Christine X. Jiang Department of Finance Fudan University Shanghai, China September 2018

Transcription

1 Active trading in passive ETFs: The role of algorithmic trading Archana Jain Saunders College of Business Rochester Institute of Technology Rochester, NY Chinmay Jain Faculty of Business and IT University of Ontario Institute of Technology Oshawa, Ontario L1H 7K Christine X. Jiang Department of Finance Fudan University Shanghai, China September

2 Active trading in passive ETFs: The role of algorithmic trading Abstract We explore algorithmic trading and its effect on Exchange-traded funds. Using cancel rate, trade-to-order volume ratio, and order fragmentation as a proxy for AT (Algorithmic trading), we find higher AT activity in ETFs results in lower deviation from NAVs, and lower persistency of those deviations for these ETFs. Using mediation analysis, we separate the direct and indirect effect of AT on NAV deviation and persistency of NAV deviation. Using arbitrage to profit from these deviations, AT directly help in reducing these deviations and their persistency. We find that AT also improves liquidity which indirectly reduces deviation from NAVs and persistency of those deviations. 2

3 1. Introduction Exchange-traded funds (ETFs) have grown exponentially in size and trading volume over the last decade. ETFs have nearly $4.3 trillion in assets by the end of July 2017 (Forbes, August 23, 2017) 1. ETFs now account for about 30 per cent of all US trading by value, and 23 per cent by share volume, according to an article in Financial Times on January 24, ETFs have numerous advantages as they provide investors with increased access to asset classes and markets, improved tax efficiency, liquidity, price discovery, and transparency (Hill, Nadig, and Hougan, 2015). However, a growing concern is that ETFs may have unintended consequences such as increased volatility in the underlying stocks (Ben-David, Franzoni, and Moussawi, 2017), a reduction in the liquidity (Hamm, 2010) and informational efficiency (Israeli, Lee, and Shridharan, 2016), increased co-movement in returns (Da and Shive, 2012, Agarwal, Hanouna, Moussawi, Stahel, 2017), and increased market fragility (Bhattacharya and O Hara, 2017). The liquidity of ETFs also makes them an attractive asset for trading by high-frequency traders. The controversial role of high-frequency traders (a subset of algorithmic traders) in the equity markets has drawn the attention of regulators, researchers and market participants. This is critically important in light of recent developments in the financial markets characterized by increasing level of algorithmic trading (AT) or high frequency trading (HFT) in highly fragmented markets. Aldridge (2016) points out that high-frequency trading (HFT) and exchange-traded funds (ETFs) are among the most recognizable and hotly debated features of today s financial markets. Although several papers have shown empirical evidence that institutions holding ETFs have a significantly shorter horizon than those holding the underlying 1 Record Inflows Boost Global ETF Assets To $4.3 Trillion, With BlackRock Leading The Way, August 23, 2017 by Trefis Team, Forbes. 2 ETFs are eating the US stock market, January 24, 2017 by Robin Wigglesworth, Financial Times. 3

4 securities (Ben David, et al., 2017; Bronman and Shun, 2017), their analyses are based on quarterly holding data. Therefore, HFT or AT that trade multiple times within a shorter time interval are not captured in these studies. Specifically, the role of algorithmic traders in ETFs and whether their trades help improve liquidity and efficiency of these ETFs have not been explored. In this paper, we aim to shed light on this gap in the current literature. ETF market is unique as arbitrage can occur through creations and redemptions of shares in the primary market when ETF prices deviate from their NAVs. Pricing differences are small generally, but they can vary over time, across ETFs and can be persistent. Petajisto (2017) report an average of 6 basis points for difference between ETFs and their net asset values, but 49 basis points for the volatility of the difference. In a related study, Madhavan (2012) finds that the prices of the Exchange-traded products (ETPs) diverged widely from the value of underlying net asset values during a market crash on May 6, Thus, to prevent ETFs propagate risk into the underlying assets, it is important that ETF prices remain aligned with their fundamental values. The persistency of the deviations is another dimension of the pricing efficiency, and has been reported to last only a few minutes (Engle and Sarkar, 2006) to about five days (Fulkerson and Jordan, 2013). Another strand of literature that is related to our study is on how the liquidity and AT activity of ETFs impact the overall equity market. Calamia, Deville, and Riva (2016) find that ETF markets are highly liquid but occasionally breakdown under turbulent conditions. The spreads on European equity ETFs are determined by inventory-risk related variables. Although ETF prices are determined by the prices of the underlying securities, they sometimes change faster than the prices of some of the underlying securities. Aldridge (2016) develops a model to show how statistical arbitrage carried out by HFT can influence the prices of the 4

