of HFT market makers. Therefore, for the purposes of the
present research, data is used on USO trades and Level 1 quotes for the 13th and the
14th of July 2015.

3.2 Data
The data for the 3 examined assets in the project shares the same basic properties.
It consists of Trade and Level 1 Quote data. For Trades, data includes: Trade price,

9
Trade quantity and Trade Time. For Level 1 Quotes, it includes Bid Price, Ask Price,
Bid Quantity, Ask Quantity and Quote Update Time. The data is consolidated from
multiple venues and spans across pre-market trading and after Hours, as well as regular
trading sessions. The precision of the time measurement is at the microsecond levels,
whereas for prices it is determined by the relevant minimum price increment. All of
this ensures the high accuracy of the information examined, and minimizes the risks of
problems such as loss of synchronization manifesting. The order to trade ratio across
the week investigated is on average of 11%.
A central challenge which emerges when working with High frequency data over a
large time period is processing the vast amount of information. The task of finding
short episodes of HFT activity in the enormous amounts of data used is similar to the
proverbial needle in a haystack. This makes clear the need for a deliberate strategy
of analysis aimed at isolating signal from noise. Otherwise, the only alternative would
be a manual approach whereas the researcher examines personally the data. A quick
simple calculation, assuming that trades and quote updates are examined in batches
of 10 consecutive and that each batch is processed by the researcher in just 1 second
(speed which is most likely unachievable in reality) indicates that it would take a total
of over 266 hours to process the entire sample used in the current project. More
realistic assumptions could result in a processing time which is even 30-60x times longer.
Clearly, this approach is not viable, and a better, more efficient solution is needed. An
obvious starting point is to look at the algorithmic techniques used by high speed
traders themselves and to utilize similar tools for the purposes of the research. This is
the main contribution of the present paper.
To begin processing the data, it is necessary to reflect on several important properties
of the instances of HFT activity. Namely, these are usually seen as occurring over a
specific time interval marked by a starting point, a time span, and an ending point.
Similarly, as these are driven by algorithms sensitive to market conditions, the period
immediately preceding an outburst of HFT activity might also be of particular interest
in the analysis. The time spans over which HFT events last can be quite flexible and
may follow a fat tailed distribution as indicated by (Tse, Lin, and Vincent 2012). This
means that a one-size fits all approach could be wrong, and a certain level of flexibility is
warranted. Additionally, different types of events might occur over drastically different
time horizons, for instance while Quote Stuffing may only last several seconds in most
cases, and Momentum Ignition typically spans over several minutes. A truly robust
strategy for HFT detection would necessitate a sufficient built in scalability to cope
with this without any fundamental alterations. It is also important to keep track of
market and institutional conventions, such as the difference between Pre-market, regular

10
hours and After Hours trading, which will have an obvious and profound impact on the
level of activity during different times of the day. All of these considerations need to
be built into the analytical strategy to ensure that it is appropriate for the current
analysis.

4 Detection of HFT strategies

The raw data is received in plain text format and includes multiple variables relating to
trades and quotes. This is processed by the analytical tool in order to extract only the
entries relevant for the purposes of the present analysis. These are stored as variables.
As a consequence, it is immediately possible to extract statistics. For the purposes of
analyzing HFT events, the best approach is to use the method of rolling time-frames
with overlap. Since the data points are unevenly distributed in time, an algorithm is
used to collate them into subsamples (windows) spanning a user specified time length.
Therefore, each data point serves as a starting point for a window which includes a
number of other data points which fall within a pre-specified time from the first one.
This is a flexible approach which allows the data to be analyzed frame by frame, and for
statistics to be estimated for each window. Furthermore, since the windows are derived
with overlap, this means that all potential frames are examined. A big advantage
of this approach is that the researcher is able to specify a desired window size (in
microseconds), which can mean that HFT events of differing length can be analyzed
by the tool. The window subsample framework is at the center of all the detection
algorithms used in this paper.
Detection of HFT activity in large data samples is the main objective of the current
research. As noted previously, this is a challenging task which calls for an algorithmic
solution where little or no human input is required, simply owing to the volumes of data
examined. Fortunately, the software tool used in this project provides an efficient envi-
ronment that can be used to implement such algorithms. The time window framework
allows for statistics to be estimated for each of the rolling subsamples. This simplifies
the task of detecting the time intervals containing HFT activity to designing statistics
which capture the similarity of the pattern of activity observed in a given subsample to
that of typical HFT patterns. The one arbitrary element in this strategy is the length
of time frames examined. Fortunately, market observations provide some hints for suit-
able time frames for both Quote Stuffing and Momentum Ignition. Alternatively, these
can be determined empirically by trial and error.

