Slow-Moving Capital and Execution Costs: Evidence from a Major Trading Glitch

Transcription

1 Slow-Moving Capital and Execution Costs: Evidence from a Major Trading Glitch Vincent Bogousslavsky Swiss Finance Institute EPFL vincent.bogousslavsky@epfl.ch Mehmet Sağlam Lindner College of Business University of Cincinnati mehmet.saglam@uc.edu May 25, 2017 Pierre Collin-Dufresne Swiss Finance Institute EPFL pierre.collin-dufresne@epfl.ch Abstract We investigate the impact of an exogenous trading glitch at a high-frequency market-making firm on standard measures of stock liquidity (effective and realized spreads) as well as on institutional trading costs (Implementation Shortfall and VWAP slippage) obtained from a proprietary data set. We find that stocks in which the firm accumulated large positions as a result of the trading glitch become substantially more illiquid on the day of the glitch. Effective spreads revert very quickly suggesting that market liquidity is resilient. Instead, institutional trading costs remain significantly higher for more than one week. We further document that all stocks for which the firm was a designated market maker become more illiquid, even if they were not heavily traded during the glitch, in the two days prior to being reassigned to another market maker. These findings are broadly consistent with slow-moving capital theories and suggest that high-frequency trading flash crashes may be associated with significant costs that are difficult to detect using standard liquidity measures. Keywords: Liquidity, Algorithmic Trading, Institutional Trading Costs, Slow-Moving Capital, Market Making. JEL Classification: G10. Preliminary draft. Comments are welcome. We thank Nanex and Eric S. Hunsader for providing the designated market maker data for NYSE-listed stocks. 1

2 1. Introduction When buyers and sellers arrive asynchronously, there is a role for providing immediacy (Demsetz (1968), Grossman and Miller (1988)). If capital is slow-moving then shocks to the risk-bearing capacity of immediacy providers should affect market liquidity (Duffie (2010)). Specifically, we expect negative shocks to the risk-bearing capacity of intermediaries to raise the trading costs of demanders of immediacy, at least temporarily, until the intermediaries capital has been replenished. Empirically, this is difficult to test however. First, it is difficult to measure exogenous shocks to the risk-bearing capacity of individual liquidity providers. Existing empirical literature finds evidence that shocks to inventory levels of NYSE specialists are related to trading costs, but such shocks could also be related to new information or changes in firm risk (Hendershott and Seasholes (2007)). To try to control for such endogeneity, Comerton-Forde et al. (2010) look at mergers between broker-dealer firms as a (positive) shock to their market making capital. Further, Hendershott and Menkveld (2014) rely on a statistical state space model to try to separately identify the pure temporary price pressure component due to inventory shocks. Second, it is not clear how one should measure market liquidity. Most existing papers use some measure of the quoted spread or of the effective spread, often decomposed into realized spread and price impact, to measure market illiquidity and its components (trading and order-processing costs, inventory costs, and adverse selection). But as pointed out by Domowitz et al. (2005) and others, the quoted spread only represents the inside quote for a small trade size. The effective spread improves on that as it is not restricted to the inside quote. However, it is still only an ex-post measure for a specific (typically small) trade size and does not correspond to an ex-ante measure of liquidity of a potential (large) trade size. Thus especially for large institutional investors it is not clear that it is an appropriate measure of liquidity. Indeed, large institutional investors typically rely on measures of implementation shortfall at the parent-order level, since their trades are broken down into multiple small trades (Perold (1988)). In this paper, we study the impact of a major trading glitch, originating from the erroneous implementation of a trading software at a large high-frequency market-making firm (Knight Cap- 2

3 ital, KC henceforth), on institutional trading costs. This exogenous shock occurred on August 1, 2012 during the first thirty minutes of trading and caused numerous erroneous trades on a set of NYSE-listed stocks. Using public data, we identify the set of stocks affected by the trading glitch and measure the impact on the cost of trading the affected stocks on the day of, as well as subsequent to, the glitch. We use two measures of trading costs. First, we compute standard market microstructure measures of effective spread, realized spread, and price impact using five-minute return data obtained from TAQ data stamped to the millisecond. Second, we use a proprietary execution data set from a large investment bank that contains measures of implementation shortfall (IS) and volume weighted average price slippage (VWAPS) for a large number of institutional investors parent-orders that are broken down into several children-orders and traded using standard execution algorithms. Since the trading glitch was due to a technical problem that became very quickly commonknowledge (Reuters published its first news feed around 10am on the day of the glitch), the dramatic long and short positions accumulated by KC on the glitch-affected stocks were unrelated to fundamental information on the underlying stocks themselves. Thus, they constitute a natural experiment to test the effect of exogenous inventory shocks for a liquidity provider on market liquidity. We expect two types of effects. First, since KC had to rapidly unload these acquired positions to be able to satisfy its margin calls, these shocks generate predictable future supply shocks that should affect current and future trading liquidity if risk-capital is limited and slowmoving (Duffie (2010)) and/or because of predatory trading (Brunnermeier and Pedersen (2008)). Interestingly, this effect is stock specific. Second, the inventory shocks together constitute a major shock to the risk-bearing capacity of KC (losses on its position were estimated at around $460 million and led to KC eventually beeing bought up by a competitor, see Section 2). Since KC was a so-called designated market maker (DMM) for a long list of NYSE stocks, we would expect, if market making capital is scarce and slow-moving, that such a shock would affect the market liquidity of all stocks for which KC was a DMM, whether or not they were heavily traded as a result of the glitch itself. To distinguish between these two effects, we report the impact of the glitch on our two sets of trading cost measures, for stocks that experience an abnormally high trading volume during the 3

