Execution Quality in Open Outcry Futures Markets Alexander Kurov May 2004 Abstract This study examines order flow composition and execution quality for different types of customer orders in six futures pits of the Chicago Mercantile Exchange (CME). We show that off-exchange customers frequently provide liquidity to other traders by submitting limit orders. We also document effective execution costs and turnaround times for customer liquiditydemanding and limit orders. The results imply that negative execution costs reported by Ferguson and Mann (2001) are likely to be explained by customer limit orders earning the bidask spread. We also find that customer limit orders do not appear to be systematically picked off by market makers. Finally, we show that common estimators of the bid-ask spread can be used as a proxy for the cost of immediacy faced by off-exchange customers. School of Management, Binghamton University (SUNY), Binghamton, New York 13902-6000. Tel: 607-777-5199. Fax: 607-777-4422. E-Mail: akurov@binghamton.edu. I thank Grigori Erenburg, Peter Locke, Tatyana Zabotina and especially Dennis Lasser for helpful comments and suggestions. Any remaining errors are my own.
1 Execution Quality in Open Outcry Futures Markets Abstract This study examines order flow composition and execution quality for different types of customer orders in six futures pits of the Chicago Mercantile Exchange (CME). We show that off-exchange customers frequently provide liquidity to other traders by submitting limit orders. We also document effective execution costs and turnaround times for customer liquiditydemanding and limit orders. The results imply that negative execution costs reported by Ferguson and Mann (2001) are likely to be explained by customer limit orders earning the bidask spread. We also find that customer limit orders do not appear to be systematically picked off by market makers. Finally, we show that common estimators of the bid-ask spread can be used as a proxy for the cost of immediacy faced by off-exchange customers.
2 I. Introduction Execution costs affect the returns of portfolio investors and profitability of trading strategies employed by active traders. Therefore, it is important for traders to be able to compare trading costs in alternative market venues. The market regulators also need to know what the transaction costs are in order to evaluate relative benefits of different trading protocols and examine the effects of rule changes on market liquidity. While execution costs in equity markets have been extensively studied, transaction costs in open outcry futures markets remain a relatively unexplored area because of limited availability of appropriate trade and quote data. Common time series estimators of bid-ask spreads, such as Roll (1984) and Smith and Whaley (1994), do not account for inter-customer trades and other microstructure effects in futures markets. Therefore, Locke and Venkatesh (1997) and Ferguson and Mann (2001) advocate using so-called execution spreads to measure effective trading costs. Execution spreads are estimated using trade data by aggregating across all customer trades during an interval. 1 Ferguson and Mann (2001) show that, surprisingly, the execution spreads in futures markets are negative in many intraday intervals, although they are positive on average. They argue that this finding may suggest that exchange locals, who tend to behave as market makers in the futures pits, often employ more complex strategies than simply buying at the bid and selling at the offer. Specifically, the locals offer customers price improvements in order to take positions ahead of anticipated price changes. Manaster and Mann (1999) use a similar argument to explain why off-exchange customers are frequently able to buy low and sell high. This study extends the prior literature by analyzing execution costs separately for liquidity-demanding and limit orders submitted by off-exchange customers in six futures pits of 1 Specifically, the execution spread is calculated as average customer buy price minus average customer sell price in a five-minute interval.
3 the Chicago Mercantile Exchange (CME). We show that, although off-floor customers tend to use liquidity-demanding orders, limit orders account for a substantial proportion of the customer order flow. Customers using limit orders compete with market makers to provide liquidity to incoming market orders or even provide liquidity to market makers. Calculating execution spread across all customer orders, an approach used by previous studies, produces an estimate of the aggregate customer transaction costs. While this measure may be useful in policy analysis, it does not tell us what the trading costs are for market orders or how much the customers can save by using limit orders. Therefore, in addition to the analysis of the aggregate customer trading costs, it is important to separately examine execution costs for customer market and limit orders. 2 The results of our analysis of customer execution costs show that the execution spreads for customer market orders and other liquidity-demanding orders are almost always positive. At the same time, execution spreads for customer limit orders tend to be negative, suggesting that limit order traders earn the bid-ask spread. Periods when executed customer limit orders dominate customer market orders are likely to produce negative aggregate execution spreads documented by Ferguson and Mann (2001). Therefore, our results explain why off-exchange customers are often able to buy low and sell high. In further analysis, we examine the notion suggested by Manaster and Mann (1999) and Ferguson and Mann (2001) that exchange locals offer customers better execution to take positions ahead of favorable price movements. In other words, customer limit orders may be picked off by market makers. We look at returns around customer limit order trades with market makers. The results show no systematic price movement away from executed limit 2 Studies of transaction costs and execution quality in equity markets (e.g., Harris and Hasbrouck (1996), Werner (2003), Battalio, Hatch and Jennings (2003), Bacidore, Ross and Sofianos (2003)) separately examine market and limit orders.