11
4.1 Quote Stuffing
Quote stuffing is an event during which high quote volatility is observed over a very
short period of time, typically lasting several seconds. During this time, rapid but
transient quote updates occur, often following several specific patterns. Figures 1 and
2 show some of these patterns.

Figure 1: Apple In-bid stuffing

Figure 2: Apple In-ask stuffing

Figure 3: Bund Combined stuffing

Additionally, Quote Stuffing can occur either on the bid side, the ask side of the market
or both simultaneously as shown in figures 1 to 3 above. In order to detect the presence
of these patterns of activity in the subsample contained in the data, the D-ratio statistic
is estimated. This can take on several alternative specifications depending on the type
of Quote Stuffing observed bid, ask or combined. The D-ratio is a geometry based
metric, that is inspired by the graphical representation of Quote Stuffing on a chart.
In essence, all Quote Stuffing episodes share several key characteristics, irrespective of
the particular pattern.

• They include small rapid movements in the bid, ask or both levels, which are
subsequently rapidly reversed.

• This is repeated a large number of times, over a small time frame.

• Over the span of the entire time frame, the actual direction movement in the
quote levels is low, if any.

12
The D-ratio has four components: (i) Carryask; (ii) Bigask; (iii) Carrybid; (iv) Bigbid.
The window (subsample) size is denoted as j, each Bid quote price is denoted as B, and
each Ask quote price is denoted as A. Therefore the four component variables are: The
D-ratio has four components: (i) Carryask; (ii) Bigask; (iii) Carrybid; (iv) Bigbid. The
window (subsample) size is denoted as j, each Bid quote price is denoted as B, and each
Ask quote price is denoted as A. Therefore the four component variables are:

1. Carryask is the sum of absolute incremental (instant-by-instant) changes in the

ask price over the period.

j
X
Carryask = |Ai − Ai−1 | (1)
i=2

2. Bigask is the absolute change in the ask price level between the starting and the
ending points of the period examined.

Bigask = |Aj − A1 | (2)

3. Carrybid is the sum of absolute incremental (instant-by-instant) changes in the

bid price over the period.
j
X
Carrybid = |Bi − Bi−1 | (3)
i=2

4. Bigbid is the absolute change in the bid price level between the ending and the
starting point of the period.

Bigbid = |Bj − B1 | (4)

These above variables are used to compute three alternative specifications of the D-
ratio:

• Ask specification aims at detecting the in-ask Quoting Stuffing activity. This
implies rapid quote volatility on the ask side, and a relatively passive bid side.
The specification is:
carryask
bigask
D= carrybid
bigbid

In order to guarantee the function’s solutions domain, several special cases are defined:
(i) if bigask=0, it is instead set at the level of the minimum tick increment at 0.01; (ii)

13
if bigbid=0, it is instead set at the level of the minimum tick increment at 0.01; (iii) i
carrybid
carrybid=0, the entire denominator bigbid
is set to equal to 0.01. This ensures that a
corresponding D-ratio can be calculated for any given subsample.

• Bid specification aims at detecting Quote Stuffing which occurs on the bid side of
the market. This is characterized by rapid quote volatility in the bid price levels,
while the ask price remains relatively inactive. The specification is given by:
carrybid
bigbid
D= carryask
bigask

For solution domain reasons, several special cases are defined: (i) if bigask=0, this is
instead set at the minimum tick increment at 0.01; if bigbid=0, this is instead set at
carryask
the minimum tick increment at 0.01; if carryask=0, the entire denominator bigask
is
set to 0.01 instead.
• Combined specification aims at detecting Quote Stuffing Activity of the combined
type which occurs on both the ask and bid sides of the market. This is characterized
by a period of high quote price volatility which occurs over a short period of time,
but it is driven by transient movements. The specification is given by:

carryask carrybid
D= +
bigask bigbid

For the purpose of solutions domain considerations, several specific cases are predefined:
If bigask = 0, it is instead set at the minimum incremental tick size at 0.01.
If bigbid = 0 , it is instead set at the minimum incremental tick size at 0.01.
Figures below present some of examples of the results derived using the D-ratio.

