Professional Documents
Culture Documents
Anton Golub
References
Strategies
Rebate Trading
Exchanges pay HF traders that post best limit sell or best limit buy
orders - Liquidity Rebate: $0,003/share.
HF traders plan:
I place a limit buy order
I if his order is matched, he turns around and place a limit sell order
at the same price
I profit from trading will be $0,00, but he will collect rebates!
Rebate Trading
Price Shares
best offer 10,02 600
best bid 10,01 100 ← HFT buy order
bid 10,00 100
Price Shares
offer 10,02 600
best offer 10,01 100 ← HFT sell order
best bid 10,00 100
Flash Trading
exchange A exchange B
Price Shares Price Shares
offer 10,02 600 10,02 400
best offer 10,01 500
best bid 10,00 100 10,00 200
Flash Trading
Flash Crash
FLASH CRASH
Example of Up Crash
INET ARCA
bid/ask price size price size
ask 32.02 500 32.02 500
ask 32.01 500
best ask 32.00 500
best bid 31.99 100 31.99 100
Table: Order book of security XYZ
Market order to buy 1000 shares is sent to ARCA; 500 shares will be
routed to INET (32.00), but the remaining 500 shares will be filled on
ARCA (32.02).
References
The aim of this paper is to study statistical properties of flash crashes using simple
data mining techniques GIVE LIST AND WHY AT SOME POINT on tick-by-tick trans-
action data. WE DON’T REALLY DISCUSS DATA MINING TECHNIQUES
1
1 millisecond is one thousand of a second
1
Previous work has suggested that one-sidedness of the market can be the cause for
flash crash; Easley et. al. [3] construct Volume-Synchronized Probability of Informed
Trading (VPIN)2 and show that this metrics had among the highest values in on May
6th. Chakravarty et. al. [13] use Intermarket Sweep Orders (ISO)3 as a proxy for
aggressive liquidity taking, and find that elevated use of ISO on May 6th. Kirilenko et.
al. [1] performed extensive analysis of May 6th Flash Crash and found that high frequency
traders did not trigger the crash, but their activities increased market volatility. Nanex
Llc., used order book data [9] of May 6th and found that quote stuffing 4 and delays in
NYSE Consolidated Quotation System caused the crash.
The aim of this paper is to study statistical properties of flash crashes on tick-by-tick
transaction data. We investigate a number of relationships that have not been reported
in the previous work: the price movement behaviour before crashes, the influence of flash
crashes on bid-ask spread, the average price before the crash compared with the crash
intensity and the effect of particular sectors and geographic region on the crash etc. Our
conclusions suggest... IN PARTICULAR, WHAT IS UNUSUAL OR UNEXPECTED?
2 Background
2.1 May 6th Flash Crash
In this section we present some of the work on May 6th Flash Crash.
Kirilenko, Kyle, Samadi and Tuzun [1] describe the market structure of the E-mini
S&P 500 stock index futures market on the day of the Flash Crash. They use audit-
trail, transaction-level data for all regular transactions to classify over 15,000 ?CHECK?
trading accounts that traded on May 6 into six categories: High Frequency Traders, Inter-
mediaries, Fundamental Buyers, Fundamental Sellers, Opportunistic Traders, and Small
Traders. They asked the following: how did High Frequency Traders and other categories
trade on May 6? what may have triggered the Flash Crash? what role did High Frequency
Traders play in the Flash Crash?. They concluded that High Frequency Traders did not
trigger the Flash Crash, but their responses to the unusually large selling pressure on that
day exacerbated market volatility. The U.S. Securities and Exchange Commission report
2
VPIN is trademark of Tudor Investment
3
Intermarket sweep orders are limit orders that require they be executed in one specific market center
even if another market center is publishing a better quote
4
Quote-stuffing is a malicious technique of placing and then almost immediately cancelling large
numbers of rapid-fire orders to buy or sell stocks
2
on May 6th Market Events[12] was based on this paper.
Easley, Lopez de Prado and O’Hara [4] highlight the role played by order toxicity
in affecting liquidity provision, and show that a measure of this toxicity, the Volume-
Synchronized Probability of Informed Trading (VPIN) [3], captures the increasing toxicity
of order flow in the hours and days prior to collapse. As the ”flash crash” might have
been avoided had liquidity providers remained in the marketplace, a solution is proposed
in the form of a ”VPIN contract” which would allow them to dynamically monitor and
manage their risks. IS THEIR DETAIL ON THE LIQUIDITY?
