Rise of the Machines: Algorithmic Trading in the Foreign...
Transcript of Rise of the Machines: Algorithmic Trading in the Foreign...
1
Rise of the Machines:
Algorithmic Trading in the
Foreign Exchange Market Alain Chaboud, Benjamin Chiquoine, Erik
Hjalmarsson, and Clara Vega
And MQL event study Preliminary results
April 2013
Rise of the Machines’ Main Question
What role do algorithmic traders (AT) play
in the price discovery process? Do ATs
make prices more or less informative?
Triangular arbitrage opportunities
Autocorrelation of high frequency returns
2
Question (continued)
What is the mechanism? Does the impact
depend on whether AT provides liquidity or
demands liquidity? AT measured in five different ways
• AT participation (VAT)
• AT liquidity provision (VCm)
• AT liquidity demand (VCt)
• AT “signed” liquidity demand (OFCt)
• Correlation of AT trading actions (ln(R))
3
Question (continued)
Main problem in answering question:
endogeneity (reverse-causality)
Granger causality and Heteroskedasticity identification
approach (Rigobon (2003), Rigobon and Sack (2003,
2004))
4
Theory
Excellent review of the literature: Biais and Woolley
(2011) and Foucault (2012)
Disagreement on the effect AT may have on price
discovery (Foucault (2012))
Positive effect: Oehmke (2009) and Kondor (2009),
competition among convergence traders makes prices more
informationally efficient
5
Theory (continued)
Positive effect: Biais, Foucault, and Moinas (2011) and
Martinez and Rosu (2011)
Computers are fast and better informed than other traders
Computers use market orders to exploit their informational
advantage
Computers make prices more informationally efficient, but
increase adverse selection costs for slow traders
Positive effect: Hoffman (2013) fast computer liquidity
providers are better informed.
6
Theory (continued)
Negative effect: Jarrow and Protter (2011): computers
reacting to a common signal, create price momentum and
push prices further away from fundamentals
Negative effect: “crowding effect”: Kozhan and Wah
Tham (2012) and Stein (2009) computers entering the
same trade at the same time push prices further away from
fundamentals
7
Theory (continued)
Negative effect: If computers are “noise” traders: Delong
et al. (1990), Froot, Scharfstein, and Stein (1992)
Positive feedback traders who predictably extrapolate past price
trends
Short-term speculators (chartist) herd and put too much emphasis
on some (short-term) information and not enough on fundamentals
AT could cause “excessive” volatility
Foucault (2012) effect may depend on strategy computers
specialize on.
8
Our Data
EBS (essentially the global site of price discovery in
interdealer FX market for several large currencies) records
when a trade is placed manually (keyboard) or by a
computer interface
Minute-by-minute data from 2003 to 2007
Three currency pairs (EUR-USD, USD-JPY, EUR-JPY)
9
Our Data (continued)
Volume and direction of trade breakdown each minute by
AT (Computer) and non-AT (Human).
We know how much computers “take” from human “makers.”
Four possible types of transactions: HH, HC, CH, CC
(maker-taker).
10
Algorithmic Trading Growth in EBS
11
0
10
20
30
40
50
60
70
80
90
Jan-03 Jan-04 Jan-05 Jan-06 Jan-07
Pa
rtic
ipa
tio
n (
Per
cen
t)
USD/EUR JPY/USD JPY/EUR
12
Our Data (continued) Five different measures of AT activity
AT participation:
Vol(CH+HC+CC)/Vol(CH+HC+CC+HH)
AT liquidity supply:
Vol(CH+CC)/Vol(CH+HC+CC+HH)
AT liquidity demand:
Vol(HC+CC)/Vol(CH+HC+CC+HH)
AT signed liquidity demand:
|OF(HC+CC)|/(|OF(HC+CC)|+|OF(CH+HH)|)
Correlation of AT trading actions: R-measure
13
What if algorithmic traders (ATs) all did the
same trade at the same time?
Correlated strategies can make prices more informationally
efficient (“convergence” trades)
Correlated strategies can cause excess volatility
e.g., Yen-Dollar market on August 16, 2007
14
15
Do algorithmic trades tend to be correlated?
We do not know strategies, we do not have orders, only
completed trades.
Instead: Do computers trade with each other as much as
expected, as much as random matching would predict? If
computer strategies are correlated, we should observe less
trading among computers than expected.
More precisely: Do computers “take” from humans and
computers in the same proportion as humans take from
humans and computers?