5 underlying securities. Apart from ETF market, HFT/AT has been shown to be associated with a reduction in trading costs (Jones, 2013; Harris, 2013) and an increase in price efficiency (Brogaard, Hendershott and Riordan, 2014; Chaboud, Chiquoine, Hjalmarsson and Vega, 2014). Thus, activity by AT and fast high-frequency traders in ETF should help in keeping the prices aligned with the underlying. An advantage in studying ETFs is that price efficiency can be measured directly by looking at the deviation in NAV from prices. We aim to study AT in the US ETF market and how the AT activity in the ETF market impacts NAV deviations and persistency of those deviations in ETFs. How pervasive is AT trading in ETFs? How does AT impact the efficiency of ETF pricings measured by the deviations between ETF prices and their net asset values, persistency of those deviations, and liquidity of these ETFs? Using the SEC Market Information Data Analytics System (MIDAS) data from 2012 to 2016, we explore algorithmic trading in Exchange-traded funds. Using cancel rate, trade-to-order volume ratio, and order fragmentation we find a high level of AT activity in ETFs. The average cancel rate is above 200, significantly higher than 25 reported for stocks in Jain, Jain, and Jiang (2017). The rich dataset from MIDAS also has order and trade data across ten different exchanges. On daily basis, trading of ETFs is highly fragmented as almost all exchanges receive orders and execute trades for all the equity ETFs in our sample. During our sample period, NYSE Arca and Nasdaq are the leading exchanges for trading ETFs. Following the literature, we focus on two measures of ETF pricing efficiency. According to law of one price, ETF price should be the same or very close to NAV. The deviation between price and NAV represents opportunities for arbitrage in both primary and secondary market. The larger the magnitude of the deviation, the less efficient is the ETF price. Another dimension of pricing efficiency that we consider is the persistency of the deviation which is how many days the discount or premium 5

6 lasts. The longer the deviation lasts, the lower the efficiency of the ETF pricing due to lack of arbitrage. Our sample shows a relatively small discount and premium, averaging about 2 basis points. Interestingly, although the average deviation for persistency is about 4.5 days, there is an asymmetric pattern between the persistency for discounts (3.9 days) versus for premiums (5.8 days). In the case of ETF premium, arbitrageurs can simultaneously sell the ETF and buy the underlying assets. The higher persistency for deviation may be due to higher cost of shorting for ETFs compared to cost of shorting for component stocks. Next, we examine the relation between AT and our pricing efficiency measures. We find that higher AT results in lower deviation from NAVs, and lower persistency of those deviations for these ETFs. This is consistent with recent empirical evidence on the positive impact of high-frequency traders and faster trading on market quality measures as short-term volatility, spreads, and displayed depth (Hasbrouck and Saar, 2010; Hendershott, Jones, and Menkveld (2011). We further perform mediation analysis to gain a better understanding of the channels through which AT could affect ETF pricing efficiency. Some ATs play the role of arbitrager where they take advantage of mispricing of ETFs with respect to the NAV. Another type of ATs play the role of market makers by posting ask and bid price while making profit based off a narrow spread. We test the effect of both these type of ATs on ETF price efficiency. We expect that AT who play the role of arbitragers should help in keeping the ETF prices aligned with NAV. Moreover AT who play the role of market maker keep the spreads low and thus should facilitate arbitragers' activity. We do not have a way to identify these two type of ATs separately, but as our AT proxy increases, we expect both type of AT activity to increase as they both rely on placing and canceling orders. We find that AT activity is positively related to price efficiency of ETFs suggesting arbitrager role of ATs in the ETF markets. We also find that AT indeed 6

7 reduces spreads suggesting market maker role of ATs. This further reduces the deviation from NAV and its persistency. A reduction in spread essentially alleviates the limits to arbitrage due to transaction costs. Around 50% of the effect of AT on deviation is because of arbitragers activity, while the other 50% is due to the market making role of AT. Our paper makes several contributions to the literature. First, we provide new insights into the relation between AT, arbitrage, and pricing efficiency of ETFs. Our study is closely related to Brown, Davies, Ringgenberg (2018) as they focus on observable arbitrage activity and its negative impact on subsequent returns. First, we characterize trading activity and AT in the ETF markets. We find that NYSE-Arca and Nasdaq are the main players for ETF trading. Second, we show that arbitrage by algorithmic trading on an intraday basis is an active force in the ETF market. AT helps to reduce the deviation between price and NAV and the persistency of the deviation. Second, using data on ETF, we collaborate the findings of the positive role of AT/HFT in reducing spread and improving market efficiency. 2. Data and descriptive statistics We download the data provided by the SEC on cancellations, trades, order volume, and trade volume on various U.S. exchanges. Our sample extends from January 2012 to December We download closing price, high price, low price, closing ask price, closing bid price, and shares outstanding from the Center for Research in Security Prices (CRSP). We download a list of U.S. equity ETFs from ETFdb.com and only keep those ETFs in our sample which are on that list. We also remove leveraged and inverse ETFs from our sample. Finally, we only keep those ETFs in our sample that trade on every day of our sample period and had at least one lit trade. 7

8 We also download the NAV data from Bloomberg and merge that with the SEC data. Following Petajisto (2017), we remove ETF-days with ETF premiums/discount greater than 20%. Following Broman (2016), when the premium based on midpoint prices is more than ten percentage points greater in absolute terms than the premium based on closing prices, we use the latter instead. We also remove the ETF XLF from our sample because it is an outlier. 3 The above steps result in a final sample of 160 ETFs. In Table 1, we provide mean and median values ETF fund characteristics in Panel A and AT in ETFs in Panel B. On each day, we take value weighted mean or median across ETFs to calculate daily values. Then, we report time series average of these daily values for our sample. Average market capitalization, calculated as price times shares outstanding, for our sample is 44 million dollars with a median of 17 million dollars. Mean values of turnover, calculated as trade volume divided by shares outstanding, is 2.90%. Mean value of volatility, calculated as difference between high price and low price divided by high price, is 1.03%. Absolute deviation from NAV is calculated as absolute difference between quote midpoint and NAV of ETF and divided by quote midpoint and multiplied by 100. For our sample, the mean value of this deviation is 2.26 basis points with a median of 1.91 basis points. Mean of raw values of this deviation is 0.19 basis points with a median of 0.10 basis points. On average a run of particular deviation persist for 4.5 days for our sample. We also report the deviation separately for ETFs traded at discount and ETFs traded at premium. The mean discount is 2.11 basis points that last for about 4 days on average, and the mean premium is 2.17 basis points which last for about 6 days on average. We also look at the persistency of premiums and discount from NAV in detail. We count the number of days in each run of deviation between ETF s NAV and price. In Figure 1, we 3 The mean deviation from NAV for XLF during our sample period is 17.74%. 8