Chakravarty, Upson and Wood [13] investigate the contribution of trading aggressive-
ness and liquidity supply to the flash crash. They find that trading aggressiveness was
significantly higher for the entire day of the flash crash, as proxied through the use of
Intermarket Sweep Orders (ISO). ISO have been found to be primarily used by informed
institutional traders and are allowed to trade through the best prices in the market. They
show that the information content of ISO trades on the day of the flash crash was highly
informed, as measured by the information shares method of Hasbrouck [8]. During the
flash crash, while the ISO volume was on 32% of total volume, ISO trades account for
over 50% of the contribution to the price variance. Faced with this large increase in in-
formed trading, liquidity suppliers withdrew from the market, decreasing overall liquidity.
They show that as liquidity decreases in the market, it leads to an increase in ISO use,
accelerating the market decline, but also speeding the market recovery - they recommend
considerations of an ISO halt during periods of high market wide volatility. HOW TYP-
ICAL IS THIS ISO BEHAVIOUR AROUND THE TIME OF THE FLASH CRASH
Lee, Cheng and Koh [2] produce 9 different simulations created by using a large-
scale computer model to reconstruct the critical elements of the market events of May 6,
2010. The resulting price distribution provides a reasonable resemblance to the descrip-
tive statistics of the second-by-second prices of S&P 500 E-mini futures from 14:30 to
15:00 on the day of the crash. Their results lead to a natural question for policy makers:
if certain prescriptive measures such as position limits have a low probability of meeting
this policy objectives on a day like May 6, will there be any other more effective counter
measures without unintended consequences?
Nanex Llc. [9] found that quote saturation and NYSE Consolidated Quotation Sys-
tem delays, combined with negative news from Greece and sale of E-mini S&P 500 fu-
tures ”...was the beginning of the freak sell-off which became known as the flash crash”.
They also disproved some of the theories about the cause of the crash: NYSE ”Slow
Quote” mode or LRP’s (Liquidity Replenishment Point); movements in price of Apple
Inc. (AAPL) and stub quotes5 . Nanex Llc. was the first to define and discover mini flash
crashes.
5
Stub orders are limit orders placed well off a stock’s market price, e.g. buy order for $0.01 or sell
order for $100,000.00. A stub quote also serves as a safety net in that if a market maker does notn’t have
enough liquidity available to trade a stock near its recent price range, then a stub quote is entered so
that the market maker complies with its requirements without extending its quotes beyond its available
liquidity.
3
2.2 Definitions and Data Characteristics
Nanex Llc., a data analytics company, was the first to identify and define mini flash
crashes and we present their definition [9]. To qualify as a down crash candidate, the
stock price change has to satisfy the following conditions:
Down crashes are usually referred to as flash crashes, but in order to avoid confusion with
the May 6th Flash Crash, we will use the term down crashes here.
Likewise, to qualify as an up crash candidate, the stock price change has to satisfy
the following conditions:
Up crashes are usually referred to as flash dashes, but we will use the term up crashes here.
4
Year Down Crashes Up Crashes
till November 2010 1041 777
2009 1462 1253
2008 4065 4354
2007 2576 2456
2006 254 208
Data used to analyze the crashes has been downloaded from Wharton Research Data
Services (WRDS) [10]6 .
The data for each stock contains the following information:
The downloaded data for each crash contains 3 minutes of trading information sur-
rounding the crash WHY NOT GO FOR A FIXED PERIOD BEFORE AND AFTER? ;
for instance if the crash occurred at 12:45:11, we downloaded the trade data from 12:44:00
to 12:46:59. As given above, the definition of a flash crash requires that 10 trades, ex-
ceeding 0.8% change are executed within 1.5 seconds. As the obtained data is second
time-stamped, we modify the second condition to require the trades to be executed within
2 seconds, i.e. there has to be at least 10 consecutive price changes, greater than 0.8%,
occurring within 2 seconds.