Figure 4: USO combined stuffing

14
 Figure 5: USO in-bid stuffing

Figure 6: Apple in-ask stuffing

The final step of the detection strategy, after the D-ratio values for each window in
the sample have been calculated, is to determine which ones are indicating a potential
period of HFT activity. Since a unique window (subsample) is associated with each
data point in the sample, and a D-ratio value is associated with each window, it is
possible to use the observed distribution of D-ratio values over the entire sample, and
subsequently select a cut-off point for the most promising ones. The Trident tool
supports a user specified cutoff point. Once a D-value has been estimated for each
window in the sample, the entire array of D-ratio values is ordered in increasing order.
The user specified percentage is then applied to select the cutoff point. This is done
via the below formula:

Dcutof f = Darraysize−(rounddown(percentile∗arraysize)) (8)

The cutoff determined using this technique has the major benefit of coming from the
distribution observed within the actual data, rather than an arbitrary level selected by
the researcher. Since a higher D-ratio should indicate higher likelihood of HFT activity,
using the observed distribution of the values within the sample allows the researcher to
focus on a subsample that appears to be most relevant within a given context. Once the
cutoff is determined, it is used to filter out only the windows (subsamples) which have
a D-ratio value above the cutoff point. These can then be examined using additional
automated techniques or manually by the researcher. For the purposes of most of the
present analysis, the second approach is taken.

4.2 Momentum Ignition

Momentum Ignition is an HFT strategy that is characterized by a specific pattern
observed in both trade prices and trade volumes. This usually comprises three main
stages:

• An initial spike in trade volume, which isnt accompanied by any significant changes
in price.

15
 • A subsequent sharp price move (positive or negative), accompanied by a new, even
larger increase in volume

• A gradual price reversal to levels observed before the event, accompanied by low
volume.
This pattern is observed over significantly longer time frames than Quote Stuffing, and
may last for up to several minutes. It is also relatively less frequent in occurrence,
although still prevalent in most traded instruments at least once per day with higher
activity in certain sub sectors of the market. The duration of the events, as well as the
severity of their market impact appears to follow a fat-tailed distribution, with a small
fraction of events having major market impact and lasting for a prolonged period of
time. Unlike Quote Stuffing where most of the economic impact is indirect resulting
from changes in the midpoint or enticing potential traders, Momentum Ignition has
a directly observable economic impact, which can be measured in relative terms (size
of the price move in bp), or, potentially even in absolute (price change x estimated
position by HFT).
The strategy used for Quote Stuffing detection is modified and applied for detecting
Momentum Ignition. While the characteristic pattern of Momentum Ignition includes
two dimensions, trade prices as well as volume for the sake of computation efficiency
and simplicity, only trade prices are used an input for the D-ratio specification used to
detect the pattern. The software has additional built in capabilities which allow the
researcher to easily examine the volume data for the characteristic patterns following the
initial screening. Furthermore, it is assumed that, as consistent with previous empirical
studies of financial markets, the distribution of asset returns exhibits leptokurtosis, and
therefore a small enough fraction of subsamples (windows) will contain the large moves
relevant for studying Momentum Ignition. The distribution derived strategy and the
specification of the D-ratio used ensure that the biggest relevant price moves present
in the data are examined. The D-ratio used for Momentum Ignition detection is based
on 3 key inputs:
1. Starting trade price is the Trade price in the starting point of the time period.

2. Ending trade price is the Trade price of the final trade in the time period.

3. Price Span defined as |EndP rice − StartP rice|. If this turns out to be 0, then it
is set to 0.01 instead for domain purposes.
As with Quote Stuffing, the metric estimated is inspired by the geometry of a graph-
ical representation of the pattern. In the case of Momentum Ignition, this involves
estimating two distances for each trade (t) included in the subsample (window):

16
 D1t = |Pt − StartP rice|
(9)
D2t = |Pt − EndP rice|
These metrics are used to derive a Total Distance:

T D1t = D1t + D2t (10)

Once this is derived for each trade in the subsample, the largest TD is determined:
n
X
T Dmax = max T Di (11)
i=0

These inputs can be used to derive the value of the D-ratio for the window:

T Dmax
Dsample = (12)
P riceSpan
Once a D-ratio value is estimated for each window in the sample, the array of D-ratio
values is ordered and a cutoff point is determined. This is then used to filter out the
top values encountered in the sample. The focus on price data only means that this
approach will select the windows with the biggest price moves which have subsequently
reversed back to their starting point. Once these are determined, the researcher can
manually examine each of these and use the built in functionality of the Trident Tool
to look for the characteristic pattern in volume, finally yielding a confirmed finding.

5 Methodology
While the detection of HFT activity is the main objective of the present research, the
study is broadened by using further analytical tools to obtain additional information
regarding the impact and performance of high speed traders on the market. The an-
alytical suite built into the Trident Tool provides an array of ways to extract that
information efficiently from the data. This includes metrics such as the Quoted Spread,
the Level 1 Quoted Depth, Midpoint data, Trading Volume, Trade Volume Moving
Average, VWAP and Implementation Shortfall.

• Quoted Spread is one of the most widely used indicators of trading costs in finan-
cial markets. While it is not necessarily indicative for very large orders, which may
deplete Level 1 liquidity and walk up or down the book, it is a generally good in-
dication for average trades.