9 report the percentage number of ETF-days for different length of time for which the deviation from NAV lasted in one run. For example, 8.30% (9.44%) of the observations had a premium (discount) only for 1 day as represented in the first black (grey) bar in figure 1. Most of these deviations last for less than 10 days. Only 19.06% of our sample observations has premium and discount that last for more than 10 days in one run % of the observation has deviations that lasted between 2 to 5 days in on run and 19.44% observations have deviation that lasted between 6 to 10 days in one run. We also plot the mean values of the premium or discount in each of these categories. We find that higher the deviation the longer it lasts. Mean discount is 2.50 basis point for when it lasts for one day, however it is 6.30 basis point when it lasts for more than 20 days. Similarly, mean premium is 2.41 basis point when it lasts for one day, however it is 5.78 basis point when it lasts for more than 20 days. [Figure 1 about here] Next, in Table 1, Panel B, we look at algorithmic trading in ETFs. First, we use cancel rate and trade to order volume %, as a proxy for AT following Jain, Jain, and Jiang (2017) and Weller (2017). Cancels is number of orders that are cancelled before execution. Lit trades is count of all trade messages for trades that are not against hidden orders. Cancel rate is calculated as cancels divided by lit trades. A higher cancel rate indicates higher algorithmic trading. Trade volume is number of shares traded. Order volume is volume of total orders placed. Trade to order volume % is calculated as trade volume divided by order volume times 100. A lower trade to order volume % indicates higher algorithmic trading. We find that the mean (median) cancel rate is 210 (110) and the average trade to order volume % is 0.51% (0.30%). 9

10 Next, we look at order fragmentation as another proxy for AT following Madhavan (2012). The Herfindahl index ranges from 0 to 1, with higher figures indicating less fragmentation in that particular ETF. Order fragmentation is one minus the Herfindahl Hirschman index of order volume across all the exchanges. Order fragmentation is , which is higher compared to one minus the quote Herfindahl of 0.35 reported by Madhavan (2012) for the baseline period of 20 days (April 7, 2010 to May 5, 2010). This suggests that markets are becoming increasingly fragmented over time. [Table 1 about here] 4. Results 4.1 AT activity in ETFs by Exchange In Table 2, we present cancels, trade, order, and AT variables by exchange. We also calculate the venue shares of the U.S. equity market based on reported trade volume and order volume. Market share by exchange for trades is defined as trade volume for the exchange divided by the total trade volume. Market share by exchange for orders is defined as order volume for the exchange divided by the total order volume. We find that, on average, NYSE Arca is the leading exchange for both trading and order placement for ETFs with 30.21% market share for trade volume and 25.38% market share for order volume. Nasdaq is the next leading exchange for both trade share and order share with values at 22.51% and 19.1%. [Table 2 about here] We also look at the total number of times each exchange dominated for a given ETF and plotted the results in figure 2. For 200,991 ETF-days in our sample, we find that Arca is the 10

11 dominating exchange 64.35% of the time, followed by NASDAQ which dominated for 15.59% of those ETF-days. The next two exchanges are Edge-X and BATS-Z, which dominate 8.14% and 7.25% of the times, respectively. Boston, CHX, NSX, and PHLX exchanges each dominated for less than 1% of those ETF-days. 4 [Figure 2 about here] 4.2 AT and fragmentation in ETFs In Table 3, we report the value weighted means, cancel rate, trade to order volume %, order fragmentation, ETF s absolute deviation from NAV, and persistency of those deviations by year and by market characteristics. In Panel A, we find that the cancel rates have increased every year since 2013 indicating higher AT over years. The trade to order volume % varies from 0.41% to 0.61% over these years. Markets have become more and more fragmented over the years. Order fragmentation has gone up from in 2012 to in The absolute deviation of NAV has declined every year from 2.84 basis points in 2012 to 1.81 basis points in We do not see any pattern for the persistency of those deviations by year. In Panel B, we present these variables for different market return percentiles. We do not see any pattern for cancel rate in this Panel. Trade to order volume % is higher on extreme return days, indicating lower algorithmic trading on those days. This is consistent with Brogaard, Carrion, Moyaert, Riordan, Shkilko, Sokolov (2017) finding that when several stocks experience simultaneous extreme price movements, high frequency trades demand liquidity instead of supplying it. We find the markets to be more fragmented both during extreme positive and negative return days. Absolute deviation from NAV is higher on days of extreme returns but persistency of those deviations is somewhat lower. In Panel C, we present these numbers by market volatility using VIX. Cancel rates are 4 In untabulated results, we also look at how many exchanges an ETF is traded on a given day. We find that on average all the ETFs are traded on at least 8 exchanges on a given day. 11