Nanex Llc. provided the following information about the crashes [9]:
The exact second of the crash, as well as confirmation that the crash occurred has
been verified by our analysis. This has though revealed certain discrepancy discrepancies
in the number of confirmed crashes between our and Nanex’s analysis. We were able to
verify 5001 out of a suggested 9048 up crashes and 4765 out of a suggested 9389 down
crashes. Though in some cases it was obvious that trading characterized as a crash might
have occurred, the condition of a 10 consecutive price changes, in either direction, was
6
We thank the University of Manchester for providing access to WRDS Trade and Quote data
5
not satisfied. For our investigation, the candidates for flash crashes that did not satisfy
the strict definition, were removed from further analysis.
Figure 3 shows the occurrences of flash crashes during the trading day. It can be seen
that most (up or down) crashes occurred during the start and the end of the trading
day. This can easily be related to intra-day volatility; the most volatile periods are at the
beginning and end of trading day [7]. Furthermore, it is interesting to note that there
appears to be no difference between the up crashes and down crashes - both histograms
look the same.
We examined next how a crash relates to the sector of the targeted stock. High fre-
quency trading has been concentrated in a few very ”liquid” stock, with low-priced stock
from financial and IT sectors7 being the favorites in for high frequency traders [11]. Fig-
ure 4 represents the most targeted sectors in the (up and down) crashes; most crashes
originated from the Finance, Insurance and Real Estate sector and Manufacturing sectors,
accounting for 31% and 30% of the crashes, respectively.
Finally at this state, we note that 39% of trades in the crashes occurred on the New
York Stock Exchange, with both NASDAQ and ARCA each accounting for 28% of the
trades and American Stock Exchange, BATS, Boston stock Exchange, Chicago Board
Options Exchange, Chicago Stock Exchange, International Securities Exchange and Na-
tional Stock Exchange accounting for the remaining 5% of trades. This segmentation is
typical and as expected.
7
Citigroup (C), Bank of America (BAC), Fannie Mae (FNM) and Freddie Mac (FRE) represent close
to 20% of U.S. equity volume. Citigroup announced 1 for 10 reserve stock split in the near future, which
will most likely change its status among HFT.
6
Figure 4:
Figure 5 shows the average price of the stock before the up and down crash. As a crash
is defined as involving at least 10 consecutive price changes, in computing the average
price before the crash we consider all of the prices leading up to the crash.
Formally, if {Pik : i ∈ 1, . . . , n} are prices considered for stock k = 1, . . . , m, and
{Pikj : j ≥ 10} are prices related to the crash, then the average price before the crash for
the stock k is computed as
1 X
hP k i = Pikl .
|{l : l < i1 }|
{l : l<i1 }
In layman terms, all of the prices before the first price that sparked the crash are taken
and their corresponding average is computed.
The bulk of stocks that were involved in the crashes were priced at less than $100,
with two notable peaks: the first one around $50 and the second around stocks priced
7
Figure 5: Average Price of the Stock before the Crash
between $10 and $25. Again, there appears to be no significant difference between the
histograms. newpage
We further explored the relationship between the average of the price before the crash
with the intensity of the crash. In Figure 5, the graph are plotted on a log-log scale,
and they are separated based on at which exchange the crash occurs. As each crash has
to have a sequence of at least 10 ticks before ticking in the other direction, we found
all exchanges from which the trades in the crash occurred. The exchange with the most
trades is defined as the dominant exchange and the crash is said to originate from the
dominant exchange. Three exchanges account for the bulk of crashes: New York Stock
Exchange, NASDAQ and ARCA. Figures 5 and 6 present the relationship between the
average price before the crash and the intensity of the crash.
It is noticeable that there is negative correlation between the average price before the
crash and the intensity of the crash in the case when crashes originated from NYSE. This
implies that flash crashes originating from NYSE, in low-priced stocks occur with greater
intensity than on the highly-priced stocks. The crashes occurring on NASDAQ, ARCA
and other exchanges do not exhibit such a relationship to the same extent.
8
INTERESTING BUT SO WHAT? GOOD POINT
We have separated the crashes according to which sector the stocks belong, as well as
to time of the crash8 and found no relationship.
The next two graphs, Figures 7 and 8, show the movement of trading in the crashes
throughout different geographical regions. The New York and greater New York area
account for trading from BATS9 , NYSE, NASDAQ, ISE and AMEX; Chicago area cov-
ers ARCA, CBOE, CSX and NSX; Boston covers the Boston stock exchange. One can
think of the graphs as dynamical systems of trading during a crash, with the empirical
percentages representing transitional probabilities. For instance, if the current trade in
the down crash occurred in Chicago, with 95.182% probability that the next trade in the
down crash will occur in New York.
