Do Proprietary Algorithmic Traders Withdraw Liquidity during Market Stress?

• Published date: 01 June 2019
• Author: Samarpan Nawn,Ashok Banerjee
• Date: 01 June 2019
• DOI: http://doi.org/10.1111/fima.12238

Do Proprietary Algorithmic Traders

Withdraw Liquidity during Market

Stress?

Samarpan Nawn and Ashok Banerjee∗

We investigate the role of proprietary algorithmic traders in facilitating liquidity in a limit order

market. Using order-leveldata from the National Stock Exchange of India, we find that proprietary

algorithmic traders increase limit order supply following periods of both high short-term stock-

specific volatility and extreme stock price movement. Even following periods of high marketwide

volatility, they do not decrease their supply of liquidity. We define orders from high-frequency

traders as a subclass of orders from proprietary algorithmic traders that are revised in less than

three milliseconds. The behavior of high-frequencytrading mimics the behavior of its parent class.

This is inconsistent with the theory that fast traders leave the market whenstress situations arise,

although their limit-order-supplying behavior becomes weaker when the increase in short-term

volatility is more informational than transitory. Agency algorithmic traders and nonalgorithmic

traders behave opposite to proprietary algorithmic traders by reducing the supply of liquidity

during stress situations. The presence of faster traders in the market possibly instills the fear of

adverse selection in them. We document that the order imbalance of agency algorithmic traders

is positively related to future short-term returns, whereas the order imbalance of proprietary

algorithmic traders is negativelyrelated to future short-term returns.

The debate surrounding the usefulness of machine traders has not reached a final verdict and

therefore warrants greater research (O’Hara, 2015). One of the few stylized facts established about

algorithmic trading (AT) and its subset high-frequency trading (HFT) is that they generate a lot

of limit orders in the market and have a high order-to-trade ratio, thereby playing an instrumental

role in the supply of liquidity in limit order markets. What is in dispute, however, is the quality and

robustness of the liquidity supplied. Although many argue that machine traders have improved

the overall liquidity of the market (Hendershott, Jones, and Menkveld, 2011; Jovanovic and

Menkveld, 2011; Boehmer,Fong, and Wu, 2014), others point out that the liquidity is “phantom”

(Shorter and Miller, 2014) and nonexistent during market crashes (Kirilenko, Kyle, Samadi, and

Tuzun, 2017).

The literature suggests that HFT has taken over the mantle of market making (Hagstr¨

omer

and Norden, 2013; Menkveld, 2013), but the focus has been on general liquidity rather than on

episodic liquidity. Although a traditional market maker increases the bid-ask spread when he

sees market stress, he remains in the market with his supply of liquidity. A natural question thus

arises whether a similar behavior is observed for machine traders, which endogenously supply

Wethank Bing Han (Editor) and an anonymous refereefor their useful comments, which have made the paper qualitatively

superior.An earlier version of the paper was presented under a differenttitle, “Market Liquidity: A Study from Proprietary

Algorithmic Traders’ Perspective” at the 28th Asian Finance Association Annual Meeting and the 5th India Finance

Conference. We thank the conference participants and discussants for their valuablecomments.

∗Samarpan Nawn is an Assistant Professor in the Department of Finance and Accounting at the Indian Institute of

Management Udaipur in Rajasthan, India. Ashok Banerjee is a Professor in the Department of Financeand Control at

the Indian Institute of Management Calcutta in WestBengal, India.

Financial Management •Summer 2019 •pages 641 – 676

642 Financial Management rSummer 2019

liquidity to the market. To comprehend the reliability and robustness of the liquidity provided by

computer-generated algorithms, it is critical to observe its supply at times of stress. Academic

research on this feature has been limited and primarily concentrated on a single event, the 2010

US flash crash (Madhavan, 2012; Kirilenko et al., 2017).

How algorithms contribute to liquidity can depend on their type. Trading algorithms can be

used in either an agency or a proprietary context (Hasbrouck and Saar, 2013). When institutional

investors use AT to minimize the cost of trading sizable amounts of shares through brokers or

“agents,” it is known as agency algorithmic trading (AAT). AAT mainly corresponds to using

algorithms to break up the required order into smaller pieces to achieve an average price better

than some benchmark (such as volume-weighted average price). Proprietary algorithmic trading

(PAT) is known for its subclass HFT, which reacts extremely rapidly to market events.1Relying

on speed, HFT applies algorithms to use very short-lived profit opportunities generated in the

trading environment (Hagstr¨

omer and Norden, 2013; Biais and Foucault, 2014). Although our

focus in this article is on PAT because recent research and arguments about algorithmic trading

have centered on HFT, we also document the liquidity-supplying behavior of AAT.

Most of the empirical research on AT/HFT has been using data from developed markets.

Biais and Foucault (2014) note that one of the major problems with this research is explicit

identification. Most of the research is done through either some proxy for machine trading,

(e.g., message traffic as in Hendershott et al., 2011; RunsInProcess using the number of linked

messages per 10-minute interval as in Hasbrouck and Saar, 2013) or some specific but not

exhaustive data set (NASDAQ data as in Brogaard, Hendershott, and Riordan, 2014).2Another

concern expressed by Biais and Foucault (2014) is that most existing studies rely on a single

market or a single asset and that the lack of cross-market data can affect results because high-

frequency traders are likely to take positions in multiple markets at the same time. The data we

employ addresses both of these concerns. We conduct our empirical study using data from the

National Stock Exchange (NSE) of India; NSE itself has flagged every order and every trade

based on whether it is generated from an algorithmic terminal or not and whether it is for a

client account or a proprietary account. Also, the stock market in India is largely unfragmented.