12 higher for higher market volatility rate indicating higher AT on those days. Order fragmentation is increasing with increase in market volatility. This indicates that markets are more fragmented on days of higher volatility. Absolute deviation from NAV is higher for higher market volatility days. We do not see a clear pattern for trade to order volume % and persistency of those deviations in this Panel. [Table 3 about here] In Table 4, we present these numbers by ETF characteristics such as market capitalization, volatility, turnover, age, and sponsor to see if AT is correlated with these variables. We divide the sample in quartile to create ranks of market capitalization, volatility, and turnover for Panels A to C. In Table 4 Panel A, we present these variables by market capitalization ranks of ETFs. Cancel rate (trade to order volume %) of ETFs with higher market capitalization is lower (higher) indicating lower algorithmic trading for larger ETFs. ETFs with higher market capitalization are more fragmented as indicated by higher order fragmentation. ETFs with lower market capitalization have higher deviation from NAV. These deviations also persist longer for smaller firms. In Panel B, we present these variables by volatility rank of ETFs. We find that trade to order volume % and order fragmentation are generally higher for ETFs with higher volatility. ETFs with higher volatility have higher deviation from NAV. We do not find any clear pattern for cancel rate and persistency of deviation in this Panel. In Panel C, we present these variables by turnover rank of ETFs. ETFs with higher turnover have lower cancel rate, higher trade to order volume %, higher fragmentation, and lower persistency of those deviations. We do not find any clear pattern for deviation from NAV in this Panel. 12

13 In Panel D, we present these variables by ETF s age. Older ETFs have lower cancel rate, higher trade to order volume %, and higher fragmentation. Deviation from NAV is lower for older ETFs, it is 2.04 basis points for the ETFs which have been around for 16 to 20 years. Whereas the deviation it is 3.7 basis points for the ETFs which were launched in the last 5 years. Persistency of these deviations is lower for older ETFs. In Panel E, we present these variables by ETF s sponsor. We see highest algorithmic trading (indicated by higher cancel rate and lower trader to order volume %) for one ETF sponsored by Fidelity and one ETF sponsored by Oppenheimer Funds followed by five ETFs sponsored by Wisdom Tree. We see highest deviation from NAV for the 1 ETFs sponsored by Fidelity at basis point followed by deviation of 6.49 basis point for the one ETF sponsored by Alerian. This is based on the sample after we removed the outlier ETF XLF. The deviation lasted longer for ETFs sponsored by First Trust, Oppenheimer Funds, and Wisdom Tree. [Table 4 about here] 4.3 Effect of AT on absolute deviation from NAV, persistency of those deviations, and spreads In Table 5, we run a regression of absolute deviation from NAV and persistency of deviation on AT and control variables. Following Jain, Jain, and Jiang (2017) we define AT as follows. We sort all stock-exchange days by trade volume and keep the top 50% to eliminate stock-exchange days with low volume. We further sort these stock-exchange days by cancel rate (trader to order volume %) to create quartiles. We then define a dummy variable AT for each stock day, which takes the value 1 if a stock falls in the highest cancel rate quartile (lowest trader to order volume% quartile) on any exchange. Results for AT variable using cancel rate are in Models 1 and 2 and results for AT variable using trade to order volume % are in Models 3 and 4. We report standardized coefficients (standardized variables have zero mean and unit variance) 13

14 with t-statistics based on White s heteroscedasticity standard errors for all regression models in this paper. Negative coefficients on AT in all models indicate that ETFs with higher algorithmic trading have lower deviation from NAV and lower persistency of those deviation. Thus, they help in bringing ETF prices more in line with the underlying stocks. Positive coefficients on spread indicates that ETFs that are less liquid have higher deviation from NAV and higher persistency of those deviations. Negative coefficients on log market capitalization in Models 3 and 4 indicate that persistency of deviation is lower for larger ETFs. Positive coefficient on volatility in Models 1 and 2 indicate that higher volatility causes deviation from NAV to be higher. However, negative coefficient on volatility in Models 3 and 4 indicate that as prices fluctuate a lot a particular run of premium or discount may not last longer as it may keep switching from premium to discount and vice versa. Negative coefficient on age in all models indicates that deviation from NAV and persistency of those deviations are lower for older ETFs. [Table 5 about here] Next, we perform a mediation analysis to look at the direct and indirect effect of AT on deviation from NAV and its persistency. NAV deviation provides an opportunity to traders to profit through arbitrage. Traders can buy ETFs trading at discount and short the underlying securities. Conversely, they can short ETFs trading at premium and buy the underlying securities. AT can reduce the NAV deviation directly through this channel of statistical arbitrage. For example, if an ETF has a NAV of $20.0, but is trading at $19.9, AT can buy the ETF and short the underlying securities. High spreads can deter this activity of statistical arbitrage. For example, if an ETF has a NAV of $20.00, and has a quote of $19.9-$20.0 (and a quote midpoint of $19.95), then it might be difficult to profit from arbitrage as a trader demanding liquidity would have to place a marketable limit order (or market order) to buy it at $20.0. If the spread 14