4 glitch (glitch-stocks) and for stocks that experience a normal trading volume, but for which KC was a designated market maker (DMM-stocks). Our main findings are as follows. First, all transaction cost measures experience dramatic increases on glitch-stocks on the day of the glitch. Standard microstructure measures (effective spread, realized spread) revert very quickly towards pre-glitch levels suggesting that market liquidity is quite resilient. However, the increase in institutional trading cost measures (IS, VWAPS) remains statistically and economically significant for more than one week after the glitch. As a result of this persistence, the impact of the trading glitch on the cumulative increase in institutional trading costs is substantial. In a back-of-the-envelope calculation we estimate that the total cost of the glitch to institutional traders on glitch affected stocks is on the order of $100 million. Second, DMM stocks experience a significant increase in their effective spreads and in institutional trading costs on the day-of and following the glitch, that reverts starting from two days after the glitch. Since KC s stocks were reassigned to another DMM (Getco) three days after the glitch, this seems consistent with the market making capital hypothesis. Comforting the idea that market making capacity was reduced on DMM stocks as a result of the glitch, we find that turnover on the DMM stocks falls in the two days after the glitch but recovers after the stocks are reassigned to a different DMM. A positive impact of DMM on liquidity is consistent with the analysis of Clark- Joseph et al. (2016). These authors use an unexpected trading halt on the NYSE as a natural experiment to show that DMM participation lowers spreads. The finding that glitch stocks liquidity is significantly reduced could be consistent with both the market making capital hypothesis as well as the persistent supply shock/predatory trading hypothesis. To better understand the mechanism and, especially, why we observe such a striking difference between traditional microstructure spread measures and institutional trading cost measures when it comes to the persistence of the effect of the glitch, we perform additional tests. First, we find that the turnover of glitch stocks increases very significantly on the day of the glitch and remains abnormally high for more than one week after the glitch. Second, we analyze the market depth on the glitch stocks and find that it is significantly decreased in the week following the glitch. Third, we distinguish between institutional trades that seek an execution in the same direction as the expected supply shock based on the glitch. As an example, consider a stock that 4

5 KC bought as a result of the glitch. Then we expect that KC (or the intermediary who bought the position from KC) will seek to offload its position in the week following the glitch. Thus, the persistent supply shock hypothesis would suggest that an institutional trade to buy would get more favorable execution than a trade to sell, since the latter would be on the same side as that desired by KC. We find strong evidence that these same side trades have persistently higher execution costs than opposite side trades, that effectively provide liquidity to KC. Our findings are thus broadly consistent with theories of slow-moving capital (Duffie (2010)). They also show that liquidity provision by high-frequency traders, in as much as it generates more instances of flash-crashes similar to the ones studied in this paper, may have adverse consequences for market participants, such as institutional investors, that are difficult to detect using traditional measures of market liquidity, such as effective spreads. 1 The rest of the paper is organized as follows. Section 2 provides brief background information on the trading glitch. Section 3 describes the liquidity measures used in the paper. Section 4 presents the data and methodology. Section 5 discusses the impact of the glitch on affected stocks. Section 6 discusses the impact of the glitch on stocks that are not affected by the glitch but for which KC is a designated market maker. Section 7 presents additional findings on the economic mechanism behind the observed changes in stock liquidity and robustness tests. Section 8 concludes. 2. The August 1, 2012, Trading Glitch This section describes the August 1, 2012, trading glitch and discusses the potential impact of the glitch on stock liquidity Timeline of Events On August 1, 2012, Knight Capital (KC) experienced significant issues with its automated order routing system. Over the previous days, KC had updated its servers so that its customers could benefit from the introduction of a new retail liquidity program on the NYSE. However, one of KC s 1 Although the link between HFT and institutional trading cost has not been heavily studied, there are numerous empirical studies documenting that HFT activity is positively correlated with improvement in market quality measures; see, for instance, Hendershott et al. (2011) and Brogaard et al. (2014). 5

6 servers was not updated correctly. The server did not recognize when an order had been executed and therefore kept sending new orders to the market. Because of the system s complexity and a lack of adequate control procedures, KC s technology team was not able to identify the source of the problem immediately. As a result, the glitch made KC s system send millions of orders to the market over the first 45 minutes of trading. Consequences were substantial. At some point, KC held a $3.5 billion net long position in 80 stocks and a $3.15 billion net short position in 74 stocks. These positions led KC to realize a $460 million loss. 2 The glitch resulted in large price movements and high volume in NYSE-listed stocks. Several NYSE-listed securities had larger share volume than SPY the most liquid S&P 500 ETF during the initial 30 minutes of trading on August 1. For instance, Juniper Networks (NYSE: JNPR) had a share volume larger than 20 million in this 30-minute period. In comparison, its daily 20-day average share volume was 8 million. Figure 1 shows average price and cumulative share volume every 15 seconds over the first hour of trading for Juniper Networks on August 1. The average price exhibits large swings, and volume is abnormally large before 10 AM. A similar pattern is observed for the number of trades. Such volume patterns can help us identify which stocks were subject to the glitch, as detailed in Section 4.2. Figure 1: Price and share volume in the first hour of trading for Juniper Networks on August 1, The figure shows average price and cumulative share volume (in thousands) every 15 seconds. Price Share Volume $ 17,000s :30 9:40 9:50 10:00 10:10 10:20 10:30 9:30 9:40 9:50 10:00 10:10 10:20 10:30 KC s traders managed to reduce the size of the positions on the day of the glitch, but KC 2 SEC Administrative Proceeding File No (October 16, 2013) available at litigation/admin/2013/ pdf. 6

