Big data in Finance
Finance Research Group, IGIDR
July 25, 2014
Introduction
I Who we are?
I A research group working in:
I Securities markets
I Corporate governance
I Household finance
I We try to answer policy questions using data and quantitative
methods.
Case study
Paper: The causal impact of algorithmic trading on market quality
Authors: Nidhi Aggarwal & Susan Thomas
URL:
ifrogs.org/releases/ThomasAggarwal2014 algorithmicTradingImpact.html
The question
I Since 2000, escalating use of technology in trading on equities
markets.
I AT now dominates exchanges worldwide. Concerns about
liquidity, ‘flash crashes’, etc.
I Regulators all over the world are contemplating interventions
on AT.
I In search of finding a market failure that justifies regulatory
intervention, numerous researchers have asked: What is the
effect of AT on liquidity and volatility?
Data
I Periods:
I Pre co-lo: Jan ’09 to Dec ’09 (260 days)
I Post co-lo: Jul ’12 to Aug ’13 (291 days)
I Criterion for securities selection: Study securities with at least
50 average daily trades in 2009 and 2012-13.
This yields a set of 552 securities.
I Frequency used: Tick by tick trades and orders data.
I Data size analysed: 3.8 Terabytes of .csv text files.
Existing literature
1. A lot of the literature uses data from U.S. markets, which
have highly fragmented liquidity.
If AT adoption was taking place in different ways in different
places, it becomes difficult to pin-point the starting point to
measure the impact on the overall market.
Existing literature
1. A lot of the literature uses data from U.S. markets, which
have highly fragmented liquidity.
If AT adoption was taking place in different ways in different
places, it becomes difficult to pin-point the starting point to
measure the impact on the overall market.
2. Datasets often do not offer clear identification of AT. Without
this, the measurement of AT activity is relatively weak.
Existing literature
1. A lot of the literature uses data from U.S. markets, which
have highly fragmented liquidity.
If AT adoption was taking place in different ways in different
places, it becomes difficult to pin-point the starting point to
measure the impact on the overall market.
2. Datasets often do not offer clear identification of AT. Without
this, the measurement of AT activity is relatively weak.
3. Two issues that are worrisome:
I Endogneity: If liquidity is a reason for ATs to choose to focus
trading on a stock, and liquidity is an outcome to be
measured, then which way does the causality flow?
I Threats to validity: Was the change in market quality because
of AT or other factors, such as macro-economics?
This paper
1. A clean microstructure: An exchange with 80% market share
of all trading, one of the largest exchange in the world by
transaction intensity.
2. Uses an exogenous event: Introduction of co-location services
in Jan 2010, which directly affected AT.
3. Data: Every order explicitly tagged as “AT” or “non-AT” for
every security at the exchange.
Consolidated trading
A big exchange by world standards
I In 2012 and 2013, NSE was the world’s #1 exchange by
number of trades on the equity market.
I The dollar value of these trades is small by world standards,
but on this question, that is not important.
A natural experiment
I NSE launched co-location (co-lo) in January 2010.
I There was an S-shaped curve of adoption thereafter.
I This was an exogenous shock to AT intensity.
AT intensity between 2009-13
Start of
70
60 co−lo
AT Intensity (%)
50
40
30
20
10
2009 2010 2011 2012 2013
Issues in establishing causality
AT adoption at the firm level
I Trading in some firms tend to become more AT while trading
in some firms do not.
I Highly liquid firms tend to be more AT.
I There is the danger of selection bias here.
Variation in adoption of AT
Q1 2009 Q3 2013
100 100
75 75
AT Intensity (%)
AT Intensity (%)
50 50
25 25
0 0
0 500 1000 1500 2000 0 1000 2000 3000
MCap (Rs ‘000s) MCap (Rs ‘000s)
Threats to validity: changes in macroeconomic conditions
Start of
co−lo
50
Nifty VIX (%)
40
30
20
2009 2010 2011 2012 2013
Start of
co−lo
0.14
0.12
Nifty IC (%)
0.10
0.08
0.06
2009 2010 2011 2012 2013
I. Research design we use
Matching at the security level
I We identify firms that got low AT adoption and firms that got
high AT adoption.
I Use propensity score matching (PSM) to identify a matched
sample.
I These are firms that are a lot like each other – but there was
an almost experimental allocation where one group got the
treatment of a surge in AT but the other group did not.
Obtaining set of matched firms
I After the launch of co-lo:
I Define
I ‘Treated’: securities with ∆ AT > 70th percentile value –
16.50% (276 firms)
I ‘Control’: securities with ∆ AT < 30th percentile value –
5.39% (276 firms)
I Leave out firms in the middle.
I Propensity score matching:
I Covariates: average daily values of market cap, price, floating
security, turnover, number of trades (for the year 2009)
Density of the propensity score, before and after matching
Before matching
Treated
3.0
Control
2.0
Density
1.0
0.0
−0.2 0.0 0.2 0.4 0.6 0.8 1.0
Propensity score
After matching
1.2
Density
0.8
0.4
0.0
0.0 0.5 1.0
Propensity score
Matching on macroeconomic conditions
I We capture changes in macroeconomic conditions by changes
in the volatility of the market index (Nifty).