15 narrows, and the quote changes to $19.94-$19.96, a trader would be immediately able to buy the ETF at $19.96 and short the underlying securities and book a profit. Thus, a lower spread can make it easier to take advantage of the arbitrage opportunity. We report the results of this analysis in Table 6. Please note that the numbers reported in column 3 of Panel A, B, C, and D, are same as those reports in Models 1, 2, 3, and 4 of Table 5, respectively. We use the same control variables (i.e. log market capitalization, volatility, and age) in this table for all regression models. Sign and significance of the control variables in all models are same as those reported in Table 5. However, we do not report the coefficients on those control variables for brevity. In Panel A, we show the mediation analysis for effect of AT calculated using cancel rate on deviation from NAV. In Model 1, we exclude spread and Model 2, we include spread. In Model 3, we run the regression of spread on AT. We find that coefficient on AT is significant in Model 1 and Model 3, indicating AT reduces the deviation from NAV and spread. In Model 2, we show that when we include spread as a control variable, the coefficient of AT remains significant, however the magnitude of that coefficient goes down. These results indicate that AT reduces deviation from NAV directly as well as indirectly through reduced spreads. In next 3 columns, we show indirect, direct, and total effect of AT on deviation from NAV. Then, we also show that proportion of total effect mediated by spread in this Panel is 48%. Ratio of indirect to direct effect is percent. Next, we also run Sobel test to confirm the mediation effect and we find the p value to be significant at 1 percent level. 5 We perform the similar mediation analysis for Models 2, 3, and 4 of Table 5 and report the results in Panels B, C, and D of Table 6, respectively. Overall, we find that AT reduces deviation from NAV and its persistency directly as well as indirectly through reduced spreads. 5 For Sobel test we use 1-AT instead of AT as an independent variable because AT and spread have coefficients of opposite sign. 15

16 Mediation effects are higher for deviation from NAV compared to persistency of those deviation. When we use AT calculated using cancel rate (trade to order volume %), mediation effect is 48% (52%) but only 9.35% (16.26%) for persistency of those deviations. Nonetheless, all these mediation effects are statistically significant at 1 percent level. [Table 6 about here] Next, we use the actual cancel rate and trade to order volume % variables as independent variables in Table 7. However, as we see in figure 2, that most of the trading for ETFs happens on Arca, Bats-Z, Edge-X, and Nasdaq, we use the volume weighted cancel rate and trade to order volume % from these 4 exchanges. We report the results using cancel rate measure in Models 1 and 2. Higher cancel rate indicates higher algorithmic trading. The negative coefficient on cancel rate indicates that absolute deviation from NAV and persistency of those deviations are both lower when algorithmic trading is higher. We report the results using trade to order volume % measure in Models 3 and 4. Higher trade to order volume % indicates lower algorithmic trading. The positive coefficient on trade to order volume % indicates that absolute deviation from NAV and persistency of those deviations are both lower when algorithmic trading is higher. Sign and significance of spread, log market capitalization, volatility, and age are same as those reported in Table 5. [Table 7 about here] Next, we look at order fragmentation as another proxy for AT following Madhavan (2012). We report the results of this measure in Table 8. Higher order fragmentation indicates higher algorithmic trading. Negative coefficient on order fragmentation indicated that absolute deviation from NAV and persistency of deviation are both lower for ETFs with higher 16

17 algorithmic trading. Sign and significance of other control variables are same as those reported in Tables 5 and 7. [Table 8 about here] Overall, using all 5 proxies of algorithmic trading in Tables 5, 7, and 8, we find that higher algorithmic trading makes ETF prices more efficient by reducing the deviation from NAV and persistency of those deviations. 5. Conclusion Using MIDAS data compiled by the SEC from 2012 to 2016, we explore AT in ETFs. Using cancel rate, trade-to-order volume ratio, and order fragmentation as proxies for AT, we find AT activity in ETFs to be significantly higher than in stocks. Jain, Jain, and Jiang (2017) find cancel rate and trade to order volume % to be 24.8 and 2.94% for stocks, respectively. We find that ETFs have a cancel rate of 210 and trade to order volume% of 0.51%, respectively. We also find that order volume in ETFs have become more fragmented over the years. NYSE Arca and Nasdaq are the top two dominating exchanges for ETF trading, followed by Edge-X and BATS-Z. We explore if AT helps in keeping the prices of ETFs aligned with the underlying securities as they use statistical arbitrage strategies to profit. Using 2 cancel rate measures, 2 trade to order volume % measures, and order fragmentation measure as a proxy of AT, we find that AT results in lower deviation from NAVs and lower persistency of deviations for these ETFs. We also perform mediation analysis and find that AT reduces deviation from NAV and its persistency directly, as well as indirectly through reduced spreads. 17