7 would not have been able to open for business on the next day because of regulatory capital requirements. As a result, KC struck a deal with Goldman Sachs Group Inc. to sell its portfolio at a 5% discount. 3 While the glitch happened on Wednesday August 1, KC s survival was still uncertain on the following Friday because of the large trading loss. In particular, KC secured a credit line to stay in business through late Friday and was actively involved with investors and rivals to secure additional funding. 4 On the next Monday morning, it was announced that KC had signed a $400 million rescue package from six other firms. In December 2012, KC was acquired by GETCO, a competing firm. The short-term survival of KC was an important issue since KC was the designated market maker for about 500 NYSE-listed stocks. Designated market makers (DMM) succeeded NYSE specialists. They have obligations to maintain a fair and orderly market in their stocks, quote at the NBBO a specified percentage of the time, and facilitate price discovery throughout the day as well as at the open, close and in periods of significant imbalances and high volatility (NYSE fact sheet, 2009). The trading glitch impaired KC s ability to act as a DMM. First, KC s market making activity appears to have dropped substantially following the glitch in part because market participants such as retail brokers stopped routing their orders to KC. Second, NYSE reassigned stocks for which KC was the DMM to GETCO on August 6, Potential Impact of the Glitch on Liquidity The KC trading glitch allows us to cleanly evaluate specific theories of stock market liquidity. Indeed, stock liquidity could be affected by the glitch for several reasons. First, the trading glitch represents a large shock to the inventory of a market participant recall from Section 2 that KC held aggregate long and short positions worth billions of dollars. If KC rushes to unload its positions, then we expect an increase in trading costs and a deterioration of liquidity for affected stocks in the following hours and days. Similarly, predatory trading could 3 Source: Knight Held $7 Billion of Stocks Due to Glitch, Wall Street Journal, August 13, Source: Knight Capital Group hustling to find a buyer or secure funding, Los Angeles Times, August 4, Sources: Knight Capital heads into make-or-break weekend, Reuters, August 4, 2012; NYSE Euronext temporarily re-assigns Knight Capital DMM responsibilities to GETCO DMM for certain securities, Business Wire, August 6,

8 worsen the impact of KC s unloading. 6 Second, the trading glitch represents an exogenous shock to a market maker s capital. Less capital should limit the ability of a market maker to provide liquidity. Comerton-Forde et al. (2010) show that lagged NYSE specialist inventories and trading revenues generated by overnight inventories predict bid-ask spreads for the next day. Relatedly, Coughenour and Saad (2004) document the existence of a common component in liquidity for all stocks assigned to a particular specialist. These papers do not rely on an exogenous shock to identify the causal relation between market maker s capital and stock liquidity. In addition, these papers cover periods prior to the apparition of NYSE s designated market makers. Both theories predict a decrease in liquidity and may be complementary. However, they also make specific predictions. The inventory shock theory predicts an asymmetric impact on stock market liquidity: it should be especially costly to trade in the opposite direction of the inventory imbalance. For instance, if KC has accumulated a large long position in a stock, it should be costly to sell the stock in the aftermath of the glitch. The market making capital theory predicts that stocks for which KC is the designated market maker but are not affected by the glitch should also experience a decrease in liquidity. 3. Liquidity Measures In this section, we describe the liquidity measures that we use in the analysis. In particular, we discuss two most frequently used cost measures for institutional trading: implementation shortfall (IS) and volume-weighted average price (VWAP) slippage. 6 Even though Goldman Sachs bought KC s portfolio at a discount, a similar reasoning applies. First, the market only learned about the deal more than two days after the glitch, on Friday afternoon (see, for instance, Goldman support boosts Knight Capital, Financial Times, August 3, 2012). Second, Goldman Sachs still had to unload the position. 8