17
• Level 1 Quoted depth is custom built indicator which captures the total quantity
offered at the best ask and the total quantity demanded at the best bid. It is a
good indicator of market liquidity as it shows the actual potential trade size which
the market can transact without too much impact. It also appears to act as an
indicator during quote stuffing episodes.

• The Midpoint is a popular indicator used by market researchers and practitioners

alike. This is often seen as a metric indicating the relative performance of a trade,
or could be used to derive more sophisticated measures of the spread. Further-
more it is incorporated into the trading strategies of institutional investors, and
thus may influence the behavior of block trades. Finally a very important usage of
the Midpoint is observed on some Dark Pools, which use a crossing rule based on
the midpoint of the lit markets. These factors make the Midpoint one of the most
important metrics for the practitioners of strategies such as Quote Stuffing.

• VWAP The Volume-Weighted Average Price is another extremely popular metric

used by industry and academia alike. It is a measure derived for a specific time
period, and may be used for the purpose of finding a realistic assessment of the
most likely trade price over that time. It is often used as a simple benchmarking
tool for the performance of trading desks at institutional investors. It is also used
a benchmark in many algorithmic strategies. For instance a trader who manages
to buy securities at below VWAP, and sell them at a price above VWAP during a
certain time period, can be assumed as having contributed value. A trader who has
followed VWAP without deviations is known as having delivered accurate execution.
A popular alternative strategy, used by many institutional investors is the constant
participation rate strategy, where a large order is worked by being split into child
orders of small sized and executed throughout the day in proportion to the average
trade sizes observed throughout a day. The VWAP can be seen as providing a more
accurate view of the prices of securities, as it assigns a lesser weight to abnormal
volatility stemming from low volume trades. In the context of the present analysis,
the VWAP over a time period (window) is calculated recursively. This applied to
recursively calculate VWAP at each moment within the window examined.

• Implementation Shortfall is a metric which builds onto VWAP, and is also widely
used by market practitioners. This is aimed at measuring the execution performance
of traders and algorithmic strategies, by benchmarking it against a hypothetical
paper portfolio executed at the midpoint once the order is received. In the context

18
 of the present research, this is coupled with VWAP to obtain of measure of market
movement. The total quantity traded during the window examined is assumed
as the order being worked by the market. The result is a metric following the
price movements during the period, but is adjusted for the initial midpoint. It is
calculated assuming a buyer point of view, therefore a positive value indicates that
a buyer would have been better off executing immediately at the midpoint at the
start of the time period examined (window), rather than delay execution partially
or fully. Similarly, negative values indicate that the price moves lower through the
window so from a seller point of view, it is ideal to execute immediately.

• Trading volume is displayed as a metric which is crucial for the correct identification
of Momentum Ignition patterns. This allows the researcher to quickly identify
trends in volume.

• Trade Volume Moving Average is a built in metric which is aimed at simplifying

detecting trends within large amounts of trade volume data. This is particularly
useful when examining time windows containing a large number of data points. The
moving sample for the average is set at 100.

Figure 7: Bund Quoted Spread

Figure 8: Apple Quoted Spread

Figure 9: Bund Level 1 Quoted Depth

19
Figure 10: Apple Level 1 Quoted Depth

Figure 11: Apple Trade size

Figure 12: USO Trade size

Figure 13: Apple trading volume

Figure 14: Bund VWAP

Figure 15: Apple VWAP

20
 Figure 16: Bund Implementation Shortfall

Figure 17: Apple Implementation Shortfall

5.1 Artificial Neural Network

The final objective in the present research is an attempt at detecting potential com-
monalities which may signal that an episode of HFT activity may be imminent. This
is achieved through performing an Artificial Neural Networks experiment. There
may be reasons to believe that at least part of the HFT activity may be predictable
to some extent. Fundamentally, algorithmic trading relies on an algorithms being
triggered by market conditions. If these conditions were known, it would also be
possible to forecast when algorithmic activity may be imminent. However, in real-
ity, this information would constitute a very closely guarded company secret, and
is almost certain to be protected as intellectual property. Therefore, an alternative
method is to attempt to estimate and detect commonalities in market conditions
immediately preceding an episode of HFT activity.
Artificial Neural Networks (ANN) are types of statistical learning models which
are designed in a way that mimics the logical structure of a biological brain. These
models are particularly good for the purposes of pattern recognition, and are used in
a variety of applications, including speech recognition, image recognition, analyzing
patterns of consumer behavior and in financial markets. Fundamentally, all ANN
models require at least two basic characteristics a topology and a transfer function
(Bishop 1995). ANNs are constructed out of nodes called neurons which act as
simple I-O transformers. Data is fed into neurons as a signal input, and this is
processed via a transfer function which generates an output signal. There are
multiple transfer functions available, which have different characteristics and may
be appropriate for analyzing specific problems. Some of the most widely used