18 References: Agarwal, Vikal, Hanouna, Paul, Moussawi, Rabih, and Christof Stahel Do ETFs increase the commonality in liquidity of underlying stocks?. Working paper. Aldridge, Irene ETFs, high-frequency trading, and flash crashes. The Journal of Portfolio Management, volume 43 (1), Ben-David Itzhak, Franzoni, Francesco, and Rabih Moussawi Do ETFs increase volatility? Working paper. Bhattacharya, Ayan and Maureen O Hara Can ETFs increase market fragility? Effects of information linkages in ETF markets. Working paper. Brogaard, J., A. Carrion, T. Moyaert, R. Riordan, A. Shkilko, and K. Sokolov. High Frequency Trading and Extreme Price Movements. Journal of Financial Economics, forthcoming. Broman, Markus Liquidity, style investing and excess comovement of exchange-traded fund returns. Journal of Financial Markets, volume 30, Calamia, Anna, Deville, Laurent, and Fabrice Riva The provision of liquidity in ETFs: Theory and evidence from European markets. Working paper. Da, Zhi and Sophie Shive Exchange traded funds and asset return correlations. Working paper. Dannhauser, Caitlin D., The impact of innovation: Evidence from corporate bond ETFs. Journal of Financial Economics, volume 125, Engle, Robert, and Debojyoti Sarkar Premiums-discounts and Exchange-traded funds. Journal of Derivatives, volume 13(4), Glosten, Lawrence, Nallareddy, Suresh, and Yuan Zou ETF activity and informational efficiency of underlying securities. Working paper. Hamm, Sophia J.W The effect of ETFs on stock liquidity. Working paper. Hill, Joanne M., Nadig, Dave, and Matt Hougan A comprehensive guide to understanding Exchange-traded funds (ETFs). CFA Institute Research Foundation. Israeli, Doron, Lee, Charles M.C., and Suhas A. Sridharan. Is there a dark side to exchangetraded funds? An information perspective. Review of Accounting studies, forthcoming. Jain, Archana, Jain, Chinmay, and Christine Jiang, Algorithmic trading and fragmentation. Journal of Trading, volume 12, number 4. 18

19 Madhavan, Ananth Exchange-traded funds, market structure, and the flash crash. Financial Analyst Journal, volume 68, number 4, Madhavan, Ananth and Aleksander Sobczyk Price dynamics and liquidity of Exchangetraded funds. Journal of Investment Management, volume 2, Marshall, Ben R., Nhut H. Nguyen, and Nuttawat Visaltanachoti ETF Arbitrage. Working paper, Massey University (November). Petajisto, Antti Inefficiencies in the pricing of exchange-traded funds. Financial Analyst Journal, volume 73, number 1, Weller, Brian Does algorithmic trading reduce information acquisition? Review of Financial Studies, Forthcoming. 19

20 Appendix A Variable definitions: Market capitalization = Market cap is calculated as price times shares outstanding. Turnover = Turnover is calculated as trade volume divided by shares outstanding. Volatility = Volatility is calculated as difference between high price and low price divided by high price. Absolute deviation from NAV = Absolute deviation from NAV is calculated as absolute difference between quote midpoint and NAV of ETF and divided by quote midpoint and multiplied by 100. Raw deviation from NAV = Deviation from NAV is calculated as absolute difference between quote midpoint and NAV of ETF and divided by quote midpoint and multiplied by 100. Persistency of deviation = Number of days in each run of deviation between ETF s NAV and price. Discount from NAV = Deviation from NAV for ETFs traded on discount. Persistency of discount = Number of days in each run of when ETF traded at a discount. Premium from NAV = Deviation from NAV for ETFs traded on premium. Persistency of premium = Number of days in each run of when ETF traded at a premium. Cancels = Cancels is number of orders that are cancelled before execution. Lit Trades = Lit trades is count of all trade messages for trades that are not against hidden orders. Cancel rate = Cancel rate is calculated as cancels divided by lit trades. A higher cancel rate indicates higher algorithmic trading. Trade volume = Trade volume is number of shares traded. Order volume = Order volume is volume of total orders placed. Trade to order volume % = Trade to order volume % is calculated as trade volume divided by order volume times 100. A lower trade to order volume % indicates higher algorithmic trading. Order fragmentation = Order fragmentation is one minus the Herfindahl Hirschman index of order volume across all the exchanges. The Herfindahl index ranges from 0 to 1, with higher figures indicating less fragmentation in that particular ETF. We create the fragmentation index and one minus the Herfindahl index for clear interpretation. AT = Following Jain, Jain, and Jiang (2017) we define AT as follows. We sort all stock-exchange days by trade volume and keep the top 50% to eliminate stock-exchange days with low volume. We further sort these stock-exchange days by cancel rate (or trade to order volume %) to create quartiles. We then define a dummy variable AT for each stock day, which takes the value 1 if a stock falls in the highest cancel rate (lowest trade to order volume %) quartile on any exchange. Spread = Spread is calculated as ask minus bid divided by price. Volatility = Volatility is calculated as highest ask lowest bid divided by highest ask. 20

21 Age = Age is calculated as calendar year minus the year of ETF inception. Data and Sample: We download the data provided by the SEC on cancellations, trades, order volume, and trade volume on various U.S. exchanges. Our sample period is from January 2012 to December We download closing price, high price, low price, closing ask price, closing bid price, and shares outstanding from the Center for Research in Security Prices (CRSP). We download a list of U.S. equity ETFs from ETFdb.com and only keep those ETFs in our sample. We also remove leveraged and inverse ETFs from our sample. Finally, we only keep those ETFs in our sample that trade on every day of our sample period and had at least one lit trade. We also download the NAV data from Bloomberg and merge that with the SEC data. We remove ETF-days with ETF premiums/discount greater than 20%. We also remove the ETF XLF because it is an outlier in terms of premium. The above steps result in a final sample of 160 ETFs. 21