4 orders. This finding seems to contradict the notion that floor traders are able to systematically pick off limit orders submitted by off-exchange customers. We also attempt to reconcile the common estimators of effective spreads with the execution spread measure. Locke and Venkatesh (1997) and Ferguson and Mann (2001) conclude that common spread estimators are unrelated to the customer trading costs in futures markets. While these studies focus on the aggregate customer trading costs, we examine whether the common estimators are able to capture trading costs for liquidity-demanding orders. This approach is consistent with Demsetz (1968), who notes that the bid-ask spread represents the cost of trading without delay. Similarly, Grossman and Miller (1988) argue that the cost of trading immediately is the essence of market liquidity. Stoll (2000) suggests that measures of trading friction should capture price concessions paid by liquidity-demanding traders. Therefore, common spread estimators may be useful if they proxy for the price of immediacy paid by offfloor customers. We find that common spread estimators are closely related to the execution costs incurred by customer market orders in futures markets. The remainder of the paper is structured as follows. Section II below describes the data used in the study, examines distribution of customer orders by order type and type of counterparty, and documents turnaround times for different types of customer orders. Section III estimates execution spreads for different types of customer orders, examines whether market makers are able to pick off customer limit orders and evaluates the performance of common spread estimators. Section IV offers a summary and conclusions.
5 II. Data and Descriptive Statistics A. Data We use time and sales and computerized trade reconstruction (CTR) data for six futures contracts traded on the floor of the Chicago Mercantile Exchange (CME). These contracts include two index futures (S&P 500 and Nasdaq-100), two currency futures (Euro and Japanese Yen) and two agricultural futures (live cattle and lean hogs). The data are obtained from the Commodity Futures Trading Commission (CFTC). CTR data contain separate records for both sides of each trade. Available CTR data include the contract ticker symbol, trade date, trade time to the nearest second, the contract month, buy/sell code, number of contracts traded, trade price, customer type indicator (CTI), CTI of the opposite side of the trade, trade type (regular or spread trade), order type, and order arrival timestamp (when the order is received on the floor). 3 CTI is assigned to four different types of traders as follows: CTI1 are trades executed for a personal account of a floor trader (local trade), CTI2 are trades executed for an account of a clearing firm, CTI3 are trades executed for a personal account of another floor trader, CTI4 are trades executed for an account of an off-exchange customer. The order type is available for customer trades but it is not available for local trades. We examine the one-year period from July 1, 2000 to June 30, 2001. Days with more than one hour of data missing, such as shortened pre-holiday days, are removed from the sample. For every trading day, only the most actively traded contract is considered in each market. We remove calendar spread trades from the analysis, because they are priced at a differential between the legs of the spread and the reported prices of individual legs can differ significantly from contemporaneous trade prices for the respective contract maturities. 3 The exact execution time is not recorded by the traders and is estimated by the CTR algorithm based on time and sales data, 15-min time brackets and trade sequence on trader cards, order arrival timestamps and other available data. While execution times are not error-free, they are used by the exchange for enforcement purposes.
6 B. Distribution of customer trading volume by order type and type of counterparty Trader s choice between market and limit orders remains an important issue in the market microstructure literature. 4 Market orders and marketable limit orders are thought of as liquiditydemanding order types (e.g., Harris (2003), Werner (2003)). Market orders offer certain execution but the execution price is uncertain. Furthermore, market order traders are likely to pay a liquidity premium by buying at the offer and selling at the bid. On the other hand, limit order submitters are patient traders willing to accept the risk of not getting a fill to get their orders executed at a better price. Standing limit orders provide liquidity by creating trading opportunities for market order traders. Limit order traders tend to earn the bid-ask spread. The composition of customer order flow in open outcry futures markets, and particularly the role of limit orders, remains unexplored in the market microstructure literature. 5 Therefore, we begin by looking at the distribution of customer trading volume by type of order and type of counterparty. This allows us to achieve two goals. First, we can document relative prevalence of liquidity-demanding and limit customer orders. Second, we can examine the degree to which market makers provide liquidity to market order traders or consume liquidity supplied by customers using limit orders. The distribution of customer trading volume by order type is shown in Table I. [Insert Table I about here] For five order types, including market, discretionary, stop, market-on-close, and marketable limit orders, brokers acting on behalf of customers trade against the standing quotes. 6 Therefore, we classify all of these order types as liquidity-demanding. In contrast, customer limit 4 See, for example, Harris and Hasbrouck (1996), Handa and Schwartz (1996) and Bae, Jang and Park (2003). 5 Locke and Sarkar (2001) provide indirect evidence on this issue by analyzing trading between customers in futures markets. 6 A stop order becomes a market order after being triggered. Discretionary orders allow the broker to use his expertise to get the best possible fill. As shown in the next section, filling these orders tends to take significant time.