21
ones include the logistic function, the linear function, a hyperbolic function and
a threshold function. For instance, a logistic transfer function implies that the
value of the potential outputs may range between 0 and 1. The derivative of the
transfer function has important implications for the performance of the module
during learning on training data sets. For the purposes of the present research, the
hyperbolic function is used:
O = tanh(I) (13)

The derivative of the hyperbolic function is approximated by:

1 − x2 (14)

This ensures that outputs can take on values between 1 and -1, as can be see below.
Additionally, a large central region of the function is characterized by a relatively
constant slope, allowing for strong learning performance in a wider region of input
values:

Figure 18: Plot of the hyperbolic transfer function

The neurons of an ANN are structured in groups called layers, and while there is
no limitation to the number of layers, and the number of neurons in a layer, most
topologies will consist of 3 layers an input layer, a hidden layer, and output layer.
Each neuron in a layer is connected to all the neurons on the layer immediately
preceding it, and to an additional bias neuron, which has a constant output. These
connections are assigned to a specific weight each, and the weighted sum of the
signals coming from all connections forms the total input. The input layer neurons
are used as input nodes, where raw data feeds into the network directly. This is
then processed via the transfer function of the neurons, and fed via connections to
the hidden layer, which then processes the signal and transmits it to the output
layer. The output layer generates the final output of the network.

22
Training is a key stage of using an ANN in a practical solution. One of the major
advantages of using this type of model, is that it doesnt require prior knowledge
of the features or relationships which are influencing the data analyzed. Instead,
these are inferred by the ANN through a process of iterative learning. This only
requires that the set of inputs and their corresponding outputs is known ex ante,
but not their actual relationship. During learning, ANN models process a data set
designated for training and utilize an algorithm to adjust their connection weights
so that their outputs converge closer to the desired values. While there are many
strategies documented in the literature, the most popular learning algorithm is
backpropagation (Rumelhart, Hinton, and Williams 1985). Backpropagation is a
strategy which adjusts network connection weights using the derivative of the trans-
fer function. The information during learning flows in the opposite direction to the
flow observed during processing. This begins at the output layer with a comparison
between the ANNs current output and the target output known ex ante. This is
used to calculate the deviation between the two also known as error. The deriva-
tive of the transfer function is then used to make adjustments to connection weights
further down the network, until all the connections are updated. The new infor-
mation learned is incorporated into the connection weights. The backpropagation
algorithm is used for the purposes of the present experiment.

Figure 19: Information flows within Artificial Neural Network

There are many reasons why ANNs may be a suitable technique for carrying out the
present experiment. The evaluation of market activity over a short period of time
can be seen, in essence, as a pattern recognition exercise. Similarly, the existing
commonalities immediately preceding an episode of HFT activity, if present, are not
known to the researcher ex-ante. However ANNs do not require such knowledge,
as long as all the necessary raw data is fed into the model. Finally, the question of
whether a certain window of activity is immediately preceding an HFT episode, can
be restated as a Boolean problem, with 1 denoting a period preceding HFT, and -1
any other. The narrower focus of the present experiment is on Quote Stuffing and
is specifically looking for a graphical pattern in quote updates immediately prior to
the episode of HFT activity. The data generated in the first section of the present
research is crucial as input sample for the ANN experiment.
One significant challenge when analyzing two-dimensional data points using ANN
models is posed by what is known as the curse of dimensionality. This is a catch

23
all phrase for many diverse issues arising from the problem of representing two
dimensional features in a format suitable for ANN processing. There is a significant
body of literature detailing alternative strategies for dealing with this set of issues.
For the purposes of the present research, a simplistic approach is adopted, based on
2-D image processing strategies (Egmont-Petersen, de Ridder, and Handels 2002).
Each window of quote updates examined, is seen as a 2 dimensional area in time and
price. This is further segmented into a number of sections of equal area. The exact
granularity of the division along the X and Y axes is determined by the researcher
and can be set within the ANN suite of the Trident tool. For the purpose of the
present research a granularity of 4 in Price and 5 in Time is selected, yielding 20
segments of equal area. Once these regions are determined, the number of quote
update events falling within each segment is estimated, and calculated as a fraction
of the total number of quote updates in the time window examined. The end result
is an array consisting of 20 fractions denoting relative event density, which sum
up to 1. This approach is very similar to the one used in image processing, where
images are segmented into areas and pixel counts are performed in each segment to
transform the shape of the image into digital form.