22 Table 1 Descriptive statistics In this table, we report the descriptive statistics of ETF fund characteristics and algorithmic trading measures. Please refer to Appendix A for variable definitions and sample details. Variable Mean Median Panel A: ETF Fund characteristics Market capitalizationi,d 44,812,486 17,528,597 Turnoveri,d 2.90% 0.76% Volatilityi,d 1.03% 0.93% Absolute deviation from NAVi,d % % Raw deviation from NAVi,d % % Persistency of deviation Discount from NAVi,d % % Persistency of discount Premium from NAVi,d % % Persistency of premium Panel B: Algorithmic trading Cancelsi,d 2,055, ,604 Lit Tradesi,d 60,120 7,313 Cancel ratei,d Trade volumei,d 16,013 1,274 Order volumei,d 2,072, ,270 Trade to order volume %i,d 0.51% 0.30% Order fragmentationi,d

23 Table 2 Descriptive statistics of AT in ETF by exchange In this table, we report the algorithmic trading variables by exchange along with market share of each exchange based on trades and orders. Market share by exchange for trades is calculated as trade volume for the exchange divided by the total trade volume. Market share by exchange for orders is calculated as order volume for the exchange divided by the total order volume. Please refer to Appendix A for other variable definitions and sample details. Cancels Lit Trades Cancel rate Trade volume Order volume Trade to order vol % Market share by exchange for trades Market share by exchange for orders Arca 401,584 14, , , % 30.21% 25.38% Bats-Y 169,084 3, , % 4.55% 5.76% Bats-Z 364,728 11, , , % 17.85% 15.22% Boston 146,074 3, , % 3.75% 5.01% CHX 132, , % 1.69% 13.29% Edge-A 149,727 3, , % 5.33% 5.96% Edge-X 203,909 6, , , % 12.45% 9.12% NSX 38, , % 0.81% 3.55% Nasdaq 428,978 16, , , % 22.51% 19.10% Phlx 119,333 1, , % 2.30% 5.92% 23

24 Table 3 AT over time and by market characteristics In this table, we report the value weighted mean values of AT variables and deviation from NAV and its persistency by year and by market characteristics variables. Please refer to Appendix A for variable definitions and sample details. Cancel rate Trade to order volume % Order fragmentation Absolute deviation from NAV Persistency of deviation Panel A: By year % % % % % % % % % % 4.2 Panel B: By market (i.e. S&P500) return category Bottom 1 percentile % % 4.3 Bottom 5 percentile % % 4.3 Bottom 10 percentile % % 4.4 Bottom 50 percentile % % 4.5 Top 50 percentile % % 4.5 Top 10 percentile % % 4.2 Top 5 percentile % % 4.1 Top 1 percentile % % 4.1 Panel C: By market volatility (i.e. Adjusted closing price of ^VIX) category Bottom 1 percentile % % 4.4 Bottom 5 percentile % % 4.4 Bottom 10 percentile % % 4.7 Bottom 50 percentile % % 4.8 Top 50 percentile % % 4.2 Top 10 percentile % % 3.9 Top 5 percentile % % 4.0 Top 1 percentile % %

25 Table 4 AT by ETF characteristics In this table, we report the value weighted mean values of AT variables and deviation from NAV and its persistency by ETF characteristic variables. Please refer to Appendix A for variable definitions and sample details. Number of ETFs Mean value of rank variable Cancel rate Trade to order volume % Order fragmentation Absolute deviation from NAV Persistency of deviation Panel A: By ETF market capitalization rank Lowest , % % , % % ,876, % % 6.2 Highest 40 53,985, % % 3.9 Panel B: By ETF volatility rank Lowest % % % % % % % % % 4.8 Highest % % % 4.6 Panel C: By ETF turnover rank Lowest % % % % % % % % % 4.9 Highest % % % 3.5 Table 4 continued 25

26 Table 4 continued Number of ETFs Mean value of rank variable Cancel rate Trade to order volume % Order fragmentation Absolute deviation from NAV Persistency of deviation Panel D: By ETF age 0 to 5 years % % to 10 years % % to 15 years % % to 20 years % % 3.5 Panel E: By ETF sponsor Alerian % % 3.9 Charles Schwab % % 7.7 Fidelity % % 8.1 FirstTrust % % 13.0 Guggenheim % % 7.7 Invesco % % 5.3 Oppenheimer Funds % % 17.6 SPDR % % 3.5 Vanguard % % 5.6 Wisdom Tree % % 12.4 ishares % %

27 Table 5 Absolute deviation from NAV and Persistency of deviation regression on AT dummy variables In this table, we report the results of regressing absolute deviation from NAV and persistency of deviation AT variables and other firm characteristic variables. In Models 1 and 2, we use cancel rate for measuring AT dummy. In Models 3 and 4, we use trade to order volume % for measuring AT dummy. Please refer to Appendix A for variable definitions and sample details. All coefficients are standardized (standardized variables have zero mean and unit variance). *** denotes statistical significance at 1% level. t-statistics are based on White s heteroscedasticity consistent standard errors. Variable Deviation from NAV Cancel rate Trade to order volume % Persistency of deviation Deviation from NAV Persistency of deviation (1) (2) (3) (4) Intercept *** *** *** *** (4.11) (60.51) (4.13) (59.09) AT i,d *** *** *** *** (-8.29) (-17.74) (-6.63) (-8.93) Spread i,d *** *** *** *** (9.08) (5.56) (9.15) (5.61) Log (Market capitalization) i,d *** *** (0.61) (-38.21) (0.63) (-38.29) Volatility i,d *** *** *** *** (5.11) (-15.45) (4.97) (-15.43) Age i,y *** *** *** *** (-25.41) (-60.85) (-25.21) (-60.77) Adjusted R Square Number of Observations 200, , , ,037 27