7 orders provide liquidity to other traders, since brokers bid or offer customer limit orders when the limit price represents the best quote. Table I shows that in all six pits liquidity-demanding orders account for approximately 60% to 70% of the customer trading volume. This finding is consistent with the notion that off-floor traders tend to act as liquidity demanders in futures markets. In the liquidity-demanding category, market orders are the most prevalent in terms of the trading volume. Table I also shows that traders often submit marketable limit orders rather than market orders. Marketable limit orders are priced so that they can be executed immediately at the standing quotes. The limit price of such orders is set at the current ask price or higher for buy orders or at the current bid price or lower for sell orders. An advantage of marketable limit orders is that they allow traders to control slippage, i.e., the difference between the market price at the time of order submission and the execution price. To classify limit orders as marketable or non-marketable, we use estimated bid and ask prices at the time of order arrival. 7 Since our data do not include bid-ask quotes, we estimate bid and ask prices using the time and sales data. Prices of up-tick trades are classified as ask prices and prices of down-tick trades are classified as bid prices. Then the last available estimated bid and ask prices are used to classify limit orders. Limit orders submitted before the market opening, as well as orders with missing arrival time, are classified using estimated bid and ask prices at the time of execution. 8 Given that many off-floor traders use the time and sales sequence disseminated by the exchange to make trading decisions, using time and sales data to estimate benchmark quotes appears reasonable. 7 Harris and Hasbrouck (1996) use the quotes at the time of order submission to classify aggressiveness of limit orders. Bessembinder (2003) and Bacidore et al. (2003) discuss the issue of choosing the time of the reference quote. 8 Werner (2003) uses the quotes at the time of execution to classify limit orders, since the order submission times are not available in her dataset.
8 Limit orders account for at least 25% of the customer trading volume in all six markets, suggesting that off-exchange customers provide liquidity to other traders in a significant proportion of their trades. 9 Given that many limit orders may remain unexecuted, it is likely that limit orders account for a larger proportion of customer order flow arriving into the trading pit. This finding does not necessarily imply that some off-exchange traders act as market makers by submitting limit orders on both sides on the market. Instead, it suggests that traders are able to use limit orders to reduce their transaction costs and supply liquidity to other traders in the process. Table I also shows that discretionary orders, as well as both marketable and nonmarketable limit orders, tend to be larger than customer market orders. Orders marked as other in the CTR data also account for a substantial proportion of the customer trading volume. These may be market-if-touched, fill-or-kill, opening only, or other infrequently used order types. Table II shows distribution of customer trading volume by type of counterparty. Consistent with the prior literature, trades with exchange locals account for between 36.7% and 67.6% of customer trading volume. Customers trade with other customers in between 26.3% and 44.0% of their volume. Trades with clearing members account for a substantial proportion of the customer trading volume in both currency pits. The proportions are similar for all customer orders, liquidity-demanding orders and limit orders. It is interesting that off-floor limit order traders frequently trade with exchange locals. In these trades the market makers demand liquidity by effectively acting as market order traders. Overall, the statistics reported in Tables I and II are consistent with the notion that the distinction between market makers and outside customers in open outcry futures markets is somewhat blurred. [Insert Table II about here] 9 Werner (2003) reports a similar proportion of limit orders for downstairs trades in NYSE stocks.
9 C. Turnaround time for different types of customer orders An important dimension of order execution quality is execution speed. While a number of papers examine execution speed in equity markets (e.g., Bacidore et al. (2003), Battalio et al. (2003)), we know of no study that has looked at the order turnaround time in open outcry futures markets. It is interesting to examine execution speed in open outcry futures markets, since the microstructure of these markets differs substantially from microstructure of equity markets. Order arrival timestamps and execution times reported in the CTR data allow us to examine turnaround time for different types of customer orders. Table III reports turnaround time statistics. [Insert Table III about here] As expected, market orders are filled faster than other types of orders, with a median time between order arrival and the reported execution time ranging from 11 sec for the S&P 500 futures to 41 sec for the live cattle futures. Trading appears to be slower in the agricultural pits. For example, while in the index and currency futures markets more than 40% of all customer market orders are executed in less than ten seconds, this proportion is much lower in the agricultural pits. Marketable limit orders follow the market orders, with the median turnaround time ranging from 15 sec to 69 sec. The median turnaround time for limit orders far exceeds that for marketable limit orders. Furthermore, many limit orders remain unexecuted for more than five minutes. Discretionary orders also tend to take a long time to execute. The median turnaround time for this order type ranges from 83 sec in the S&P 500 futures to 145 sec in the live cattle futures. Between 23% and 36% of discretionary orders remain unexecuted for at least five minutes. For all order types the means of the turnaround times are much larger than the medians.