9 3.1. Standard Liquidity Measures (Spreads) We compute the following standard liquidity measures for a transaction on stock i at time t: (1) (2) (3) Effective Spread i,t = 2 ln P i,t ln M i,t, Realized Spread i,t = 2q i,t (ln P i,t ln M i,t+5mn ), Price Impact i,t = 2q i,t (ln M i,t+5mn ln M i,t ), and where M i,t denotes the midpoint of the best quote available immediately preceding the transaction, and q i,t equals 1 for buy orders and -1 for sell orders. Trades are signed with the Lee and Ready (1991) algorithm using the first quote available prior to a trade. The liquidity measures are also computed in dollar terms. We follow the literature and compute realized spread and price impact using the quote midpoint five minutes after a trade Implementation Shortfall As a first measure of institutional trading costs, we use implementation shortfall (IS) as introduced by Perold (1988). IS is the widely preferred measure of trading cost for institutions and has been frequently employed in the literature to proxy institutional trading cost. This approach is based on comparing the weighted average price of the actual trades to a benchmark price observed prior to the execution period. Formally, implementation shortfall of the kth execution in our data is given by (4) Implementation Shortfall k = sgn (Q k ) P avg k P k,0 P k,0, where P avg k is the value-weighted execution price of the parent order and P k,0 is the mid-quote price of the security (arrival price) when the parent order starts being executed VWAP Slippage Implementation shortfall considers an ex ante static benchmark and does not consider any permanent price changes during the execution period. For this reason, one popular ex-post benchmark for 9

10 average execution price is the volume-weighted average price during the trading interval. Formally, VWAP in a particular trading interval period is defined as follows. Suppose that there are N trades during this period. Given the sequence of trades at prices P 1,..., P N with corresponding quantities V 1,..., V N, VWAP is given by (5) P = Ni=1 P i V i Ni=1 V i. Using this definition, VWAP slippage for the kth execution in our data equals (6) VWAP Slippage k = sgn (Q k ) P avg k P k, P k where P k is the realized VWAP over the kth execution period. Minimizing VWAP slippage is also the main objective when the investor chooses VWAP as the algorithmic trading strategy. About 70% of the executions in our dataset are executed according to the VWAP algorithm. 4. Data and Methodology 4.1. Data We obtain stock information from the Center for Research in Security Prices (CRSP) and focus on common stocks (CRSP share code 10 or 11) listed on the NYSE or NYSE MKT (CRSP exchange code 1 or 2). Stocks are required to have a price greater than $2 and lower than $1000, as well as a market capitalization greater than $100 million at the end of June For these stocks, trades and NBBO quotes are obtained from the Trade and Quote (TAQ) millisecond data base. Locked and crossed quotes are excluded. Transaction data are filtered as in Chordia et al. (2001). In total, we obtain data for more than 1,300 stocks from June 1 to September 28, Earnings announcement dates come from I/B/E/S. We also use novel proprietary execution data from the historical order database of a large 7 The TAQ millisecond trade file is missing trade data for most symbols between D and FDX on August 1, We use TAQ data stamped to the second to compute liquidity measures on these stocks with the corrections proposed by Holden and Jacobsen (2014). Five stocks affected by the glitch are concerned. 10

11 investment bank providing algorithmic trading services ( The Bank ). The orders originate from a diverse pool of investors, such as institutional portfolio managers, quantitative investment funds, internal trading desks and retail customers. The Bank offers a large selection of algorithms to match the investors trading styles and expectations. The dataset consists of two frequently used algorithms, the volume weighted average price (VWAP) and the percentage of volume (POV). The VWAP algorithm is designed to achieve an average execution price that is as close as possible to the volume weighted average price over the trading interval. Similarly, the main objective of the POV algorithm is to have constant participation rate in the market within the trading interval. This proprietary dataset provides a rich set of attributes. For each order, the data contains trade- and stock-level statistics. Trade-level statistics include order size, direction of the order (buy or sell), order start and end times, participation rate (the ratio of order size to the total volume during the trading interval), average execution price, proportional bid-ask spread and mid-quote volatility based on the duration of the execution. Stock-level information includes average daily volume, proportional bid-offer spread and mid-volatility of the stock on the trading day along with their rolling averages over the last 20 trading days prior to the execution. The traded asset universe includes all S&P 500 stocks. All executions occur between January 2012 and December 2012 inclusive and last between 5 minutes and 6.5 hours, the duration of a regular trading day. All orders have been fully filled without intermediate replacements or cancellations. Our sample consists of 39,570 executions coming from 18,357 buy and 21,213 sell orders. The trading algorithms used are 29,027 VWAP and 10,543 POV. There are orders per trading day on average. The highest number of executions on a single stock is 330 which corresponds to 0.83% of all executions. Panel (a) of Table 1 provides additional summary statistics for our complete execution data. [Insert Table 1 here] There is a wide range of participation rates across executions with an average (median) of 5.24% (0.89%). Most of the orders make up less than 1% of daily volume. This is expected as the dataset contains the most liquid S&P 500 stocks. Average and median implementation shortfall are 3 bps. 11