9
Figure 8: Trading between geographical regions - DOWN crashes
For each type of crash, up and down, we computed the difference between the sample
median and the sample mean of up and down empirical probabilities, i.e. we computed
\
median(X) − mean(X),
[
10
Figure 9: Trading between geographical regions - UP crashes
This finding allows us to conjecture a hypothesis for relating to the cause and oc-
currence of flash crashes. High frequency traders in essence appear to act as market
makers, i.e. they provide liquidity by buying and selling the stocks rapidly, trying to
earn the difference between the bid (best buying price) and the ask (best selling price),
while maintaining a low exposure - usually a couple of thousand shares. When a stock
price has a distinct price movement, market makers receive an overabundance of orders
that work against them, for instance, if there is a distinct down trend in the stock price
movement, market makers will receive sell orders from other market participants - market
makers will then be the buyers. As the price moves further down, market makers will
have larger and larger inventory exposure obtained at less and less favourable prices. As
soon as the market maker’s risk management limits are breached - this limit, is comprised
of the size of the inventory and the (unrealized) profit & loss - , the market maker has to
stop providing liquidity and start to aggressively take liquidity, by selling back the shares
bought moments earlier. This way they push the price further down and thus exaggerate
the downward movement. As waiting for a price to revert to favourable levels is not a
feasible option due to the speed of trading, high frequency traders have to ”dump” their
accumulated inventory on the market as soon as possible -, i.e. they have to immediately
sell shares in their inventory. This action is, in turn, likely to cause a sharp downward
spike in the price movement. HOW OFTEN MIGHT ONE EXPECT THIS TO HAP-
PEN?
11
Figure 10: Empirical Probabilities of Up and Down Price Movements before the DOWN
Crash
12
Figure 11: Cumulative Distributions of the First 10 Price Changes - DOWN crash
and the strict inequality holds for at least one value x∗ ∈ R. As, for random variable Z
with cumulative distribution function FZ , the following holds
Z
E[Z] = (1 − FZ (z))dz,
{z : z=Z(ω), ω∈Ω}
and we immediately obtain that if X dominates Y in the sense of first order, then
Therefore, we also obtain that the expected size of the first price change in the flash crash
is the largest.
Finally, we report the pattern found in the average price and the intensity of the
crashes that occurred at NYSE. We separated the crashes that occurred before and after
the introduction of the Supplemental Liquidity Provider program on NYSE.
Supplemental Liquidity Providers (SLPs) are market participants that who use sophis-
ticated high-speed computers and algorithms to create high volume on exchanges in order
to add liquidity to the markets. As an incentive for providing liquidity, the exchange pays
the SLP a rebate or fee, which was 0.15 cents per share as of 2009. The SLP program
was introduced shortly after the collapse of Lehman Brothers.
At this moment we do not have an understanding why this pattern exists and whether
the disappearance of the pattern after introduction of the SLP program is merely a coin-
cidence.
Corresponding scatter plots for other type of crash are given in the appendix.
13
Figure 12: Average price vs. intensity UP crash - zoomed in
that the bid-ask spread should widen once the crash has occurred - traders should be
unwilling to trade knowing that an uncommon event recently occurred. This assumption
is not backed up by our findings, the average spread does not seem to alter after the crash.
Figure 13 shows in log-log scale the scatter plot of average bid-ask spread before and after
the flash crashes.
We must keep in mind that the average bid-ask spread is computed; it might may be
the case that any deviation can be drowned by the amount of the data when taking the
average. Furthermore, the analysis is done on a sub-sample of data. At this stage we
cannot conclude that there is no impact on bid-ask spread due to the crash.
4 Legal Issues
11
11
This issue was brought up by Dennis Dick of Trading Defenders
12
Trade-through is an execution of an order at a price inferior to NBBO.