10 Our estimates of the turnaround times are likely to be downward-biased for two reasons. First, because the order arrival time is one of the data items used by the CTR algorithm to determine the execution time, it is possible that for some trades the reported execution time is the same as the order arrival time, while in reality the order was not executed instantaneously. Second, when the execution times are estimated by the CTR algorithm, multiple trades are usually assigned to the same tick in the time and sales sequence. Most of those trades actually occurred during the interval between the tick they are matched to and the next tick. Given that the average time interval between ticks ranges from about 8 sec in the S&P 500 futures to about 42 sec in the lean hog futures, the impact of this feature of the CTR data on our estimates of order turnaround times is likely to be limited. The CME offers side by side electronic and open outcry trading in the futures contracts considered in our paper. The execution times for different types of orders that we document are useful to traders who consider the choice between the open outcry and electronic trading venues. The execution speed statistics may also be used by the CME in the analysis of execution quality in the CME s trading pits. III. Empirical Results A. Execution costs for different types of customer orders Analysis of trading costs and market frictions is central to market microstructure research. While transaction costs in equity markets have been extensively studied, the literature on execution costs in futures markets is more limited. Bid-ask quotes in open outcry futures markets are good only as long as the breath is warm and are not recorded systematically. Consequently, studies
11 of transaction costs in futures markets have used time and sales data or trade reconstruction data to estimate effective bid-ask spreads. Locke and Venkatesh (1997) and Ferguson and Mann (2001) estimate effective execution costs for off-floor customers (or execution spreads) by aggregating across all customer orders during an interval. 10 They argue that this social cost, i.e., the cash flow from customers to market makers, is the relevant measure of transaction costs. This approach produces a useful measure of aggregate customer trading costs. We extend the prior literature by separately analyzing execution spreads for different types of customer orders. Trades between customers have zero social cost but liquidity-demanding customer orders have to pay a liquidity premium, even when they are executed against orders of other customers. The execution costs for each trader depend on the trader s choice of order types. Therefore, it is important to be able to estimate execution costs for liquidity-demanding and limit customer orders. Similar to Locke and Venkatesh (1997) and Ferguson and Mann (2001), we calculate customer execution spread as average customer buy price minus average customer sell price over a five-minute interval. 11 Table IV shows execution spreads calculated for all customer orders, as well as for customer liquidity-demanding and limit orders. Unsurprisingly, execution spreads are very different for different types of orders. For example, the all-trade execution spread in the S&P 500 futures market is about one basis point. At the same time, the execution spread is about 3.2 basis points for liquidity-demanding orders and about 4.3 basis points for limit orders. This finding is consistent with the notion that traders using liquidity-demanding orders tend to pay the bid-ask spread, while customers using limit orders tend to earn the spread. [Insert Table IV about here] 10 This measure is also used by Manaster and Mann (1996) and Chang and Locke (1996). 11 We repeated calculations using average prices weighted by trade size. The results, which are not reported for brevity but available upon request, appeared similar.
12 In the S&P 500 market, about 20% of the five-minute spread estimates for all customer orders are negative. The corresponding proportion is only 4.1% for liquidity-demanding orders and 89.3% for limit orders. The proportions are qualitatively similar in the other pits. These results suggest that negative execution spreads observed by Ferguson and Mann (2001) are likely to be explained by frequent executions of customer limit orders. In essence, the aggregate execution spread is a weighted average of the spreads for liquidity-demanding and limit orders. As shown in the previous section, customers tend to use liquidity-demanding orders. Therefore, the aggregate execution spread is positive on average, although it becomes negative in intervals in which executions of customer limit orders dominate. The execution spreads reported in Table IV imply that the aggregate measure of transaction costs needs to be used with caution in evaluation of market rule changes. 12 For example, if some improvements in execution quality (e.g., faster execution) make market orders relatively more attractive, their proportion in the customer trading volume is likely to increase, resulting in a higher estimate of the aggregate transaction costs. While such a change may be Pareto improving, based on the aggregate measure of transaction costs it will appear to have a negative effect on the market liquidity. Manaster and Mann (1999) and Ferguson and Mann (2001) explain the apparent ability of off-exchange traders to buy low and sell high by price concessions offered by market makers. This argument implicitly assumes that off-exchange traders have no control over the spreads they pay. In contrast, the spread estimates reported in Table IV suggest that the choice of order type has a large effect on customer execution costs. Essentially, by submitting limit orders off- 12 Locke and Venkatesh (1997, p. 236) argue that the aggregate execution spread should be used to evaluate the effect of changes affecting market microstructure. Chang and Locke (1996) use this measure to evaluate the effects of dual trading restrictions.
13 exchange traders are able to compete with locals for market orders. Limit order traders provide liquidity to other traders and earn the bid-ask spread in the process. Our category of liquidity-demanding orders includes several types of orders that may face different execution costs. Therefore, we separately examine the execution spreads for three common types of liquidity-demanding orders including market, discretionary, and marketable limit orders. The results are presented in Table V. Interestingly, the execution spreads appear to be lowest for discretionary orders, even though these orders tend to be relatively large. The additional time spent by brokers to execute discretionary orders appears to translate into lower execution costs. In the index futures markets, marketable limit orders have the highest execution spreads: 3.6 basis points in the S&P 500 futures and 10.2 basis points in the Nasdaq-100 futures. This finding is consistent with Peterson and Sirri (2002) and Werner (2003), who examine marketable limit orders in equity markets. However, in the currency and agricultural pits the execution spreads of marketable limit orders are below those of market orders. [Insert Table V about here] It is also noteworthy that the execution spreads for limit orders tend to be greater in absolute value than the spreads for market orders. This finding is likely to be explained by the difference between customer limit orders and locals quotes. Exchange locals compete for market orders through an open outcry auction and the trade goes to the local who bids or offers the best price. As a result, a market order is often executed at a better price than the best bid or offer that existed before the order was announced. In contrast, when brokers bid or offer customer limit orders, they may not be able to quote aggressively, i.e., the brokers cannot bid a price higher or offer a price lower than the limit price. Therefore, conditional on execution, limit order traders tend to earn higher spreads than the spreads paid by market order traders.