12 Average execution duration is 2.67 hours. Finally, average and median percentage return realized during an execution are roughly zero. As a point of comparison, Table 2 provides descriptive statistics for the standard liquidity measures as well as various stock characteristics for the full sample and the sample of stocks for which we have execution data. Stocks that are in our execution data set tend to be larger and more liquid than stocks in the full sample. [Insert Table 2 here] 4.2. Identification of Affected Stocks Available public data does not allow us to identify with certainty stocks affected by the trading glitch. According to a KC s statement on August 2, the erroneous orders are limited to NYSE stocks. In addition, several stocks experience a tremendous increase in trading volume in the first half-hour of trading on August 1 absent any specific corporate news. This abnormal increase in volume can help us identify affected stocks. Given that the number of trades drops significantly after 10:00 AM, we focus on the number of trades occurring in the first half-hour of trading and compute it for each stock in the dataset. We only consider stocks with at least 5,000 trades in this interval. As a benchmark, the number of trades for the most active S&P 500 ETF, SPY, is 31,341. We find that 50 stocks have more trades than SPY, which is highly unusual. We label these stocks as glitch-affected. In addition, we compute for each stock the average number of trades between 9:30 and 10:00 AM over the past five trading days. We label a security as affected by the glitch if its ratio of August 1 number of trades (between 9:30 and 10:00 AM) over average past five trading days number of trades (between 9:30 and 10:00 AM) is larger than 10. Table 3 reports the list of affected stocks with their number of trades. In total, we obtain 61 affected stocks, among which 35 are S&P 500 stocks. 8 We verify that all of these stocks display similar abnormal volume pattern as JNPR in Figure 1. Among affected stocks, 35 stocks are in our 8 We exclude KC s own stock (NYSE: KCG) from our analysis because it does not appear to be subject to a similar anomalous volume pattern. Trading volume increases only around 9:45 and is associated with a price decrease. These variations can potentially indicate when the market starts realizing that something is wrong with KC. 12

13 execution dataset. More precisely, 31 out of 35 securities are executed on the day of the glitch, and there are 50 different parent executions implemented on the affected group. [Insert Table 3 here] To provide a clear picture of the composition of our data, Table 4 reports the allocation of stocks among the following groups: stocks affected by the glitch; stocks that have KC as DMM; stocks for which we have institutional execution data; and stocks that are part of the S&P500. Interestingly, the vast majority of stocks affected by the glitch do not have KC as DMM. This property of the sample is valuable since it allows us to test whether stocks that are not affected by the glitch and have KC as DMM experience an increase in trading costs. [Insert Table 4 here] 4.3. Methodology To formally whether liquidity measures increase for affected stocks, we run panel regressions for each daily liquidity measure. Let L i,t be the daily liquidity measure of stock i on date t. Our baseline panel regression is (7) L i,t = 5 k=0 δ k Glitch k,i,t + δ P GlitchPost i,t + j β j Control j,i,t + ɛ i,t, where Glitch k,i,t equals one for affected stocks on the date k trading days after the glitch (k = 0,..., 5), and GlitchPost i,t equals one for affected stocks on any date more than five trading days after the glitch. Hence, the regressions include separate dummy variables to identify changes in the liquidity of affected stocks in the week following the glitch. In addition to stock and day fixed effects, we use the following stock-specific control variables at each date: log price, log turnover, log market capitalization, and volatility (volatility is computed based on daily high and low prices as in Parkinson (1980)). The spreads and control variables are winsorized at 0.05% and 99.95% as in Hendershott et al. (2011). To test the market making capital theory, we augment regression (7) to include dummies for nonaffected stocks that have KC as DMM. More precisely, we add a dummy for the day of the 13

14 glitch, each of the five trading days following the glitch, and the post glitch period. (8) 5 5 L i,t = δ k Glitch k,i,t + δ P GlitchPost i,t + γ k DMM k,i,t + γ P DMMPost i,t k=0 k=0 + j β j Control j,i,t + ɛ i,t, where DMM k,i,t equals one on the k th trading days after the glitch, if stock i has KC as DMM and is not affected by the glitch. This double restriction ensures that the coefficients can be directly interpreted with respect to the market making capital theory. As shown in Table 4, only five affected stocks have KC as DMM. Regressions (7) and (8) give similar estimates of δ k. Hence, we only report results for the second regression. For institutional trading costs, we run a slightly modified version of regression (8) where we do not estimate a separate coefficient for each day following the glitch because of the limited number of executions available on each day for affected stocks. 9 Instead, we estimate the following regression: (9) C k = δ 0 GlitchDay0 k + δ W GlitchNextWeek k + δ M GlitchAug9ToAug31 k + δ P GlitchPost k + γ 0 DMM0 k + γ W DMMNextWeek k + γ M DMMAug9ToAug31 k + γ P DMMPost k + j β j Control j,k + ɛ k, where C k is VWAP slippage or implementation shortfall, GlitchDay0 k equals one for affected stocks on the day of the trading glitch, GlitchNextWeek k equals one for affected stocks executed within the next 5 business days after the trading glitch, i.e., before August 9th, GlitchAug9ToAug31 k equals one for affected stocks executed on or after August 9th and on or before August 31st. Finally, GlitchPost k equals one for affected stocks on any date after August 31st. To test the market making capital theory, DMM0 k equals one for Knight-DMM-affected, but not-glitch-affected, stocks on the day of the trading glitch, DMMNextWeek k equals one for Knight-DMM-affected stocks executed within the next 5 business days after the trading glitch, i.e., before August 9th, DMMAug9ToAug31 k 9 We do not have a complete panel as some stocks may be executed more than once in a day or may not be executed at all. 14