I We then match dates in the period before and after co-lo on
volatility.
I This yields a set of dates in both periods which are alike in
macroeconomic conditions.
Matching dates on macro-economic conditions
I Pick dates in the post co-lo period when market volatility
matched the levels in the pre co-lo period (using Mahalnobis
distance).
I This gives a set of 59 dates in each period that are alike.
Final sample characteristics
I Starting sample: Observations on 552 securities; Period of 260
days before co-lo and 291 days after co-lo.
I After matching on security level co-variates: 91 securities with
high AT and 73 securities with low AT.
I After matching on macro-economic conditions: 59 days before
co-lo and after co-lo.
II. Empirical setting
Market quality measures
I Liquidity
1. Transactions costs
1.1 qspread (in %): (best ask - best sell)× 100 / mid-quote
price.
1.2 Impact cost (ic, %): execution cost of a market order at a
size of Rs 25,000 relative to the mid-quote price.
2. Depth
2.1 top1depth (in Rs.): Rupee depth available at the best bid
and ask prices.
2.2 top5depth (in Rs.): Cumulated Rupee depth available at
top five best bid and ask prices.
2.3 depth (# of shares): Average of the outstanding buy side
and sell side number of shares.
2.4 |oib| (in %): Difference in buy and sell side depth as a
percentage of the total depth, on average.
Market quality measures (contd..)
I Volatility
1. Price risk, rvol: Standard deviation of five-minutes returns.
2. Price risk, range: Difference in highest and lowest mid-quote
price in a five-minutes interval.
3. Liquidity risk, lrisk: Standard deviation of ic in five-minutes
intervals.
I Efficiency
1. vr: Ratio of 10-min variance of returns to 5-min returns
2. kurtosis: Value of kurtosis in a five minute interval (absolute
value).
What we find
Estimation using a Difference-in-Difference regression with
matched securities and matched dates.
mkt-qualityi,t = α + β1 at-dummyi + β2 co-lo-dummyt +
β3 (at-dummyi × co-lo-dummyt ) + i,t
β3 Expected
sign
qspread -0.35+ −
ic -0.80+ −
|oib| -14.34+ −
depth -0.08 +
top1depth 0.09 +
top5depth 0.25∗ +
|vr-1| -0.03+ −
kurtosis 6.81+ −
rvol -2.88+ −
range -19.86+ −
lrisk -0.02+ −
What we find, contd.
I Kurtosis is the incidence of extreme returns.
Does higher kurtosis mean more flash crashes?
I We analyse how frequently:
1. Traded prices move by 2%, 5% or 10%
2. In a period of 5 minutes
before co-lo and after co-lo.
I What we find:
in %
Pre co-lo Post co-lo
High-AT Not High-AT Not
two-excess 33.35 33.46 29.36 36.84
five-excess 5.21 5.65 5.30 7.85
ten-excess 1.01 0.91 1.42 1.29
Some more facts
Are ATs consumer or providers of liquidity?
I A well-accepted hypothesis is that ATs trade at the cost of
non ATs. They are assumed to take away liquidity, and do not
supply it.
Are ATs consumer or providers of liquidity?
I A well-accepted hypothesis is that ATs trade at the cost of
non ATs. They are assumed to take away liquidity, and do not
supply it.
I We investigate this hypothesis. We define:
I AT liquidity demand: % of trades that were initiated by ATs
irrespective of who provided the liquidity.
I We calculate out of total trades:
I AT2AT:% of AT trades where ATs were liquidity suppliers.
I nAT2AT:% of AT trades where non-AT supplied liquidity.
I Separate and similar calculations for Non ATs.
I Done for the period between Jan 2013 to Dec 2013.
The facts
Overall cash market:
in %
Mean Median SD Min Max
at-demand 37.45 37.75 3.86 11.22 48.28
at-supply 39.92 40.11 5.17 8.76 53.24
at2at 18.60 18.80 3.47 1.42 29.24
at2nat 21.33 21.34 2.03 7.34 25.91
nat2at 18.85 18.91 1.19 9.80 21.75
nat2nat 41.23 40.92 5.41 29.60 81.44
Nifty stocks
in %
Mean Median SD Min Max
at-demand 47.12 47.30 4.21 17.12 58.41
at-supply 56.12 56.36 5.34 18.31 67.98
at2at 28.95 29.13 4.36 3.64 40.89
at2nat 27.17 27.21 2.08 14.67 32.86
nat2at 18.17 18.22 1.38 13.48 22.56
nat2nat 25.71 25.27 4.92 15.96 68.20
Case study: Conclusion
I The world has shifted from manual to computer-supported
trading in an extremely short time.
I A major new phenomenon that requires analysis.
I All the regulators of the world are interested.
I Rapidly growing literature.
I Four identified flaws: (a) Fragmented microstructure (b) No
clear identification in data infrastructure (c) Lack of
exogenous change in AT and (d) Problems of causal
identification.
I Our research design addresses these four problems.
I Main result: AT is good for market quality, but a) no
significant impact on the depth though, b) no evidence in
support of increase in flash crashes.
Thank you
[email protected]
http://www.ifrogs.org/