Figure 20: Initial segmentation of a chart sample. For demonstration purposes

granularity is set to 4 in both dimensions

Figure 21: Representation of the 4 input nodes corresponding to the 4 chart

segments

Figure 22: Event Density within each region. There are a total of 30 events (quote
updates) over the sample period

24
 Figure 23: Representation of the result by filling each region with a % pf black
color in accordance with the event density calculated

The resulting set of inputs is readily processed by an ANN model. For the purposes
of the present research a training sample of 638 observations is used, with 319
windows immediately preceding a previously detected Quote Stuffing episode, which
are assigned a target value of 1, and 319 randomly selected alternative samples which
are assigned a desired output value of -1. The ANN models are used to process
600 iteration of the training dataset, and once this is accomplished, a final holdout
sample consisting of 50 periods with a target value of +1 and 50 periods with a
target value of -1, is used for evaluation purposes.

6 Results and discussion

6.1 HFT strategies and their impact

At a first glance, the present analysis appears to be successful in achieving its main
objective of HFT detection. Throughout the data samples examined, a total of 484
potential HFT events are identified, among which 372 episode of Quote Stuffing and
112 episodes of Momentum Ignition. While there is no possibility of verifying these
findings within the scope of the present research, some of the patterns observed
seem to closely match previous findings in the literature. This may indicate that
the detection techniques utilized are appropriate. A further breakdown by time
frame show the window sizes used to detect each observation. For Quote Stuffing
the majority of events occur within the bigger time scale examined (10s), whereas
Momentum Events appear to be more transient, with the majority occurring within
30s, followed by 60s and 90s, as shown in figure 24.
Furthermore, a breakdown within the group of Quote Stuffing events shows that
the distribution by strategy subcategory between In-Ask, In-Bid, and Combined are
similar in terms of occurrence. A conjecture suggests that Quote Stuffing is used
to manipulate the market environment in order to provoke institutional investors
to trade. Therefore, trading activity during episodes of Quote Stuffing is expected

25
to be adversely affected by the HFT strategy. For instance, (Tse, Lin, and Vincent
2012) note that during quote stuffing events, the trade prices tend to move in the
direction of the stuffing activity i.e. increase when stuffing occurs in-ask or decline
when it is happening on the bid side. These effects should be observable in the data,
and are therefore tested for. Further results indicate that a majority of ask and bid
events are accompanied by trading activity. Furthermore, the characteristic pattern
expected is confirmed by using the Implementation Shortfall metric. When IS drift
lower to negative values during the time window examined, this is an indication of
declining prices. Similarly, as it increases and remains positive, this is an indication
of prices rising. Another event is shown below:
It appears that the majority of Quote Stuffing episodes accompanied by trading
exhibit the characteristic pattern. These results are especially important since they
confirm previous observations. Also they confirm the potential indirect impact
of the strategy on market prices. In the process of analyzing the results, a new
characteristic pattern is observed which seems to link quote price volatility with
Level 1 quoted depth volatility. A frequent observation during Quote Stuffing is
that quote updates which narrow the quoted spread appear to be associated by a
significant decrease in Level 1 quoted depth. This pattern is very pronounced and
may have important implications for correctly interpreting the market impact of
HFT activity. Examples are presented below:

Figure 25: Apple quote stuffing in-bid and The level 1 depth at the same interval

26
 Figure 26: Bund stuffing in-bid / combined and Level 1 depth

While these results could potentially be caused by trading activity depleting quoted
depth in the order book, the characteristic pattern is also similarly observed during
episodes which involve no trades at all. This suggests that the change in volume
levels is due to new quotes being posted rather than old ones being depleted. Addi-
tionally, these newly posted orders are characterized by very low quantities offered,
which is another evidence of HFT activity. One of the most significant impacts of
the increasing presence of HFT on markets is a steady decline in the average trade
size. This result may be interpreted as indirect proof that the events observed may
be caused by the involvement of low latency traders. The observation of small or-
ders being posted and cancelled rapidly over a very short period of time fits the
patterns expected of HFT market makers. Moreover, this also lends support to the
argument that liquidity provision by HFT may be transient in nature. Finally, this
pattern is also consistent with the technique of pinging, since the orders posted
narrow the spread and may be intended to entice block traders into execution.
A final pattern is observed at the event level which confirms the intuition that
the quoted spread also experiences volatility particularly during one-sided (in-ask
or in-bid only) Quote Stuffing episodes. This finding seems to suggest that while
HFT seems to provide liquidity through posting small but competitive orders which
initially narrow the bid-ask spread, their rapid cancellation increases the volatility
and may actually increase trading costs over the long run or even introduce an
additional risk-premium for traders. This could potentially offset some or all of the
benefits of added liquidity by HFT market makers. A thorough investigation of
this hypothesis is beyond the scope of this paper, and may serve as a suggestion for
future research in the area of asset pricing (Pastor and Stambaugh 2001)