15 equals one for Knight-DMM-affected stocks executed on or after August 9th and on or before August 31st. Finally, DMMPost k equals one for Knight-DMM-affected stocks on any date after August 31st. In addition to stock and day dummies, we use the stock- and execution specific control variables as described in Section 4.4. We are concerned with heteroskedasticity, contemporaneous correlation across stocks, and autocorrelation within each stock and adjust our standard errors by clustering on calendar day and stock throughout the analysis as suggested by Petersen (2009) Determinants of Institutional Trading Costs A broad literature analyzing the variation in trading costs finds that relative order size, market capitalization, asset volatility, share turnover, and bid-offer spread are the main determinants (see e.g., Domowitz et al. (2001)). Based on this primary list, we use the following order- and stock-level characteristics as control variables in the main analysis: Order Size. Price impact of a trade increases with the size of the order. In order to capture variations related to this, we use participation rate (the ratio of order size to interval volume) and fraction of daily volume (the ratio of order size to daily volume). Share Volume and Turnover. All else equal, higher volume and turnover may signal higher liquidity and lower trading costs. We use relative daily volume (the ratio of daily volume to average daily volume over the past month), and interval turnover (the ratio of interval volume to number of shares outstanding) to control for differences in volume profiles associated with an execution. Volatility. In the presence of higher volatility, market makers have the incentive to widen the spread between their quotes leading to higher trading costs. In order to control for differences in volatility, we use interval volatility (mid-quote volatility during the interval expressed in annualized percentage), and daily volatility (mid-quote volatility during the trading day expressed in annualized percentage). Bid-offer spread. Higher bid-offer spread will translate into higher trading costs. We use interval spread (average bid-offer spread during the interval expressed in basis points), and 15

16 daily spread (average bid-offer spread during the trading day expressed in basis points) to control for variations in this common liquidity metric. Execution Duration. Higher urgency in trading may result in higher costs too. We use execution duration expressed as a fraction of total trading hours to control for differences in urgency level of the executions. Stock Returns. In order to control for extreme contemporaneous price movements, we use absolute and raw values of the stock daily returns. Earnings Announcement Days. On earnings announcement days, firms may also have increase in trading volume due to the mere desire of opening or closing position in anticipation of good or bad quarterly results. Therefore, we also include a control variable, IsEarningsDay, which takes a value of 1 if the executed stock has an earnings announcement on the execution day. Type of Algorithm. VWAP and POV strategy may differ in their expected trading costs allelse equaling. POV strategy is usually employed in larger orders and thus can be intrinsically more costlier. Thus, we include a control variable, IsVWAP, which takes a value of 1 if the execution algorithm is aimed to minimize VWAP slippage. These various stock- and order-level characteristics serve as the main control variables for all models explaining institutional trading costs. We find that logarithm of market capitalization or prices are highly correlated with the remaining independent variables (with the presence of firm dummies) and are thus omitted to mitigate multi-collinearity. 5. Impact of the Glitch on Affected Stocks In this section, we examine the impact of the glitch on the liquidity of affected stocks on the day of the glitch and its aftermath. 16

17 5.1. Standard Liquidity Measures We start with a visual inspection of these liquidity measures on the day of the glitch. Figure 2 plots the average 1-minute effective spread, realized spread, and adverse selection for affected and nonaffected stocks on August 1. The glitch has a significant but short-lived impact on liquidity measures of affected stocks (left chart). Effective spreads stabilize after 10 AM and do not exhibit any marked variation for the rest of the day. Realized spreads spike during the glitch but then fall back rapidly. Market liquidity, as measured by spreads, appears to be resilient. The right chart of Figure 2 plots liquidity measures for nonaffected stocks. As can be seen, spreads tend to be large at the open and decline gradually. By comparing the two charts, it is clear that the glitch generates extreme movements in realized spreads and price impact. [Insert Figure 2 here] Next, we check whether there is a persistent increase in liquidity measures in the days following the glitch. Daily liquidity measures for each stock are obtained by averaging dollar-weighted intraday measures. Results are similar when using share-weighted intraday measures. Figure 3 plots the cross-sectional quartiles of each liquidity measure for affected and nonaffected stocks around the day of the glitch. Visually, there is no evidence of a persistent increase in trading costs in the days following the glitch. [Insert Figure 3 here] To test more formally whether spreads increase for affected stocks, regression (8) is estimated using daily liquidity measures from June 1 to September 28, 2012 (84 days). To be included, a stock is required to have valid data for all variables every day of the sample. Standard errors are clustered by firm. 10 Table 5 shows the results. The results confirm the visual evidence in Figure 3. There is no significant increase in the spreads of affected stocks in the aftermath of the glitch. If anything, effective spreads of affected stocks appear to be lower on the day following the glitch. This conclusion is unchanged when we restrict the sample to NYSE-listed S&P 500 stocks only. 10 The results are similar using White s standard errors. Double clustering by day and firm using the method of Thompson (2011) gives similar results, but several of the standard errors are not well defined. The coefficients on realized spreads and price impact may not sum up exactly to the coefficient on effective spreads because the liquidity measures are winsorized as described above. 17

