Professional Documents
Culture Documents
Volume 70 Number 3
2014 CFA Institute
PERSPECTIVES
High-frequency trading is not just lowfrequency trading on steroids; markets actually behave very differently now. HFTs have
been characterized as cheetah traders by Bart
Chilton of the US Commodity Futures Trading
Commission (CFTC). I do not agree with that
assessment. As in all areas of the markets, some
HFTs (or, more precisely, the computer programmers behind some HFTs) are behaving and some
are misbehaving. Actual high-frequency trading
is done by computers or silicon traders, but
low-frequency traders (LFTs) are still humans.
Today, a battle is being waged between the silicon
traders and the humans.
HFTs use super-high-speed computers and
co-located servers. To understand what they do,
consider all the stocks or all the futures contracts
out there and think about calculating a giant
variancecovariance matrix of how each security
moves. When a stock goes up, an HFT instantly
computes a probability that another stock will go
up on the basis of predicted correlations and then
trades in that stock. A lot of what HFTs do, such
as predicting relationships between assets, is based
on strategy. Because they are predictive, they also
act on information revealed by LFTs (LFTs are the
rest of us). HFTs engage in sequential games, and
sometimes they behave like predators and take
advantage of LFTs. They can do so because they
are working under a different paradigm.
0.20
0.15
0.10
0.05
0
5
Standardized Return
Time Clock
Volume Clock
Sources: Easley, Lpez de Prado, and OHara (2012b); based on Bloomberg data
from 1 January 2008 to 22 October 2010.
May/June 2014
www.cfapubs.org
19
distribution behaves more like the normal distribution, which matters because it is easier to predict
something that has a better-behaved distribution.
The machines predict on the basis of the betterbehaved distribution because they are thinking
with a volume clock, whereas humans are thinking
with a time clock. The new paradigm is thus event
based, where volume takes the place of time.
11:00 a.m.
12:00 p.m.
1:00 p.m.
2:00 p.m.
3:00 p.m.
20 www.cfapubs.org
Big Data
One reason why HFTs are so successful is that they
base decision making on big data. For example, HFTs look at so-called Level III quotes. The
rest of us are simply looking at the trade price or
quoted spread, but HFTs are looking at the order
book depth. Again, HFTs are strategic, and they
want to figure out where their orders are going
to be. They know the exact way that exchangematching engines work, so they optimize against
the exchange-matching engine where they trade.
They estimate the orders position in the queue
and other players liquidity needs and consider
asymmetric information.
HFTs also use natural language programing
(NLP), which allows for trading on the basis of
new fundamental information by reading millions of webpages at once. NLP derives an aggregate statistic, like an index, based on all that reading and translates it into an immediate buy or sell.
RavenPack and Reuters are big players in developing these market sentiment indices. Remember
Watson, IBMs supercomputer that played on the
television game show Jeopardy? Watson now has
a job working for the hedge funds, reading webpages and deciding how to trade.
Figure 3 shows the US stock market reaction
to the Tweet Crash of 23 April 2013, in which
a hacker posted a false tweet on the Associated
Press (AP) Twitter account stating that an explosion had occurred at the White House, injuring
14,700
14,650
14,600
10:00 a.m.
11:00 a.m.
12:00 p.m.
1:00 p.m.
2:00 p.m.
3:00 p.m.
May/June 2014
www.cfapubs.org
21
10
100
1,000
Trade Size
informed trade. PIN models provide a way to estimate this probability, but this empirical technique
is difficult to implement given the huge trade volumes in todays high-frequency markets.
Our new trading tool is VPIN (volume synchronized probability of informed trading), which
captures the same notion that when there is new
information, people who have the information can
benefit only if they trade on it. If a trader believes
that a stock is undervalued, she will buy it. If other
people share that nonpublic information, then
more buys than sells will occur. This imbalance
moves prices. VPIN is a way to measure this imbalance, and it provides a measure of the markets
toxicity. VPIN is estimated using a volume clock,
and it works by capturing the imbalance between
buy-initiated volume and sell-initiated volume
over volume increments.
Figure 5 shows the e-mini S&P 500 futures
during the flash crash of 6 May 2010. Our VPIN
calculation itself does not seem to be doing anything particularly spectacular, but the cumulative
distribution function (CDF) of the VPIN explains
what is to come (a CDF essentially tells you the
probability that a random variable is less than or
equal to a given number X). As the VPIN measure
goes up, its cumulative distribution function enters
unusual territory. In fact, by 11:56 a.m., the realized
value of the VPIN was in the 10% tail of its distribution (meaning that 90% of the time, VPINs are
below this level); that is, this market was very toxic.
And the VPIN keeps going up: By 1:08 p.m., it is at
the 1% tail of the distribution (or 99% based on the
CDF)almost an hour and a half before the market
crashes at 2:32 p.m.
2014 CFA Institute
1.0
0.9
59,000
0.8
58,000
0.7
Market Value
0.6
0.5
57,000
CDF(VPIN)
56,000
0.4
0.3
0.2
55,000
VPIN
54,000
0.1
0
2:31 a.m.
53,000
9:55 a.m.
11:18 a.m.
1:12 p.m.
2:24 p.m.
2:48 p.m.
3:19 p.m.
4:02 p.m.
23
1.0
42
0.9
40
CDF(VPIN)
0.8
VPIN
0.7
38
36
0.6
34
0.5
32
CDF(VIX)
0.4
30
0.3
28
0.2
VIX
0.1
26
0
24
12:00 a.m. 2:52 a.m. 5:45 a.m. 8:38 a.m. 11:31 a.m. 2:24 p.m. 5:16 p.m. 8:09 p.m. 11:02 p.m.
10,000
8,000
6,000
4,000
Probabilistic Loss
2,000
Timing Component
Liquidity Component
0
0
10,000
20,000
30,000
40,000
50,000
Volume Horizon
24 www.cfapubs.org
Conclusion
We need to recognize that a new paradigm exists.
Markets are not just faster; they are different. LFTs
have to change; they do not have to be faster, but
they do have to be smarter. HFTs cannot survive
by trading only with other HFTs. Doing so has
driven some HFTs out of business.
Portfolio managers and traders need to think
about the high-frequency trading world expanding beyond trading to information. The same technology that can send orders so swiftly can also get
fundamental information more quickly than you
10,000
8,000
6,000
Probabilistic Loss
4,000
Timing Component
2,000
Liquidity Component
0
0
10,000
20,000
30,000
40,000
50,000
Volume Horizon
Note: This picture is drawn for particular parameters for the model developed in Easley, Lpez
de Prado, and OHara (forthcoming).
Source: Easley et al. (forthcoming).
May/June 2014
www.cfapubs.org
25
Notes
1. See, e.g., Easley, Lpez de Prado, and OHara (2011, 2012a,
2012b, 2013, forthcoming).
2. CFTC and SEC, Findings Regarding the Market Events of
May 6, 2010 (30 September 2010): www.sec.gov/news/
studies/2010/marketevents-report.pdf.
References
Almgren, Robert, and Neil Chriss. 2000/2001. Optimal
Execution of Portfolio Transactions. Journal of Risk, vol. 3, no.
2 (Winter):539.
26 www.cfapubs.org
Optimal
Execution
[ADVERTISEMENT]
Horizon.
May/June 2014
www.cfapubs.org
27