Informed Traders and Convergence to Market Efficiency: Evidence from a New Commodities Exchange∗ Stefan Hunt† October 15, 2010
Abstract I document the existence and disappearance of momentum in commodity prices on a large, new agricultural futures exchange in India, and I investigate which mechanisms caused momentum to disappear. Using a complete record of transactions on the exchange for over four years, containing identifiers for buyers and sellers, I reconstruct the daily positions and profits of the universe of participants, and find three main results. First, after the launch of the market, momentum trading strategies using prices from the previous two weeks earn substantial excess returns, and over a four year period the returns to these strategies steadily decline. Second, traders with large position sizes (large traders) have a similar pattern of decline for their investment returns, and much of their returns can be explained by momentum. As profits from trading are zero sum, large traders gain at the expense of small and medium traders, the majority of market participants, who make considerable losses. Third, the decline in momentum profits coincides with a relatively high rate of market exit by small and medium traders, supporting Friedman’s (1953) claim that, in an inefficient market, naive investors should make losses and drop out, leading prices to become more efficient. In addition, over time the losses of medium traders decrease more than those of small traders, consistent with learning by medium traders, a second mechanism leading to increased efficiency. I also present evidence that the root cause of the momentum is gradual diffusion of information through market participants.
Keywords: market efficiency, emerging markets, commodity futures, agricultural commodities JEL Classification: G13, G14, Q11
I thank my advisors John Campbell, Shawn Cole, and Sendhil Mullainathan for many invaluable conversations. I also thank Ryan Bubb, Josh Coval, Esther Duflo, Stefano Giglio, Asim Khwaja, Paul Niehaus, Rohini Pande, Laura Serban, Jeremy Stein, Andrei Shleifer and participants of Harvard Finance Lunch for helpful comments. I am grateful to the management and staff of NCDEX, Raj Benahalkar, Ankur Garg, Anand Iyer, Nirmalendu Jajodia, Uma Mohan, Ravinder Sachdev, and Siddharth Surana. Special thanks go to Jagdish Choudhry and Mr. Ramaseshan. This paper does not necessarily reflect the views of NCDEX. I acknowledge the support of the Warburg Fund and Harvard South Asia Initiative. All errors are mine. † Department of Finance, School of Management, Yale University
Introduction Recent research on financial markets in developed countries finds that many asset-pricing anoma-
lies have weakened or disappeared since their initial discovery (Schwert, 2003). Two explanations are typically given for these findings. First, markets are efficient and supposed anomalies are spurious findings and the result of misspecified asset-pricing models (Fama and French, 1996) or data mining (Sullivan, Timmermann, and White, 2001). Second, risk-adjusted returns can at times be predictable using public information; but informed traders exploiting the profit opportunity causes other traders to make losses, and then market participants to adjust their trading behavior and the market to become more efficient. This second “adaptive markets” explanation is an obvious market adjustment mechanism and is frequently referred to, for example Campbell, Ramadorai, and Schwartz (2009) suggest capital deployed by institutional investors may have caused the “post earnings announcement drift” to diminish. But surprisingly we have little concrete evidence that some traders successfully exploit anomalies and even less evidence on which processes lead markets to become more efficient.1 The lack of evidence is in part because we have limited and incomplete data on the positions of traders in markets with anomalous behavior. With such data we can address important open questions. Do savvy traders profit from anomalies as predicted by the adaptive markets view? If so, do inefficiencies persist for a long period of time before, and after, the better-informed traders begin to exploit them? Precisely which processes cause anomalies to fade: do the uninformed exit the market (a mechanism that Friedman, 1953 and Fama, 1965 invoke to argue that markets must be efficient), the uninformed become more knowledgeable, or the informed trade more aggressively? The answers to these questions may be particularly pertinent in the context of developing countries, where securities markets may be more prone to being inefficient and we want to know how and how quickly these markets will mature. This paper studies trading on a new commodity futures market launched in India in late 2003, the National Commodity and Derivatives Exchange (NCDEX), using a unique transaction-level dataset identifying buyers and sellers. After the launch, I find that short duration momentum trading strategies, buying futures whose price has risen in the past two weeks and selling futures whose prices has fallen, earn substantial and persistent excess returns. Traders with the largest positions (“large traders”) are believed to be more informed and they earn substantial excess returns, which strongly correlate with the returns to the momentum strategies. Because profits from trading are zero sum, medium and small traders make substantial losses. As the exchange matures, the momentum strategies become less profitable. Large traders earn lower returns as well, which become increasingly less associated with momentum, indicating that they trade less aggressively on the anomaly. Throughout this period small and medium traders 1
Lo (2004) outlines an Adaptive Markets Hypothesis, a qualitative theory that, because of the behavioral biases of investors, markets can be inefficient, but that over time markets evolve. My use of the term “adaptive markets” is broadly consistent with his usage of the term.