28 Table 6 Mediation analysis for table 5: direct and indirect effect (through reduced spread) of AT on Absolute deviation of NAV and Persistency of Deviation. In this table, we report the mediation analysis to test the direct and indirect effect (through reduced spread) of AT on Absolute deviation of NAV and Persistency of Deviation. In Panel A, we report the mediation analysis for Model 1 of Table 5. In Panels B, C, and D, we report the same analysis for models 2, 3, and 4 of table 5, respectively. All coefficients in Models 1 to 3 of this table are standardized (standardized variables have zero mean and unit variance). t-statistics are based on White s heteroscedasticity consistent standard errors. *** denotes statistical significance at 1% level. 28 Indirect effect of AT mediated by spread Direct effect of AT Total effect (direct + indirect) of AT Proportion of total effect mediated Ratio of indirect to direct Sobel test significance level (1) (2) (3) Panel A: Cancel rate Deviation from NAV Spread AT i,d *** *** *** *** (-25.32) (-8.29) (-48.24) Spread i,d *** (9.08) Control variables Yes Yes Yes Persistency of deviation Spread *** AT i,d *** *** *** (-20.64) (-17.74) (-48.24) Spread i,d *** (5.56) Control variables Yes Yes Yes Panel B: Trade to order volume % Deviation from NAV Spread AT i,d *** *** *** *** (-21.53) (-6.63) (-42.72) Spread i,d *** (9.15) Control variables Yes Yes Yes Persistency of deviation Spread AT i,d *** *** *** *** (-11.22) (-8.93) (-42.72) Spread i,d *** (5.61) Control variables Yes Yes Yes

29 Table 7 Absolute deviation from NAV and Persistency of deviation regression on AT volume weighted by exchange In this table, we report the results of regressing absolute deviation from NAV and persistency of deviation AT variables and other firm characteristic variables. In Models 1 and 2, we use cancel rate as a proxy AT. In Models 2 and 3, we use trade to order volume % as a proxy for AT. Please refer to Appendix A for variable definitions and sample details. All coefficients are standardized (standardized variables have zero mean and unit variance). *** denotes statistical significance at 1% level. t-statistics are based on White s heteroscedasticity consistent standard errors. Variable Cancel rate Trade to order volume % Deviation from Persistency of Deviation Persistency of NAV deviation from NAV deviation (1) (2) (3) (4) Intercept *** *** *** *** (4.73) (59.03) (4.25) (58.09) Cancel ratei,d ** *** (-2.15) (-10.81) Trade to order volume %i,d *** *** (10.51) (6.16) Spreadi,d *** *** *** *** (9.36) (5.67) (9.11) (5.64) Log (Market capitalization)i,d *** *** (-0.42) (-39.69) (-0.11) (-37.94) Volatilityi,d *** *** *** *** (4.86) (-15.45) (4.68) (-15.39) Agei,y *** *** *** *** (-25.16) (-58.44) (-25.57) (-60.25) Adjusted R Square Number of Observations 200, , , ,027 29

30 Table 8 Absolute deviation from NAV and Persistency of deviation regression on order fragmentation In this table, we report the results of regressing absolute deviation from NAV and persistency of deviation order fragmentation, a proxy for AT, and other firm characteristic variables. Please refer to Appendix A for variable definitions and sample details. All coefficients are standardized (standardized variables have zero mean and unit variance). *** denotes statistical significance at 1% level. t-statistics are based on White s heteroscedasticity consistent standard errors. Variable Deviation from NAV Persistency of deviation Intercept *** *** (4.90) (62.10) Order fragmentationi,d *** *** (-12.06) (-27.87) Spreadi,d *** *** (9.27) (5.67) Log (Market capitalization)i,d *** (0.77) (-34.11) Volatilityi,d *** *** (5.77) (-14.91) Agei,y *** *** (-25.08) (-57.27) Adjusted R Square Number of Observations 200, ,037 30

31 10.00% 9.00% 8.00% 7.00% 6.00% 5.00% 4.00% 3.00% Discount Premium Mean premium Mean discount 2.00% 1.00% 0.00% >20 Figure 1: Persistency of premiums and discounts from NAV We count the number of days in each run of deviation between ETF s NAV and price. We report the percentage number of ETF-days for different length of time for which the deviation from NAV lasted in one run. For example, 8.30% (9.44%) of the observations had ETF trading at a premium (discount) only for 1 day as represented in the first black (grey) bar in this figure. We also plot the mean values of the discount or premium in each of those categories in the line chart. 31

32 Percentage number of ETF-days for which exchange is dominating exchange 70% 60% 50% 40% 30% 20% 10% 0% Arca Bats-Y Bats-Z Boston CHX Edge-A Edge-X NSX Nasdaq Phlx Figure 2: Frequency distribution of dominating exchange We find the dominating exchange for each ETF day. We report the percentage number of days for which each exchange was the dominating exchange for our sample. 32