In the Indian cash equities segment, the NSE records nearly 75% of the traded volume3.The

only other major exchange in the same segment is the Bombay Stock Exchange, which is a long

way behind the NSE regarding turnover. The situation provides an ideal setting to trace AT/PAT

behavior, as this trading activity would be primarily concentrated in the market we examine.

Although the exchange flags help us to exactly identify orders and trades from three mutually

exclusive and exhaustive trading classes (PAT, AAT, and nonalgorithmic trading [NAT]), we also

apply a criterion to identify order messages from HFT in an approximate way. Wherever possible,

we perform our tests on the identified “proxy” HFT group, and in almost all cases, the results

for the HFT subgroup mimics the results of its superset, PAT. Our technique for identifying

HFT, which is based on the time elapsed between two occurrences (submission, modification,

or cancellation) of the same order number, ensures that we detect almost all liquidity-supplying

orders from HFT; however, it also ensures that we miss most of the liquidity-demanding orders

from this group. Still, this limitation should have little effect on our analysis because our focus is

1PAT can trade at lowerfrequencies; thus, not all PATis HFT. However,because our data do not explicitly identify HFT,

we concentrate on PAT to get a sense of the behaviorof HFT. We also introduce a technique to identify orders from HFT

and we include the results of the “proxy” HFT group in our analysis whereverpossible.

2The term “message,” here and in the rest of the article refers to the order messages and includes new orders, cancellations,

and revisions.

3Handbook of Statistics on Indian Securities Market 2014 published by Securities and Exchange Board of India (SEBI).

Nawn & Banerjee rRole of Proprietary Algorithmic Traders during Market Stress 643

on examining the liquidity-supplying behavior of various trading groups. Since we have both the

liquidity-demanding and liquidity-supplying orders from PAT, we also ensure that our analysis

remains robust to controlling for the liquidity-demanding orders of this group.

Providing liquidity is one of the two primary functions of a capital market (O’Hara, 2003).

Commonly,liquidity is measured by relative spread or effective spread. For our purposes though,

we do not need a general measure of liquidity; rather, we need to capture the variation in the

supply of liquidity coming from each trading group. Our measure of liquidity is thus computed

from the limit order book (LOB) snapshots as the sum of the size of the orders (depth) from each

trading group, stationed at first few steps of the LOB. The orders near the top of the book represent

genuine trading interest compared to orders awayfrom the top, because the execution probabilities

of limit orders fall significantly as we move away from the top (Aitken and Comerton-Forde,

2003). We reconstruct the LOB snapshots at one-minute interval for the 50 largest stocks on the

NSE for all trading days of 2013.

Employing severaldef initions of market stress such as high volatility, extreme stock movement,

and so on, we examine how PAT changes its supplied depth. We use two measures to determine

short-term volatility: the absolute return and the residual volatility from a market model. Following

the technique employed by Ahn, Bae, and Chan (2001), we explore how the depth supplied by

PAT reacts to short-term volatility and find that PAT depth near the top of the book is positively

related to stock-specific volatility under either definition. The change in depth supplied by PAT

in reaction to marketwide volatility is positive under the first definition and is neutral under

the second. This behavior, though, is limited only to PAT as AAT and NAT do not increase the

supply of liquidity in times of stress. The presence of faster traders in the market possibly instills

the fear of adverse selection in AAT and NAT. Furthermore, in the spirit of Bae, Jang, and Park

(2003), we distinguish between informational and transitory volatility and find that PATincreases

limit order supply more during high transitory volatility than during high informational volatility.

These results indicate that proprietary algorithmic traders, in general, do not panic in times of

high volatility but react more positively to transitory volatility because this volatility has less of

an adverse selection component.

To examine whetherPATprovides liquidity at all times, specifically at the most diff icult times,

we look at the extreme positive and extreme negative return minutes. When any stock is experi-

encing a high, negative, short-term return, limit buy orders become critical to the functioning of

the market. When everyone is looking to sell, there must be someone who is looking to buy—a

role served by market makers in traditional markets. We find that the proportion of buy-side depth

supplied by PAT in the top few quotes increases during these times. Similar behavior is observed

in sell-side depth when any stock is experiencing a high, positive,short-ter m return. These results

indicate that PAT supports the market when price movement is extreme.

Order imbalance is known to affect stock returns (Chordia, Roll, and Subrahmanyam, 2002;

Chordia and Subrahmanyam, 2004). Narayan, Narayan, and Westerlund (2015) document these

effects for intraday data. We test whether PATorder imbalance is related to future returns, as high-

frequency traders are considered informed and sophisticated and are often accused of causing

abrupt price fluctuations and trading ahead of the market. Although we find a negative long-term

price impact for PAT order imbalance, AAT order imbalance has significant positive predictive

power for future returns. The evidence suggests that proprietary algorithmic traders do not seem

to be informed traders, whereas agency algorithmic traders, who primarily execute the orders of

institutional traders, seem to be informed.

Our work contributes to the literature in several ways. First, we show that in a modern stock

market, a positive relation between short-term volatility and depth (Ahn et al., 2001) holds for

one trading class (PAT) but not for the other two(AAT and NAT). We then segregatethe effect of