14 B. Are customer limit orders systematically picked off by market makers? An issue related to customer trading costs is whether customer limit orders are systematically picked off by market makers. It may take significant time for an off-exchange customer to cancel a limit order. Because of their direct access to the trading pit, exchange locals are able to react to changing market conditions much faster than off-exchange customers. Manaster and Mann (1999) argue that exclusive access to customer order flow may also allow locals to time their trades better than off-floor traders. Consequently, limit orders submitted by off-exchange traders are likely to face increased risk of adverse selection. 13 Manaster and Mann (1999) and Ferguson and Mann (2001) suggest that when locals want to take a position to profit from an anticipated price movement they sometimes allow offexchange customers to buy low and sell high. Under this scenario, we would expect to see a systematic price movement away from executed customer limit orders. We test this hypothesis by looking at returns surrounding customer limit orders trades with market makers. Customer limit order trades are likely to be clustered. This clustering may affect the results if we simply calculate average returns around such trades. To account for clustering, we use a method similar to the one suggested by Harris, Sofianos, and Shapiro (1994). We estimate the following regression of tick-by-tick returns on 9 leads and 30 lags of customer limit order trades with exchange locals: 14 R t 9 9 buy sell 0 + β i Lt + i γ i Lt + i + i= 30 i= 30 = α + ε t, (1) 13 Handa and Schwartz (1996) and Harris and Hasbrouck (1996), among others, discuss the adverse selection problem faced by limit order traders. In an empirical study of NYSE stocks, Werner (2003) shows that limit orders tend to get picked off. 14 Given that Manaster and Mann (1996) show that locals tend to reduce inventory by half in the next trade, a 30-tick post-trade interval is likely to be sufficient. This interval ranges from about 3.9 min in the S&P 500 futures to about 20.8 min in the lean hog futures. We repeated the analysis using 1-min returns and a 15-min post-trade interval. The results appeared similar.
15 where R t represents tick-by-tick returns, 15 and L buy and L sell are (0, 1) indicator variables for customer buy and sell limit order trades with locals, respectively. If several limit order trades of the same type occurred at the same tick, only one trade is used in calculations. Equation (1) is estimated using the generalized method of moments (GMM) and the Newey and West (1987) heteroscedasticity and autocorrelation consistent covariance matrix. Figure 1 plots the cumulative average returns (CARs) surrounding customer limit order trades with locals. The CARs are obtained by estimating equation (1). The figure shows positive return drifts before limit order sells and negative return drifts before limit order buys. These price changes are likely to cause the exposure and subsequent execution of the limit orders by making the limit price the best quote. In all markets after the trade there appears to be a steady price movement in the direction of the executed limit order. [Insert Figure 1 about here] A summary of lead-lag regression results is shown in Table VI. The absolute values of post-trade cumulative returns range from about one basis point in the Euro futures to about 4.5 basis points in the lean hog futures. These post-trade returns lead us to reject the hypothesis proposed by Manaster and Mann (1999) and Ferguson and Mann (2001) that locals allow offfloor customers to buy low and sell high to take positions ahead of anticipated price movements. Customer limit orders do not appear to be systematically picked off by exchange locals. Furthermore, the post-trade return drifts in the direction of the limit orders suggest that customer limit orders in these markets are informative. [Insert Table VI about here] 15 We calculate returns based on trade prices obtained from the time and sales data. The negative autocorrelation of returns induced by the bid-ask bounce may affect the results. We repeated the calculations using pseudoequilibrium prices as suggested by Ederington and Lee (1995). Pseudo-equilibrium price is calculated as a moving average of the last two prices in the time and sales sequence. The results, which are not reported for brevity but available upon request, were qualitatively unchanged.
16 C. Do common spread estimators capture the cost of immediacy? Most of the studies that examine the bid-ask spreads in open outcry futures markets estimate the spreads using time and sales data that contain execution times and prices of trades that result in price changes. 16 Commonly used spread estimators include Roll (1984) serial covariance estimator, the mean absolute price change estimator suggested by Thompson and Waller (1988), and the method of moments estimator of Smith and Whaley (1994). ap Gwilym and Thomas (2002) examine performance of common spread estimators using trade and quote data. They conclude that although such estimators appear to produce biased estimates, they can be used as a proxy for effective and realized spreads in futures markets. ap Gwilym and Thomas (2002) look at effective and realized spreads for all trades, including trades initiated by exchange locals. It is reasonable to suggest, however, that the focus in estimation of trading costs should be on costs incurred by off-exchange customers, who are the end-users of markets. Our dataset allows us to distinguish between local trades and trades initiated by off-exchange customers. The buy/sell codes and order type identifiers available in the CTR data for both sides of each trade also allow us to avoid having to infer trade direction. 17 We examine whether common spread estimators and time and sales data can be used to reliably estimate the cost of immediacy for off-floor customers in futures markets. Our study complements Locke and Venkatesh (1997) and Ferguson and Mann (2001), who look at the relationship between the execution spread for all customer orders and some of the common estimators of futures spreads. Specifically, Locke and Venkatesh (1997) examine 16 Hasbrouck (2003) develops an econometric methodology for making inferences about market liquidity using time and sales data. 17 ap Gwilym and Thomas (2002) and most studies of effective and realized spreads in equity markets use trade classification algorithms, such as the Lee and Ready (1991) algorithm or the tick rule, to sign trades as buyer- or seller-initiated. Aitken and Frino (1996), Finucane (2000) and Ellis, Michaely and O Hara (2000) examine the accuracy of several trade classification algorithms and show that the existing methodologies incorrectly classify a substantial proportion of trades.