drop out of the market at a faster rate than large traders, who consequently increase their market share. The losses of medium traders reduce considerably while those of small traders reduce much less, consistent with medium traders learning more about the anomaly. Together these findings suggest that the profitability of momentum strategies is not due to a misspecified model of risk nor data-mining but rather that prices on NCDEX were initially weak-form inefficient and that traders with information advantage exploited their knowledge. The findings also provide some preliminary evidence on the mechanisms for adjustment, suggesting that differential drop out by uninformed traders and learning by uninformed traders were important. The NCDEX market for commodity futures provides a unique opportunity to study how a major new financial market adapts and evolves. First, the data set that is available is unusually detailed: it consists of the entire history of transactions on the exchange from inception for over four years, and includes trader identifiers and some demographic information. The data allow me to reconstruct the positions of any participant at any point in time. Second, the exchange was, and is, a sizeable and important market. During the sample period, it was the dominant agricultural exchange in the world’s second most populous country, and it became the third largest market for agricultural futures globally2 Third, in a setting with many inexperienced traders, it may be more likely that simple market inefficiencies persist for some time, which can be readily studied. NCDEX opened following legislative change that paved the way for the creation for new commodity futures markets, which attracted, for the first time, small retail investors throughout India. Such a setting may be comparable to new markets in the US; for example, when index futures began in the early 1980s, simple arbitrage opportunities persisted for at least eighteen months (Thomas, 2002; Chung, 1991). I begin my empirical analysis with a test for weak-form inefficiency, focusing on momentum because of its simplicity and its ubiquity across asset classes. The momentum strategies that I consider buy commodities with price rises and sell commodities with price falls over a period in the recent past, up to 50 days beforehand. While many studies investigate longer periods than this, I focus on periods of shorter length partly because a new exchange has no price history and partly because it seems plausible that when an exchange begins and there are few services that provide market news , information may be embedded in prices slowly, over some days.3 Controlling for risk using a three-factor model used by Miffre and Rallis (2007) that includes equity returns, bond returns, and commodity index returns, I find substantial short duration momentum and modest longer term price reversal; prices continue on trend for one month and then reverse over 2 For 19 of its 24 principal agricultural futures contracts it was the largest market domestically. Globally, it was smaller than the Dalian Commodity Exchange and the Chicago Board of Trade but larger than the New York Board of Trade (UNCTAD, 2006). Size is measured by number of contracts traded, the standard size metric used to compare exchanges worldwide. Data compiled by the Futures Industry Association. 3 Reuters, for example, only began covering prices for Indian commodity futures in December 2004, launched news feeds in November 2005, and continued to develop the depth of news and number of commodities covered through the sample.