16
Prob(HC)/Prob(CC) = computer taker ratio = RC
Prob(HH)/ Prob(CH) = human taker ratio = RH
In a world with more human makers than computer makers
(our world), we expect Prob(HC)/Prob(CC) > 1, i.e.,
computers take more from humans than from other
computers. And we expect Prob(HH)/ Prob(CH) > 1, i.e.,
humans take more from humans than from computers.
However we expect RC/RH=1, i.e., humans take more
from humans in a similar proportion that computers take
more from humans.
17
Assuming Prob(HH)= (1 - αm)(1 – αt), RC/RH>1 implies
that either:
Prob(HC) > (1 - αm) αt
or
Prob(CH) > αm(1 – αt)
or
Prob(CC) < αm αt
If we find that RC/RH>1, then we conclude that computers
take more from humans, than humans themselves, in other
words, computer trading is more correlated than expected,
as computers trade less with other computers than expected
or computers trade more with humans than expected.
18
We estimate:
R= RC/RH= 𝑉𝑜𝑙(𝐻𝐶)
𝑉𝑜𝑙(𝐶𝐶)
𝑉𝑜𝑙(𝐻𝐻)𝑉𝑜𝑙(𝐶𝐻)
At 1-minute, 5-minute and daily frequency
Report ln(R)
If we find that ln(R)>0, then we conclude that computer
trading is more correlated than expected
19
20
ln(R ) ln(R ) ln(R )
1-min 5-min Daily
EUR/USD
Mean 0.222*** 0.367*** 0.531***
(std. err.) (0.0031) (0.0036) (0.0118)
Fraction of obs.>0 0.599 0.723 0.99
No. of non-missing observations 143421 52101 880
Total no. of obs. 512160 102432 1067
JPY/USD
Mean 0.310*** 0.392*** 0.5849***
(std. err.) (0.0037) (0.0043) (0.0131)
Fraction of obs.>0 0.611 0.703 0.99
No. of non-missing observations 114449 49906 953
Total no. of obs. 512160 102432 1067
JPY/EUR
Mean 0.698*** 0.687*** 0.813***
(std. err.) (0.0049) (0.0051) (0.0173)
Fraction of obs.>0 0.696 0.758 0.98
No. of non-missing observations 71783 45846 987
Total no. of obs. 512160 102432 1067
Do algorithmic trades tend to be correlated?
Answer: Yes. It seems that computers do not trade with
each other as much as random matching would predict.
21
Relationship between algorithmic trading
activity and triangular arbitrage
opportunities
Graphical evidence
22
23
Percent of seconds with triangular arbitrage profit
greater than 1 basis point, in 3-11 time interval
What is the effect of algorithmic trading on
triangular arbitrage opportunities?
Endogeneity (reverse causality) problem:
Triangular arbitrage must clearly also cause
AT Granger Causality at high frequency (minute-by-
minute)
and
Heteroskedasticity identification
24
Structural VAR Estimation
𝑌𝑡 = (𝐴𝑟𝑏𝑡 , 𝐴𝑇𝑡)
Φ 𝐿 Lag-polynomial, 20-lags
𝑋𝑡−1:𝑡−20 Controls for past volatility and liquidity (volume)
𝐺𝑡 deterministic intra-daily patterns and time trend
25
𝐴𝑌𝑡 = Φ 𝐿 𝑌𝑡 + Λ𝑋𝑡−1:𝑡−20 +Ψ𝐺𝑡 + 𝜀𝑡
Test of AT causing triangular arbitrage opp.
26
Test of AT Causing Triangular Arbitrage VAT VCt VCm OFCt ln(R )
Sum of coeffs. on AT lags -0.0027*** -0.0061*** 0.0049*** -0.0117*** -0.1202***
Chi-squared (Sum=0) 7.7079*** 25.2501*** 14.435*** 142.46*** 15.39***
p-value 0.0055 0 0.0001 0 0.0001
Chi-squared (All coeffs. on AT lags=0) 91.4066*** 149.588*** 46.51*** 317.38*** 40.99***
p-value 0 0 0.0007 0 0.0037
Contemporaneous coeff. X σ(AT) -0.1932*** -0.813*** -0.0406*** -0.4813*** -1.5193**
No. of obs. 512016 512016 512016 511688 176713
Test of triangular arbitrage opportunities
causing AT
27
Test of Triangular Arbitrage causing AT VAT VCt VCm OFCt ln(R )
Sum of coeffs. on AT lags 0.0140*** 0.0219*** -0.0022 0.0345*** 0.0005
Chi-squared (Sum=0) 19.5814*** 56.9040*** 0.82 75.95*** 1.69
p-value 0 0 0.36 0 0.1925
Chi-squared (All coeffs. on AT lags=0) 83.0479*** 118.9288*** 31.00* 459.30*** 84.92***
p-value 0 0 0.055 0 0
Contemporaneous coeff. X σ(AT) 2.0405*** 1.6185*** 0.1832*** 4.3776*** 0.5308***
No. of obs. 512016 512016 512016 511688 176713
Triangular Arbitrage Causality Tests
AT reduces triangular arbitrage opportunities
Predominantly AT acts on posted quotes by other traders
that enable the profit opportunity
Increase the speed of price discovery, but increase adverse
selection costs of slow traders
Some evidence that algorithmic traders make prices more
efficient by posting quotes that reflect new information
quickly
28
Does algorithmic trading increase or
decrease “excess” volatility: autocorrelation
of high frequency returns?