17 the Roll (1984) serial covariance estimator and the Smith and Whaley (1994) method of moments estimator, while Ferguson and Mann (2001) consider only the Roll s estimator. Both studies come to the same conclusion: there is no relationship between the aggregate customer transaction costs in futures markets and the common time series spread estimators. The bid-ask spread represents compensation for liquidity provision, or cost of immediacy, which is the essence of market liquidity (e.g., Grossman and Miller (1988)). Common spread estimators may be useful if they capture the cost of immediacy for off-floor customers. The CTR data allow us to directly measure the cost of immediacy for off-floor customers by calculating the execution spread for customer market orders. This approach is consistent with Stoll (2000), who uses a similar measure to estimate trading costs in equity markets. 18 We examine whether the common spread estimators can be used as a proxy for the cost of immediacy faced by off-exchange customers. We then look at the relative performance of four different estimators to determine which estimator represents the best proxy for the cost of immediacy. Thus, in contrast to Locke and Venkatesh (1997) and Ferguson and Mann (2001), we focus on the cost of immediacy rather than on the aggregate customer trading costs. Roll (1984) serial covariance estimator is calculated using trade prices as S R 2 (cov( Pt, P 1 ). A commonly encountered weakness of this estimator is that it is = t undefined in intervals with positive serial correlation of price changes. The spread estimator suggested by Thompson and Waller (1988) is calculated simply as the average absolute value of price changes during an interval. A potential drawback of this estimator is that fundamental price changes are treated as bid-ask bounce, leading to possible biases. The method of moments 18 This measure, called traded spread, is calculated as the average price of trades on the ask side minus the average price of trades on the bid side. Therefore, it measures execution costs for liquidity-demanding orders.
18 estimator proposed by Smith and Whaley (1994) is designed to correct this problem. 19 This estimator attempts to purge the effects of fundamental price changes on the estimated effective spread. Assuming normal distribution for the fundamental price changes, Smith and Whaley (1994) show that the effective spread can be estimated using the first two moments of the empirical distribution of absolute price changes. Another estimator, which is used by the CFTC, is calculated as the mean absolute value of price reversals. 20 Price changes in the same direction as immediately preceding price changes are removed, as they are likely to represent fundamental price changes. Finally, the spread estimator suggested by Bhattacharya (1983) is similar to the CFTC estimator in that it attempts to eliminate fundamental price changes. However, this estimator uses only non-overlapping consecutively reverting price changes in groups of three or more. 21 Table VII reports means of four different spread estimators and their biases relative to the execution spread for customer market orders. 22 The spread estimators appear to underestimate the cost of liquidity in the index and currency futures markets. For example, for the S&P 500 futures the estimator bias ranges from about 24.6% for the Smith and Whaley (1994) estimator to about 6.4% for the Bhattacharya (1983) measure. At the same time, all of the estimators appear to overestimate the cost of immediacy in the agricultural futures markets. For all spread estimators (with the exception of the CFTC estimator for the Nasdaq-100 futures and the Bhattacharya (1983) measure for both index futures markets) the difference between the estimators and the execution spread for customer market orders is statistically significant. 19 This estimator is used by Fleming, Ostdiek and Whaley (1996) and Bollen, Smith and Whaley (2003), among others. 20 The CFTC estimator is used by Wang, Michalski, Jordan and Moriarty (1994) and Wang, Yau and Baptiste (1997). 21 Ma, Peterson and Sears (1992) use Bhattacharya (1983) and Thompson and Waller (1988) estimators to estimate bid-ask spreads in futures markets. 22 We do not consider Roll (1984) serial covariance estimator, since ap Gwilym and Thomas (2002) show that this estimator is frequently undefined and exhibits erratic behavior.
19 [Insert Table VII about here] Why do the common spread estimators tend to underestimate the cost of immediacy faced by off-floor customers? A possible explanation is that the common spread estimators have a downward bias. Alternatively, off-exchange customers demanding liquidity may tend to pay higher than average spreads. Many trades between locals are likely to be inventory control scratch trades. It is possible that when locals trade with other locals they often meet half-way between the standing quotes, paying less for liquidity than off-floor customers. 23 The results in Table VII show that the negative bias tends to be smaller for estimators that attempt to purge fundamental price changes. This finding suggests that the underlying price changes are usually smaller than the price fluctuations between the bid-ask quotes. The results discussed above suggest that common spread estimators can be used as a proxy for the execution costs incurred by customer market orders, although all of the estimators we consider tend to produce biased estimates. In the financial futures markets, the Bhattacharya (1983) measure appears to be closest to customer execution spread for market orders. This finding is consistent with ap Gwilym and Thomas (2002), who show that the Bhattacharya (1983) measure produces better estimates of effective spreads than other common spread estimators. A potential weakness of the Bhattacharya (1983) method is that it eliminates a large proportion of the data and therefore can be used only for frequently traded markets. The CFTC estimator, which discards fewer prices, may be a reasonable alternative in less active markets. We also examine correlations of the CFTC spread estimator with execution spreads for all customer orders, as well as for market and limit orders. The results presented in Table VIII show that the CFTC bid-ask spread estimator is highly correlated with the execution spread for 23 Pirrong (1996) suggests that locals trade less aggressively with other locals than with brokers.