the following month. These effects decline markedly as the exchange becomes more established. For example, a strategy based on price changes over the past one day earns an excess return of 55% per annum with a Sharpe ratio of 4.5 in 2004, but this falls to 11% with a Sharpe ratio of 1.7 by 2007. To understand whether some traders exploit return predictability, I identify and classify traders likely to be informed. Given the limited demographic information available, I focus on using past trading to categorize investors and use recent position size in the commodity as a proxy for informedness, categorizing traders as small, medium or large. There are two motivations for the use of recent position size: in a market evolving towards efficiency, traders that have successful strategies can be expected to grow in size both due to prior gains and due to incentives (see, for example, Luo, 1998); and the vast majority of traders on NCDEX are households, and they might be expected to be both smaller and less informed than other participants, such as agricultural trading firms and commodity futures brokers. As predicted by size proxying informedness, large traders make substantial gains of 23% in the initial year, at the expense of medium and small traders, who lose 16% and 22%. As the exchange matures, the magnitude of gains and losses diminishes monotonically through to 2007: though while large trader gains and medium trader losses fall considerably to 3%, small trader losses remain substantial at 12%. These changes in return coincide with the large trader share of total capital increasing from 43% to 66% indicating differential drop out of small and medium traders from the market. To assess whether these gains and losses are due to the anomaly, I regress large trader returns on the returns from each of the momentum strategies. The beta coefficients on these factors are highly significant and together the factors account for almost two thirds of the average inter-day returns of large traders.4 Thus, the traders believed informed make substantial profits in large part because their portfolio returns correlate with the returns of momentum strategies, consistent with informed traders exploiting inefficiencies. As large traders trade on momentum less over time not more, the anomaly appears to decrease because the uninformed traders either exit the market or learn about the inefficiency. I extend the analysis further to understand why short duration momentum existed at the beginning of NCDEX. Theories that seek to explain momentum fall into two camps. On the one hand, momentum may be caused by investor reactions to public information, by under or over-reaction due to conservatism and representativeness (Barberis, Shleifer, and Vishny, 1998) or by responses to gains and losses caused by price changes (Grinblatt and Han, 2001) . On the other hand, momentum may be caused by investor reactions to private information, from gradual diffusion of 4
Inter-day returns are appropriate because the momentum strategies rebalance at a daily frequency.
private information coupled with imperfect filtering of the information of others from prices (Hong and Stein, 1999) or from investor overconfidence in the value of their private information (Daniel, Hirshleifer, and Subrahmanyam, 1998) . I provide support for private information being the cause of momentum in this setting by testing two predictions. First, traders may be informed because they have private information or because they conduct technical analysis to identify inefficiencies. If the latter is true, we should expect that aggregate, as opposed to commodity-specific, trader size should proxy for informedness. The results demonstrate however that aggregate size has no predictive power for trader returns over and above commodity-specific trader size, which suggests informedness is commodity-specific. Second, I test whether commodities with a greater fraction of commonly-observed information have lower momentum. The test is directly analogous to the “low analyst coverage” test of Hong, Lim, and Stein (2000), Zhang (2006), and Hwang (2008) who find higher momentum in US stocks with fewer analysts and interpret this findings as supporting theories that explain momentum using private information, such as gradual information diffusion. My indicator for commonlyobserved information is whether the commodity has a pre-existing international futures markets. As predicted, these commodities have both reduced momentum and reduced returns for large traders. Overall the evidence supports gradual information diffusion over and above investor overconfidence as there is only modest evidence for price reversal, a key component of the model of Daniel, Hirshleifer, and Subrahmanyam (1998). To my knowledge the only other paper that investigates the trading of participants on a developing country exchange is Khwaja and Mian (2005), who provide evidence for a particular type of “pump and dump” price manipulation by equity brokers in Pakistan. That paper, similar to this one, uncovers sizeable losses for the majority of exchange participants, and it has numerous regulatory policy implications for emerging markets. The findings in this paper may suggest that in the context of fledgling markets the World Bank, which has championed the expansion of futures exchanges in developing countries since the late 1990s, and developing country governments consider measures to ameliorate small investor losses, in addition to more traditional regulatory considerations such as managing counterparty risk and preventing manipulation (see Morgan, Rayner, and Vaillant, 1999, and Morgan, 2001). I return to these regulatory considerations in the discussion in the penultimate section below. This paper advances a large literature that identifies anomalies and evaluates whether they are market inefficiencies, reviewed by Daniel, Hirshleifer, and Subrahmanyam (1998), Fama (1998) and Keim and Ziemba (2000). The current paper relates most closely to the subset of this research that examines whether anomalies persist following their discovery, typically finding that they fade. Schwert (2003) provides evidence that the size effect, value effect, weekend effect, and dividend