Graphical evidence
29
30
5-second return autocorrelation
31
5-second return autocorrelation
32
5-second return autocorrelation
Test of AT participation causing HF return
autocorrelation
33
VAT
Test of AT Causing HF return Autocorrelation USD/EUR JPY/USD JPY/EUR
Sum of coeffs. on AT lags -0.039*** -0.025*** -0.01539***
Chi-squared (Sum=0) 33.75*** 25.26*** 14.96***
p-value 0 0 0.0001
Chi-squared (All coeffs. on AT lags=0) 43.05*** 45.54*** 15.1881***
p-value 0 0 0.0043
Contemporaneous coeff. X σ(AT) -0.3438*** -0.244*** -0.5011***
No. of obs. 102432 102427 102113
Test of AT liquidity demand causing HF return
autocorrelation
34
VCt
Test of AT Causing HF return Autocorrelation USD/EUR JPY/USD JPY/EUR
Sum of coeffs. on AT lags -0.043*** -0.021*** -0.01069**
Chi-squared (Sum=0) 23.01*** 11.88*** 6.12*
p-value 0 0.0006 0.0134
Chi-squared (All coeffs. on AT lags=0) 24.79*** 20.87*** 6.65
p-value 0.0001 0.0003 0.155
Contemporaneous coeff. X σ(AT) -0.1568*** -0.050*** -0.0369
No. of obs. 102432 102427 102113
Test of AT liquidity supply causing HF return
autocorrelation
35
VCm
Test of AT Causing HF return Autocorrelation USD/EUR JPY/USD JPY/EUR
Sum of coeffs. on AT lags -0.046*** -0.0362*** -0.02146***
Chi-squared (Sum=0) 29.72*** 28.84*** 17.88***
p-value 0 0 0
Chi-squared (All coeffs. on AT lags=0) 44.16*** 38.86*** 18.90***
p-value 0 0 0.0008
Contemporaneous coeff. X σ(AT) -0.3844*** -0.311*** -0.295***
No. of obs. 102432 102427 102113
Test of AT correlated actions causing HF
return autocorrelation
36
OFCt
Test of AT Causing HF return Autocorrelation USD/EUR JPY/USD JPY/EUR
Sum of coeffs. on AT lags -0.00481 -0.0015 0.00013
Chi-squared (Sum=0) 2.3 0.2637 0.0021
p-value 0.12 0.607 0.963
Chi-squared (All coeffs. on AT lags=0) 2.82 1.85 1.38
p-value 0.58 0.76 0.84
Contemporaneous coeff. X σ(AT) 0.098 -0.144 -0.032
No. of obs. 102250 1019992 100815
Test of HF return autocorrelation causing AT
37
VAT
Test of AT Causing HF return Autocorrelation USD/EUR JPY/USD JPY/EUR
Sum of coeffs. on AT lags -0.0068* -0.010* -0.0208***
Chi-squared (Sum=0) 3.0155* 3.49** 7.8843***
p-value 0.0825 0.0616 0.005
Chi-squared (All coeffs. on AT lags=0) 12.87** 4.026 12.21**
p-value 0.0119 0.4 0.0158
Contemporaneous coeff. X σ(Autocorrelation) 0.036* 0.011 0.608***
No. of obs. 102432 102427 102113
38
Conclusion We find evidence of algorithmic trading improving price efficiency:
Reduces triangular arbitrage opportunities: mainly by acting on the posted quotes of other traders that enable the profit opportunity
Reduces HF return autocorrelation: mainly by providing liquidity
39
Event Study:
Minimum Quote Life on EBS
Work in progress
EBS Imposed a MQL of 250 milliseconds on
June 15, 2009
Is it a binding constraint? Why did EBS impose it?
Theory
What happened to AT volume?
What happened to bid-ask spreads? Depth? Auto-correlation
of 5-second returns? Triangular-arbitrage opportunities?