18 [Insert Table 5 here] Standard liquidity measures may not pick up the dimension of liquidity that is affected by the glitch. Not everything remains stable following the glitch. As can be seen in the last column of Table 5, affected stocks turnover increases in the aftermath of the glitch. The increase in turnover is economically large and strongly significant for up to four days after the glitch. 11 Strikingly, the increase in turnover for affected stocks for each of the four days after the glitch remains close to two-third of the increase in turnover observed on the day of the glitch. Standard liquidity measures do not appear to reflect this surge in trading activity Institutional Trading Costs We first describe the institutional trading data during the glitch period. Panel (b) of Table 1 provides the summary statistics for executions occurred on August 1, There are 662 parent orders executed by The Bank on that day which is substantially higher than the daily average of We observe that when compared with the full sample the order sizes are smaller relative to both interval and daily volume. Consequently, the average values for implementation shortfall and execution duration are also smaller. We have 50 executions implemented on the affected group of the stocks on August 1. More precisely, 31 affected securities are executed on the day of the glitch. The maximum number of executions in a single security from the affected group is merely 3 (SWN and WFC) resulting in a homogeneous distribution between the affected stocks. We also have executions from the affected group in the days following the trading glitch. There are 10, 6, 6, 14 and 20 executions from the affected group respectively in the next five business days totaling 56 executions. Furthermore, we have 208 and 1001 affected executions in the remaining days of August and during the last quarter of 2012, respectively. We find that more than 90% of the executions start being traded after the end of the glitch, specifically after 10:30 AM. Therefore, our data will mainly capture the aftermath impact of the trading glitch. 11 This is for S&P 500 stocks. The results are similar when we examine the full sample. 18

19 We first examine visually whether there is any pronounced increase in the institutional trading costs of glitch-affected stocks. For this purpose, we examine the difference in VWAP slippage means between the affected and unaffected groups of stocks during the period of July 2012 and September 2012 (inclusive). [Insert Figure 4 here] Figure 4 illustrates that the mean difference between the slippage values spikes up during the week following the day of the glitch. We observe that the difference seems to lie at a marginally elevated level for the rest of the month compared to its pre-glitch period. The mean difference seems to revert its pre-glitch level during the month of September. Using the complete dataset, we run the model specified in equation (9) using VWAP slippage and implementation shortfall as our cost measures. [Insert Table 6 here] Table 6 illustrates the regression estimates. We find that participation rate, fraction of daily volume, daily bid-offer spread, execution duration and the type of the trading strategy (VWAP or POV) are the statistically significant variables for both cost metrics. As expected, we find that execution costs are positively correlated with participation rate, fraction of daily volume, daily bid-offer spread, and execution duration and VWAP orders are cheaper to execute. Overall, we observe that our cost model is able to explain much higher variation of VWAP Slippage. The main variables of interest are of course the dummy variables associated with affected executions during and after the trading glitch. Overall, the results confirm the visual evidence in Figure 4. We find that institutional trading costs are economically and statistically significantly higher on the day of the trading glitch. The estimated cost increase due to the trading glitch is approximately 6 bps and 15 bps for VWAP slippage and implementation shortfall, respectively. Since the median VWAP slippage (IS) in our dataset is 1 bp (3.1 bps), the estimated cost increase is indeed economically substantial. We can also interpret the estimated coefficient by comparing it with the most important execution-level parameter, participation rate. A 6 bps cost increase in VWAP slippage or a 15 bps cost increase in IS roughly corresponds to an additional 45% increase 19

20 in participation rate. That is, all else equal, a glitch-affected execution with a participation rate of 10% is expected to have a similar implementation shortfall compared to an unaffected execution with a participation rate of 55%. In the previous section, we observed that the standard liquidity (spread) measures revert quickly to pre-glitch levels and, in fact, do not detect any abnormal change in liquidity in the days following the trading glitch. Instead, the institutional trading cost measures exhibit high persistence for more than one week. We find that for both cost measures, executions occurring in the following week (within the next 5 business days) are exposed to roughly the same cost increase as on the glitch-day itself. As measured by VWAP slippage (IS), executions in the affected group realize on average an additional 7 bps (15 bps) in the following week. Furthermore, VWAP slippage seems to imply even higher persistence. Our dummy variable, GlitchAug9ToAug31, is also significant for VWAP slippage indicating additional 1.2 bps of cost increase for the rest of August. Note that although this increase is roughly 20% of our earlier estimates of the cost increase, this is still a substantial cost considering the additional three-week horizon and the median VWAP slippage is merely 1 bp. We observe that IS does not pick up any cost increase for this period, which might be related to the fact that most of our executions use as their objective to minimize VWAP slippage. Using these estimates for the cost increase, a back of the envelope calculation can be computed for the total welfare loss for institutional traders. Using July 2012 data, the average dollar volume per day on the glitch-affected set of stocks was roughly 20 billion. Given that a majority of the daily trades are due to institutional investors, we roughly attribute 10 billion of dollar volume to institutional investors. Thus, using implementation shortfall (VWAP Slippage) as our cost measure, the total loss due to the trading glitch is approximately $18 ($7) million per day during the first five days. Thus, the total IS loss for the first five days amounts to $90 million. The findings documented in this section are consistent with the slow-moving capital theory presented in Duffie (2010). They also highlight that the effects of a trading glitch, such as a flash-crash, may have persistent adverse effect on the trading liquidity faced by institutional investors, which are difficult to detect using standard (spread-based) market liquidity measures. In the following section we provide further evidence to explain why institutional trading cost remain 20