20 customer market orders. 24 Correlation between the CFTC estimator and the execution spread for market orders ranges from about 0.25 for the Japanese Yen futures to about 0.89 for the S&P 500 futures. 25 At the same time, there is a strong negative correlation between the CFTC estimator and the execution spread for customer limit orders. This negative correlation suggests that when bid-ask spreads increase limit order traders earn greater liquidity premium. The findings presented in Tables VII and VIII show that common spread estimators are closely related to the cost of immediacy faced by off-exchange customers in futures markets, although they tend to produce biased estimates. [Insert Table VIII about here] IV. Summary and Conclusion Recent studies of trading costs in futures markets, including Locke and Venkatesh (1997) and Ferguson and Mann (2001), argue that the appropriate measure of trading costs can be obtained by aggregating across all customer trades. This measure, called execution spread, reflects the aggregate customer trading costs. Surprisingly, Ferguson and Mann (2001) show that in many intraday intervals the customer execution spreads are negative. They suggest that negative execution spreads may be related to trading strategies of market makers, who offer customers price concessions to take positions ahead of favorable price movements. Our empirical results show, however, that customer limit orders are not systematically picked off by market makers. The choice of the order type is an important determinant of trading costs. Therefore, in addition to the analysis of the aggregate customer trading costs, it is important to examine 24 We also examined correlations of the execution spreads with Thomson and Waller (1988), Smith and Whaley (1994) and Bhattacharya (1983) estimators. Since these estimators are highly correlated with the CFTC estimator, we do not report these results. 25 The relatively low correlation for the Japanese Yen futures is likely to be explained by the relatively little variation in the bid-ask spreads during our sample period.
21 execution costs separately for liquidity-demanding and limit orders. We show that limit orders account for a significant proportion of the customer order flow in open outcry futures markets. Limit order traders supply liquidity to other customers, in effect competing with market makers. Furthermore, the limit order traders frequently provide liquidity to market makers. We also find that off-floor customers frequently submit marketable limit orders rather than market orders, presumably to control the execution price uncertainty. We know of no other study that directly examines composition of the customer order flow in open outcry futures markets. Our study is also the first to examine execution costs and execution speed for different types of customer orders in futures markets. The results show that, while the execution spreads for liquidity-demanding orders are almost always positive, limit order traders tend to earn the liquidity premium by buying low and selling high. Therefore, the finding of Ferguson and Mann (2001) that customer execution spreads are often negative is likely to be explained by frequent trades of market makers against customer limit orders. The execution quality statistics that we document are useful to futures traders and policy makers. Our study also makes a methodological contribution by reconciling common time series spread estimators with the execution spread measure. We show that the common spread estimators can be used as a proxy for the execution costs incurred by off-exchange traders using market orders. We conclude that the common spread estimators are closely related to the cost of immediacy in futures markets.
22 References Aitken, M., and Frino, A., 1996. The accuracy of the tick test: Evidence from the Australian stock exchange. Journal of Banking and Finance 20, 1715-1729. ap Gwilym, O., and Thomas, S. 2002. An empirical comparison of quoted and implied bid-ask spreads on futures contracts. Journal of International Financial Markets, Institutions and Money 12, 81-99. Bacidore, J., Ross, K., and Sofianos, G. 2003. Quantifying market order execution quality at the New York Stock Exchange. Journal of Financial Markets 6, 281-307. Bae, K. H., Jang, H., and Park, K. S. 2003. Traders choice between limit and market orders: Evidence from NYSE stocks. Journal of Financial Markets 6, 517-538. Battalio, R., Hatch, B., and Jennings, R. 2003. All else equal?: A multidimensional analysis of retail, market order execution quality. Journal of Financial Markets 6, 143-162. Bessembinder, H. 2003. Issues in assessing trade execution costs. Journal of Financial Markets 6, 233-257. Bhattacharya, M., 1983. Transaction data tests of efficiency of the Chicago Board of Options Exchange. Journal of Financial Economics 12, 161-185. Bollen, N. P. B.; Smith, T.; and Whaley, R. E., 2003. Optimal contract design: For whom? Journal of Futures Markets 23, 719-750. Chang, E. C., and Locke, P. R., 1996. The performance and market impact of dual trading: CME rule 552. Journal of Financial Intermediation 5, 23-48. Demsetz, H. 1968. The cost of transacting. Quarterly Journal of Economics 82, 33-53. Ederington, L. H., and Lee, J. H. 1995. The short-run dynamics of the price adjustment to new information. Journal of Financial and Quantitative Analysis 30, 117-134. Ellis, K., Michaely, R., O Hara, M. 2000. The accuracy of trade classification rules: Evidence from Nasdaq. Journal of Financial and Quantitative Analysis 35, 529-552. Ferguson, M. F., and Mann, S. C., 2001. Execution costs and their intraday variation in futures markets. Journal of Business 74, 125-160. Finucane, T. J., 2000. A direct test of methods of inferring trade direction from intraday data. Journal of Financial and Quantitative Analysis 35, 553-576. Fleming, J., Ostdiek, B., and Whaley, R. E. 1996. Trading costs and the relative rates of price discovery in stock, futures and options markets, Journal of Futures Markets 16, 353-387.