Can we find a counterfactual to compare it to?
Are the results applicable to equity markets?
BIG CAVEAT, this is a market where EBS has certain monopoly
power: investors “do not have much of a choice” of trading venues.
40
41
It is a binding constraint
I II III IV I II III IV I II III IV I II III IV I II III IV I II III IV
2007 2008 2009 2010 2011 2012
Date
Percent of QI< 250ms out of all QI
42
EBS wants to promote “genuine interest in trading”
I II III IV I II III IV I II III IV I II III IV I II III IV I II III IV
2007 2008 2009 2010 2011 2012
Date
Fill Ratio
43
Theory regarding the effects of MQL
Imposing a MQL will make it more costly for computers to post a quote:
Increase bid-ask spreads
Decrease depth (because there is less willingness to post quotes)
Increase high-frequency return autocorrelation
Imposing a MQL may lower adverse selection costs for some humans (slow traders)
Ambiguous effect on triangular arbitrage opportunities
Increased bid-ask spreads makes triangular arbitrage opportunities less likely to happen
Decreased trading activity by computers increases triangular arbitrage opportunities (as we showed previously)
44
Caveats
Caveat 1: If trading venue has a “monopoly” then the effect on the willingness to post quotes may be very small
Caveat 2: Quotes interrupted within 250ms were not intended to lead to transactions anyway, so getting rid off these quotes may have a minimal effect on: “transactable” bid-ask spreads, “transactable” depth.
45
What is the counterfactual?
Reuters (the other large FX interbank dealer trading venue) imposed MQL in 2005
EBS currencies where MQL is not a binding constraint, e.g., EUR-JPY currency pair
46
MQL may affect EUR/JPY the least
47
Diff -in-diff
methodology:
trend needs to
be similar
prior to
“event”
I II III IV I II III IV
2006 2007
Date
Volume CH
EUR/USD
USD/JPY
EUR/JPY
I II III IV I II III IV
2006 2007
Date
Volume HC
EUR/USD
USD/JPY
EUR/JPY
48
Diff -in-diff
methodology:
during event 8 22 5 19 3 17 31 14 28 12 26 9 23 6 20
M3 M4 M5 M6 M7 M8 M9
Date
Volume CH
EUR/USD
USD/JPY
EUR/JPY
8 22 5 19 3 17 31 14 28 12 26 9 23 6 20
M3 M4 M5 M6 M7 M8 M9
Volu
me
Date
Volume HC
EUR/USD
USD/JPY
EUR/JPY
49
Diff -in-diff results CC HC CH Total
EUR/USD
diff -14.6% -16.0% -16.0% -16.7%
p-value 0.280 0.177 0.232 0.142
diff in diff -21.8% -32.5% -35.5% -32.2%
p-value 0.150 0.055 0.083 0.051
UDS/JPY
diff -20.4% -17.4% -18.0% -14.5%
p-value 0.173 0.228 0.248 0.295
diff in diff -28% -34% -38% -30%
p-value 0.059 0.028 0.035 0.043
EUR/JPY
diff 7.1% 16.4% 19.5% 15.5%
p-value 0.382 0.232 0.248 0.252
50
Diff -in-diff results
Bid-Ask Spread Depth Autocorrelation Triangular Arbitrage
EUR/USD
diff 3% -4% 8% 24%
p-value 0.13 0.33 0.09 0.15
diff in diff 5% -5% 0%
p-value 0.17 0.24 0.48
UDS/JPY
diff -2% -9% -3%
p-value 0.26 0.11 0.26
diff in diff 0% -10% -11%
p-value 0.46 0.11 0.09
EUR/JPY
diff -1% 1% 8%
p-value 0.35 0.35 0.13
51
Bid-Ask Spread
.010
.015
.020
.025
.030
.035
.040
.0001350
.0001400
.0001450
.0001500
.0001550
.0001600
8 22 5 19 3 17 31 14 28 12 26 9 23 6 20
M3 M4 M5 M6 M7 M8 M9
Sp
rea
d
Date
Bid Ask Spread
EUR/USD
USD/JPY
EUR/JPY
52
Depth
2
3
4
5
6
7
8 22 5 19 3 17 31 14 28 12 26 9 23 6 20
M3 M4 M5 M6 M7 M8 M9
Depth
0
Date
Depth0
EUR/USD
USD/JPY
EUR/JPY
53
Auto-correlation
.09
.10
.11
.12
.13
.14
.15
8 22 5 19 3 17 31 14 28 12 26 9 23 6 20
M3 M4 M5 M6 M7 M8 M9
Auto
corr
ela
tion
Date
Autocorrelation
EUR/USD
USD/JPY
EUR/JPY
54
Triangular Arbitrage Opportunities
.1
.2
.3
.4
.5
.6
.7
8 22 5 19 3 17 31 14 28 12 26 9 23 6 20
M3 M4 M5 M6 M7 M8 M9
Pro
port
ion
Date
Triangular Arbitrage Opportunities
55
Need to find a better counterfactual
Preliminary evidence indicates
Trading volume decreased around the imposition of MQL
There was no statistically significant change on other variables: bid-ask spreads, depth, HF return autocorrelation, triangular arb opportunities
Caveat: the conclusions cannot be generalized to a fragmented equity market. Unless, perhaps MQL is imposed simultaneously on all trading venues.