21 persistently higher after the trading glitch, and why traditional spread-based measures do not reflect this increase in trading costs. 6. Impact of the Glitch on DMM Stocks In this section, we interpret our findings from the perspective of market making capital theory. One potential reason for the decrease in liquidity can be attributed to the KC s role of designated market maker in these stocks. When KC s capital decreases, its liquidity provision may drop subsequently. As indicated by Table 4, only 5 glitch-affected stocks are under KC s designated market making roles: NAVB, MHR, P, PL and SWN. Only SWN is part of S&P 500 Index and our institutional data set has executions on this stock. On the contrary, 114 stocks are not affected by the trading glitch and their designated market maker is KC. We call these stocks the DMM-stocks. 17 of these stocks appear in our institutional data set in the post-glitch period Impact on Standard Liquidity Measures We now focus on the coefficients for nonaffected stocks that have KC as DMM in regression (8). Table 5 reports the results. Effective spreads on nonaffected stocks that have KC as DMM do not increase on the day of the glitch relative to unaffected stocks that do not have KC as DMM. However, effective spreads increase significantly on the day after the glitch. This increase is strongly significant for S&P 500 stocks (fourth column). For S&P500 stocks, effective spreads decline markedly following the takeover of KC s stocks by GETCO (three days after the glitch). This decline appears to be driven by a decrease in realized spreads. In terms of turnover (last column), the signs of the coefficients are consistent with a decrease in market making capacity. Turnover decreases following the glitch and increases when GETCO steps in. The coefficients seem, however, economically small and are not statistically significant. Overall this analysis appears partly in line with the results of Comerton-Forde et al. (2010). Even though designated market makers may play a less important role than NYSE specialists, we find an increase in trading costs on the day after the glitch for nonaffected stocks that have KC as DMM. The importance of DMM for liquidity is also consistent with the recent study of 21

22 Clark-Joseph et al. (2016). At the same time, it is puzzling that the coefficient for S&P500 stocks is negative and significant on the day of the glitch. To get a better sense of what happens, we estimate a regression similar to (8) but using 5-minute average dollar-weighted effective spreads. We restrict the sample to S&P 500 stocks to ensure that liquidity measures can be computed for each five-minute interval of the trading day. The regression includes indicators for glitch-affected stocks and DMM stocks on August 1, as well as an FOMC indicator from 14:10 until 15:10 on August 1. The DMM stocks indicator for five-minute interval k equals one in interval k on August 1 for stocks that have KC as DMM and are not affected by the glitch. Figure 5 reports the coefficients and t-statistics on the DMM indicators. 12 Spreads on DMM stocks spike at the time of the FOMC announcement on August 1. This evidence fits the story that DMM matter and that their importance increases when voluntary liquidity providers withdraw (i.e., at the time of the announcement). We cannot, however, rule out that these stocks are not more sensitive to FOMC announcement than other S&P500 stocks. As suggested by Figure 5, the negative and significant coefficient on DMM0 in Table 5 for S&P500 stocks is driven by the first five minutes of trading. Excluding this interval in the estimation of equation (8), the DMM0 coefficient increases to with a standard error of [Insert Figure 5 here] 6.2. Impact on Institutional Trading Costs In this section, we document that institutional trading costs increase on the day of the glitch for DMM stocks. We have 14 parent-order executions occurring on August 1, 2012 and we have additional 20 parent-order executions during the next 5 trading days. We focus on the DMM coefficients in regression (9). These coefficients identify the impact on institutional cost measures for executions which are not on stocks affected by the glitch but still affected indirectly via KC s DMM activity. Table 6 reports the regression estimates. We find that institutional trading costs are significantly higher, both economically and statistically, for KC s DMM-stocks on the day of the glitch. However, 12 The coefficients for glitch-affected stocks confirm the visual evidence in Figure 2 and are not reported. 22

23 we do not observe significant changes in the post-glitch period. This is expected as KC s DMM responsibilities are transferred to GETCO on August 3, Overall, the sign of the dummies are negative in the post-glitch period suggesting mean-reversion after the transfer. 7. Additional Findings and Robustness Tests 7.1. Depth at the Best Prices One potential explanation for the increase in institutional trading costs can be due to decrease in the available depth at the best prices. Standard liquidity measures may not fully capture the variation in the depth so such an outcome would also be a potential resolution for the discrepancy between institutional trading costs and standard liquidity measures. [Insert Table 7 here] We obtain our depth measure using the time-weighted average of available share volume (in $K dollars) at the best prices during a trading day. Table 7 illustrates that available depth has strictly decreased in the days following the glitch Predatory Trading We test whether some of the cost increase is due to potential predatory trading as KC or Goldman Sachs will be forced to liquidate the inventory that it has built up during the glitch. In order to test this hypothesis, we test for any significant cost increase for executions involving stocks that experienced large positive and negative returns during the trading glitch. If a stock had large positive return during the glitch, this suggests that KC formed a large long inventory on the stock that it needs then to liquidate. We test whether institutions that are trading in the same desired direction as KCG or GS subsequent to the glitch pay larger execution costs on these stocks with extreme returns during the glitch period. We create a dummy variable, KnightSameSide, that equals 1 if the stock realized ±2.5% return during the glitch period and if the institutional investor trades in the opposite direction of the extreme return, that is in the same direction as KCG or GS if they try to trade out 23