27
 Figure 27: Apple in-ask stuffing and spread volatility

Figure 28: Apple in-bid stuffing and spread volatility

An examination of the intraday patterns of the D-ratio values reveals characteristic

peaks around the beginning and the end of the regular trading hours from 9:00 a.m.
till 16:00. This was observable across all the three assets as shown below in the
figures:

Figure 29: Intraday patterns of D-ratio - Bund

Figure 30: Intraday patterns of D-ratio - APPLE

An examination of the distribution of D-ratio values seems to strongly suggest

that these follow a Chi-squared distribution, characterized by a fat tail. This is a
finding which warrants further investigations and could potentially lead to a more

28
formalized quantitative method of detecting HFT activities or their probability of
occurrence.

Figure 31: Bund combined stuffing D-ratio histogram

Momentum Ignition events are observed on time frames of 30, 60 and 90s. The
total size of the price movement for each even is recorded in basis points, as well as
the potential volume traded by the HFT traders. The method used for this is an
approximation based on the volume observed during the initial volume peak and
the secondary volume peak. It is assumed that the strategy practitioners use the
initial increase in volume to position themselves accordingly, and once they succeed
in instigating a price move, unwinding their position during the secondary volume
peak. This is a strong assumption, but the aim of the analysis is simply to provide
a potential estimate of the direct economic result derived by the practitioners of
the strategy. Future research could refine the methodology used in this estimation,
or obtain confirmation of the results from industry.

Figure 32: Momentum Ignition Patterns - Apple

While the average return per event observed in the sample is 23,09 bp, this number
is significant when considered within the context of the extremely short time frames
of its occurrence between 0,5 and 1,5 minutes. The largest observed relative return
in the sample is 106,7 bp, during an event on the Bund futures market. Using
the methodology outlined above, the potential gross profit generated during this is
EUR 82692,5. This is calculated via the below formula:

29
 Returninbp
GrossP rof it = Quantitytraded ∗ IndicativeP rice ∗ ∗ T ickSize (15)
10000
Indicative price is a rough indication of the relevant assets price. For Bund futures,
this is assumed at a constant EUR 155, for Apple shares at $113, and for the USO
ETF $17,36. The tick size of the Bund Futures contract is for a nominal value of
EUR 10 per 0.01 change in price.
An examination of the intraday D-ratio values chart for Momentum Ignition reveals
a similar pattern to the one observed for Quote Stuffing. The pattern observed for
Bund contracts is different than the one for the exchange traded assets in the US.
This is due to the longer regular hours trading session on EUREX, as opposed to
the market hours observed on US exchanges, where both Apple and USO are listed.
A start of trading, end of trading and mid-day peaks are observable for most assets.

Figure 33: Intraday patterns of D-ratio - Bund

Figure 34: Intraday patterns of D-ratio - USO

The distribution of the D-ratio values appears to follow the Chi-square distribution
pattern observed for Quote Stuffing, but with even longer fat tails and greater
skewness, especially for Apple. This suggests granularity in the data, and the
presence of extreme outliers as could expected during the episodes of HFT activity.
This also provides additional basis for a further future quantitative research on
this distribution, which could shed more light on the probability distribution of
HFT activity occurrences. The total profits generated by the Momentum Ignition
strategy over the sample studied amounts to about $ 5,25million. A glance at
the breakdown of event count by day of the week, reveals no particular pattern,
although it seems to suggest that midweek days may tend to contain higher HFT
activity.

30
 Figure 35: USO D-ratio histogram

Figure 36: Bund D-ratio histogram

6.2 Artificial Neural Networks Experiment

The ANN experiment is carried out as per the methodology detailed above. This is
a simplistic approach, aimed mainly at providing a proof of concept and guidance
for further future research. The results seem to suggest that some of the models
examined may have forecasting potential. Nevertheless, there are significant caveats
with the approach used in this project. A typical problem when using ANN tools,
is the large number of potential specifications, and the sensitivity of results to the
network architecture used. Small changes to topology may yield extreme differences
in performance. Therefore prior selection of the optimal topology is extremely
difficult, especially when there is little or no knowledge of the relationship being
studied ex-ante. Furthermore, ANN models are very sensitive to the quality of input
data. For instance, the present approach focuses exclusively on graphical patterns
observed on charts of market activity. The model isnt supplied data about the actual
price levels or time frames examined, neither of global long run statistics, such as
typical average volatility and volume levels, or other indicators. Similarly, the 20
regions defined in the current experiment may be an insufficient resolution to fully
capture the relevant graphical patterns. The experiment detailed here could easily
be expanded to include additional information. However, care should be taken to
limit the number of input variables, as too many could lead to introducing noise
in the data, or to over-fitting. A similar tradeoff is faced when adjusting network
parameters such as the learning rate and the learning momentum , as well as the
number of neurons in the hidden layers. Adjusting all of these parameters may
potentially lead to faster convergence, or the inference of additional features from