23 Grossman, S, J., and Miller, M. H. 1988. Liquidity and market structure. Journal of Finance 43, 617-633. Handa, P., and Schwartz, R. 1996. Limit order trading. Journal of Finance 51, 1835-1861. Harris, L. 2003. Trading and exchanges: Market microstructure for practitioners. Oxford University Press. Harris, L., and Hasbrouck, J. 1996. Market vs. limit orders: The SuperDOT evidence on order submission strategy. Journal of Financial and Quantitative Analysis 31, 213-231. Harris, L., Sofianos, G., and Shapiro, J. E. 1994. Program trading and intraday volatility. Review of Financial Studies 7, 653-685. Hasbrouck, J. 2003. Liquidity in the futures pits: Inferring market dynamics from incomplete data. Journal of Financial and Quantitative Analysis, forthcoming. Lee, C. M., and Ready, M. J. 1991. Inferring trade direction from intraday data. Journal of Finance 46, 733-746. Locke, P. R., and Sarkar, A., 2001. Liquidity supply and volatility: Futures market evidence. Journal of Futures Markets 21, 1-17. Locke P. R., and Venkatesh, P. C., 1997. Futures markets transaction costs. Journal of Futures Markets 17, 229-245. Ma, C. K., Peterson R. L., and Sears, R. S. 1992. Trading noise, adverse selection, and intraday bid-ask spreads in futures markets. Journal of Futures Markets 12, 519-538. Manaster, S., and Mann, S. C. 1996. Life in the pits: Competitive market making and inventory control. Review of Financial Studies 9, 953-975. Manaster, S., and Mann, S. C. 1999. Sources of market making profits: Man does not live by spread alone. Working Paper. Texas Christian University. Newey, W., and West, K. 1987. A simple positive semi-definite, heteroscedasticity and autocorrelation consistent covariance matrix. Econometrica 55, 703-708. Peterson, M., and Sirri, E. 2002. Order submission strategy and the curious case of marketable limit orders. Journal of Financial and Quantitative Analysis 37, 221-241. Pirrong, C. 1996. Market liquidity and depth on computerized and open outcry trading systems: A comparison of DTB and LIFFE Bund contracts. Journal of Futures Markets 16, 519-543.
24 Roll, R. 1984. A simple measure of effective bid-ask spread in an efficient market. Journal of Finance 39, 1127-1139. Smith, T., and Whaley, R. E. 1994. Estimating the effective bid/ask spread from time and sales data. Journal of Futures Markets 14, 437-455. Stoll, H. R. 2000. Friction. Journal of Finance 55, 1479-1514. Thompson, S., and Waller, M. 1988. Determinants of liquidity costs in commodity futures markets. Review of Futures Markets 7, 110-126. Wang, G., Michalski, R., Jordan, J., and Moriarty, E. 1994. An intraday analysis of bid-ask spreads and price volatility in the S&P 500 index futures markets. Journal of Futures Markets 14, 837-859. Wang, G., Yau, J., and Baptiste, T. 1997. Trading volume and transaction costs in futures markets. Journal of Futures Markets 17, 757-780. Werner, I. M. 2003. NYSE order flow, spreads, and information. Journal of Financial Markets 6, 309-335.
25 Table I Distribution of customer trading volume by type of order Liquidity-demanding orders Market on Market Discretionary Stop close Marketable limit Limit Other S&P 500 Mean size 3.72 9.22 2.75 5.46 6.10 5.25 3.24 Percent of total volume 29.9% 14.8% 4.0% 1.1% 14.9% 31.0% 4.2% Nasdaq-100 Mean size 3.33 5.86 2.48 2.83 4.27 3.64 2.63 Percent of total volume 29.3% 11.3% 3.0% 0.4% 14.4% 26.2% 15.6% Euro Mean size 3.53 7.65 3.11 5.60 7.03 5.07 8.45 Percent of total volume 30.6% 1.2% 5.1% 2.1% 23.1% 28.3% 9.6% Yen Mean size 3.41 10.50 3.30 4.96 6.86 5.58 5.94 Percent of total volume 28.3% 2.5% 7.5% 2.7% 23.3% 27.8% 7.7% Lean Hogs Mean size 2.43 5.48 2.24 3.88 4.15 3.00 3.06 Percent of total volume 32.6% 15.8% 7.8% 3.0% 11.8% 25.0% 4.0% Live Cattle Mean size 3.43 7.55 3.01 5.15 6.34 4.31 4.19 Percent of total volume 29.0% 18.8% 5.8% 2.1% 14.3% 26.1% 3.9% Orders are classified as marketable limit orders if the order price is equal to or exceeds estimated ask price for buy orders or if the order price is equal to or below the estimated bid price for sell orders. Bid and ask prices at the time of order arrival are estimated from the time and sales data as follows. The prices of up-tick trades are classified as ask prices and prices of down-tick trades are classified as bid prices. Then the last available estimated bid and ask prices are used to classify limit orders. Limit orders submitted before the market opening, as well as orders with missing arrival time are classified using estimated bid and ask prices at time of execution. Trades for which the order arrival time exceeds the reported execution time are removed.