Need more research in this area. Analyze other events, other markets.
Conclusion
56
MQL Effect on Quote Posting Behavior
2006 2007 2008 2009 2010 2011 2012
Date
Quotes Submitted
57
MQL Effect on Volume-Transactions
2006 2007 2008 2009 2010 2011 2012
Date
Transactions
58
Backup Slides
59
6
PM
12
AM
6
AM
12
PM
6
PM
12
AM
6
AM
12
PM
-2500
-2000
-1500
-1000
-500
0
500
1000
1500
111
112
113
114
115
116
117$ Millions Yen/$
Human-Maker / Computer-Taker Order Flow
Order Flow
Dollar-Yen
6
PM
12
AM
6
AM
12
PM
6
PM
12
AM
6
AM
12
PM
-2500
-2000
-1500
-1000
-500
0
500
1000
1500
111
112
113
114
115
116
117$ Millions Yen/$Computer-Maker / Computer-Taker Order Flow
Order Flow
Dollar-Yen
60
6
PM
12
AM
6
AM
12
PM
6
PM
12
AM
6
AM
12
PM
-2500
-2000
-1500
-1000
-500
0
500
1000
1500
111
112
113
114
115
116
117$ Millions Yen/$
Computer-Maker / Human-Taker Order Flow
Order Flow
Dollar-Yen
6
PM
12
AM
6
AM
12
PM
6
PM
12
AM
6
AM
12
PM
-2500
-2000
-1500
-1000
-500
0
500
1000
1500
111
112
113
114
115
116
117$ Millions Yen/$Human-Maker / Human-Taker Order Flow
Order Flow
Dollar-Yen
Theory (continued) Foucault, Kadan, and Kandel (2009) model AT as
lowering monitoring costs
Pareto optimal (lower trading costs, increase trading
rate)
Ambiguous effect on bid-ask spread (liquidity)
• When monitoring costs for market-makers
liquidity informational efficiency
• When monitoring costs for market-takers liquidity
informational efficiency
61
We have four events: H-make, C-make, H-take, C-take.
The probability of each event at time k is
Prob(C-take) = αt , Prob(H-take) = 1 – αt
Prob(C-make) = αm, Prob(H-make) = 1 - αm
Assuming each event is independent, the probabilities of
each trading event are:
Prob(HH) = (1 - αm)(1 – αt)
Prob(HC) = (1 - αm) αt
Prob(CH) = αm(1 – αt)
Prob(CC) = αm αt
We can write the following identities:
Prob(CH)×Prob(HC) ≡ Prob(CC)×Prob(HH)
Prob(HC)/Prob(CC) ≡ Prob(HH)/ Prob(CH) 62
Heteroskedasticity identification
𝑦 = 𝛽1𝑥 + 𝜀 𝑥 = 𝛽2𝑦 + 𝜂
4 parameters: 𝛽1, 𝛽2, 𝜎𝜀 , 𝜎𝜂
3 moments: Var y , Var x , 𝐶𝑜𝑣(𝑥, 𝑦)
System is not identified
63
Heteroskedasticity identification
Two regimes, keep coefficients constant across regimes
𝑦1 = 𝛽1𝑥1 + 𝜀1 𝑦2 = 𝛽1𝑥2 + 𝜀2 𝑥1 = 𝛽2𝑦1 + 𝜂1 𝑥2 = 𝛽2𝑦2 + 𝜂2
6 parameters: 𝛽1, 𝛽2, 𝜎𝜀1, 𝜎𝜂1, 𝜎𝜀2, 𝜎𝜂2
6 moments:
Var 𝑦1 , Var 𝑦2 , Var 𝑥1 , Var 𝑥2 , 𝐶𝑜𝑣 𝑥1, 𝑦1 , 𝐶𝑜𝑣(𝑥2, 𝑦2)
System is exactly identified
64