14
Figure 13: Average bid-ask spread, before vs. after the flash crash
Consider the following example of a fictitious order book for security XYZ:
INET ARCA
bid/ask price size price size
ask 32.02 500 32.02 500
ask 32.01 500
best ask 32.00 500
best bid 31.99 100 31.99 100
Suppose ARCA receives a market order to buy 1000 shares of XYZ, and the best avail-
able offer is on INET for 500 shares. ARCA must route the first 500 shares to INET, so as
to not violate the order protection rule (because the 500 shares on INET is at the top of
the book and is therefore a ”protected quotation”). After that, the remaining 500 shares
can be filled back on the original exchange that received the order, in this case ARCA,
regardless of whether there is better quotations on other trading centers. In this case,
500 shares would be executed against the 32.00 INET offer, and the remaining 500 shares
could then be executed against the 32.02 ARCA offer. The 500 shares offered at 32.01
on INET would remain unfilled. The buyer in this case actually pays a penny more than
they should of had to for the last 500 shares. That is because the only ”protected quo-
tation” is the top of the book which in this case is the 500 shares offered on INET at 32.00.
Previous example demonstrates the issues that might arise if a large buy or sell order
would be routed to an exchange; the ”domestic” exchange would route a part of the order
to the exchange with the NBBO, so as to not violate the order protection rule, and then
would filled the remaining shares on its own quotes; it could easily result in an dramatic
price change if there is lack of liquidity on ”domestic” exchange.
15
This series of trades exemplifies the lack of protection for depth of book quotations -
there might have been a significant amount of additional liquidity on the other exchanges
during the time frame under question. However, because the order protection rule only
protects the top of the book, trades can occur at inferior prices despite there being liquidity
on other trading centers. The domestic exchange did nothing wrong in our example and
they should not be villainized. The problem lies in the order protection rule, and the SEC
should consider expanding this rule to provide protection to the depth of book quotations.
These trade-throughs occur every day, just usually on a smaller scale as in the first
example. They can sometimes occur on larger scales, especially on exchanges with less
liquidity - the media would report these incidents as ”mini flash crashes” (or mini flash
dashes in the case that the stock spikes higher). These incidents might be nothing more
than a testament to the fragmented liquidity that is indicative of the current market
structure.
As more and more exchanges are created, traders need to be more aware of how their
orders are routed. In the example, the trader could of sent the order using some type of
”smart” order router. This order type would have sought out liquidity on all exchanges as
well as on dark pools and ECN’s. Unfortunately, not all participants of the capital markets
understand how their orders are routed and therefore it may become necessary for the SEC
to expand the order protection rule to include depth of book quotations. Otherwise as
the market structure becomes more and more complex, expect trade-throughs to become
the norm.
YOU NEED TO TIDY=UP THE REFERENCES - NEED FULL REF and PAGE
NUMBERS; NEED web-page and time and data accessed,
5 Contribution
16
References
[1] Mehrdad Samadi Tugkan Tuzun Andrei Kirilenko, Albert S. Kyle. The flash crash:
The impact of high frequency trading on an electronic market. January 12 2011.
[2] Annie Koh Bernard Lee, Shih-fen Cheng. Would position limits have made any
difference to the flash crash on may 6, 2010.
[3] Maureen O’Hara David Easley, Marcos M. Lopez de Prado. Measuring flow toxicity
in a high frequency world. October 21 2010.
[4] Maureen O’Hara David Easley, Marcos M. Lopez de Prado. The microstructure
of the flash crash: Flow toxicity, liquidity crashes and the probability of informed
trading. The Journal of Portfolio Management, 2011.
[5] Michael Durbin. All About High Frequency Trading. McGraw Hill, 2010.
[6] Cristina McEachern Gibbs. Breaking it down: An overview of high frequency trading.
Advanced Trading, 2009.
[7] L. Harris. A transaction data survey of weekly and intraday patterns in stock prices.
Journal of Financial Economics, 1986.
[8] J. Hasbrouck. One security, many markets: Determining the contributions to price
discovery. Journal of Finance, 1995.
[10] The Wharton School of the University of Pennsylvania. Wharton research data
service.
[12] U.S. Commodities Futures Trading Commission U.S. Securities and Exchange Com-
mission. Findings regarding the market events of may 6, 2010. September 2010.
[13] Robert Wood Sugato Chakravarty, James Upson. The flash crash: Trading aggres-
siveness, liquidity supply, and the impact of intermarket sweep orders. January 2011.
6 Appendix
17
Figure 14: Average Price vs. Intensity of the crash - UP crash
18
Figure 16: Empirical Probabilities of Up and Down Price Movements before the UP Crash
19
Figure 18: Average price vs. intensity UP crash - zoomed in
20