31
the data, however as a downside these may lead to focus on spurious patterns, over-
fitting or even complete failure to achieve convergence. The table below summarizes
the results of several alternative specifications, along with the basic parameters and
topology used. Forecasts of negative output, are taken as signaling a result of -1,
while positive outputs are seen as signaling +1.While it is difficult to benchmark
precisely the forecasting performance observed, a simple rule of thumb is to check
whether the forecasts add any incremental value to a nave forecast of 50%.

7 Conclusion

This paper attempts to provide a fresh empirical perspective on the topical area of
High Frequency Trading and has four main objectives. A novel empirical method-
ology is introduced aimed at detecting the presence of two commonly encountered
HFT strategies in financial market Quote Stuffing and Momentum Ignition. Fur-
ther analytical tools are used to assess their performance, impact on the wider
market and empirical implications for future research. An experiment using ANNs
is carried out to test potential forecasting techniques which may be used to de-
tect these phenomena. All of these instruments are implemented via a proprietary
analytical tool developed in the C++ language which has the merits of high per-
formance in processing large data sets, and a lightweight user interface. Section
1 introduced briefly current topics relating to HFT and their general implications
for finance academia and industry. Section 2 introduces in detail different HFT
strategies, their objectives, implementation mechanisms and potential impact. Sec-
tion 3 introduces the data and analytical instruments used in the current project.
Section 4 presents the results of this analysis and discusses the most important find-
ings and their implications. There are several important weakness of the present
project. The main one is the lack of ability to validate the findings presented.
Without access to additional data from the industry, it may not be possible to
conclude with certainty whether particular market activities can be attributed to
HFT or not. Furthermore it is possible that some of the analytical tools used are
inappropriate or not calibrated correctly to deal with the objectives of the project.
Finally during the course of the current project, several important findings emerged
which may serve as a stepping stone for future research. Some these are the spread
and volume patterns observed during Quote Stuffing activities, the direct economic
performance of Momentum Ignition and the nature of the underlying distribution
followed by the D-ratio indicator. The ANN experiment, while serving as a proof

32
of concept, could be vastly improved by introducing additional variables, improv-
ing network topology and calibrating its parameters accordingly. Nevertheless, the
current project provides some useful analytical tools which can hopefully be of help
for other researchers in the field of High Frequency Trading.

33
References
Biais, Bruno, and Thierry Foucault. 2014. “HFT and market quality.” Bankers,
Markets & Investors 128:5–19.
Biais, Bruno, Thierry Foucault, and Sophie Moinas. 2015. “Equilibrium fast
trading.” Journal of Financial Economics 116 (2): 292–313.
Bishop, Christopher M. 1995. Neural networks for pattern recognition. Oxford
university press.
Egmont-Petersen, Michael, Dick de Ridder, and Heinz Handels. 2002. “Image
processing with neural networksa review.” Pattern recognition 35 (10): 2279–
2301.
Foucault, Thierry, Johan Hombert, and Ioanid Roşu. 2015. “News trading and
speed.” The Journal of Finance.
Kyle, Albert S. 1985. “Continuous auctions and insider trading.” Econometrica:
Journal of the Econometric Society, pp. 1315–1335.
Lewis, Michael. 2014. Flash boys: a Wall Street revolt. WW Norton & Company.
OHara, Maureen. 2015. “High frequency market microstructure.” Journal of
Financial Economics 116 (2): 257–270.
Pastor, Lubos, and Robert F Stambaugh. 2001. “Liquidity risk and expected
stock returns.” Technical Report, National Bureau of Economic Research.
Patterson, Scott. 2013. Dark pools: The rise of the machine traders and the rigging
of the US stock market. Crown Business.
Rijper, Thierry, Willem Sprenkeler, and Stefan Kip. 2010. “High frequency trad-
ing.” Position Paper. Optiver Holding BV, Strawinskylaan 3095:1077.
Rumelhart, David E, Geoffrey E Hinton, and Ronald J Williams. 1985. “Learn-
ing internal representations by error propagation.” Technical Report, DTIC
Document.
Tse, Jonathan, Xiang Lin, and Drew Vincent. 2012. “High frequency trading–
measurement, detection and response.” Credit Suisse, Zürich, Switzerland,
Tech. Rep.
Turbeville, Wallace C. 2013. “Cracks in the Pipeline, Part Two: High Frequency
Trading.” Demos, March.

34