yield effect cease to exist in more recent sample periods. There have been similar findings for the weekend effect; the small-firm effect; the January effect, the day-of-the-week effect; and the holiday effect, time-of-the month effect and the turn-of-the-month effect (Chang, Pinegar, and Ravichandran, 1993, Dimson and Marsh, 1999, Mehdian and Perry, 2002, Hudson, Keasey, and Littler, 2002, and Marquering, Nisser, and Valla, 2006 respectively). Papers that find short-term mispricings decrease over time include Chordia, Roll, and Subrahmanyam (2005) on delayed price response to order imbalances, and Thomas (2002) and Chung (1991) on index arbitrage. Schwert (2003) notes that most anomalies waned soon after publication of the original research findings, and he offers anecdotes that this coincides with investment firms trading on the phenomenon in some cases, but he admits that the available evidence, mostly based on the profitability of trading strategies, does not differentiate between the efficient and adaptive market explanations. This paper provides direct support that markets can adapt by examining the trading and profits of investors and by providing initial evidence on which channels operate to bring the market towards an efficient equilibrium. The results here also contribute to research that examines behavioral biases and trading losses for individual investors or, on the flip side, gains for institutions who exploit inefficiencies. Odean (1999) and Barber and Odean (2000) demonstrate that individuals have poor security selection in US equities and earn returns well below standard benchmarks. Grinblatt and Keloharju (2000), Linnainmaa (2003a) and Linnainmaa (2003b) show individuals lose in the Finnish stock market, while foreign investors trade on momentum and gain. Barber, Lee, Liu, and Odean (2008a) and Barber, Lee, Liu, and Odean (2008b) find, for Taiwanese stocks, that there are systematic and economically large losses for individuals and gains for institutions. Several papers suggest that institutions are momentum traders in the United States, including Campbell, Ramadorai, and Schwartz (2009) on the New York Stock Exchange, Griffin and Topaloglu (2003) on Nasdaq 100 stocks and Grinblatt, Titman, and Wermers (1995) on NYSE and AMEX stocks. This paper advances the literature by directly relating losses for uninformed traders to a specific market inefficiency that informed traders exploit, and by associating the behavior of investors with the anomaly as the market evolves. The paper proceeds as follows. Section 2 presents the institutional background, data and descriptive statistics, and Section 3 assesses the returns to momentum strategies. Section 4 analyzes large traders returns and relates them to momentum; and Section 5 investigates the cause of momentum. Section 6 discusses results, and Section 7 concludes.
Institutional Background, Data and Descriptive Statistics
While India historically had commodity futures exchanges, by the 1970s, markets for virtually all commodities had been banned, as part of the agenda of the socialist government to exert influence on the prices and quantities of goods produced and consumed in India. In the 1980s and 1990s, the Indian government slowly allowed the creation of small regional exchanges, which were poorly regulated and technologically backward. These exchanges, located in the main production area for the commodity traded, focused on one or two agricultural commodities and attracted a small clientele of agricultural traders specializing in the commodity.5 In 2003 a major shift in the Indian regulatory landscape permitted the launch of modern electronic “multi-commodity” exchanges, and contracts in non-agricultural commodities and 54 “essential” agricultural commodities, which had previously been prohibited. The legislative change led to the creation of NCDEX and two competitor exchanges, the Multi-Commodity eXchange (MCX) and the National Multi-Commodity Exchange (NMCE) that operate similarly to commodity futures exchanges in established markets (see Appendix A for the market rules of NCDEX). These organizations reached out to two new classes of traders, equity brokers and Indian retail investors, who are important investors in India and major participants in equity derivatives (Sarkar, 2006). The rules, however, prevented foreign individuals or institutions, banks, and any institutions using outside investors’ money from trading. The major equity brokerages established proprietary trading and commodity broking businesses, which were advertized extensively and provided market access to retail investors through more than 20,000 terminals in over 800 towns and cities (of Consumer Affairs, 2008). As overseas futures and options trading was illegal, the majority of these new investors had little or no experience in trading commodity derivatives. NCDEX quickly became the dominant domestic exchange for agricultural futures and an important global exchange. From 2002 to 2007 there was a twenty-fold increase in the trading volume of agricultural futures in India, almost 80% of which was due to NCDEX. By 2005 the exchange had become the third largest agricultural futures market in the world (UNCTAD, 2006). In 2006 NCDEX was the main exchange globally for 11 out of its 24 principal agricultural commodities. A combination of the nationally-prominent success of the commodity exchanges in India and high global agricultural commodity prices led to a regulatory backlash and a move away from liberalization.6 Crop prices are an important political issue in India, and the Communist Party 5
See Kolamkar (2003) and Thomas (2003) on the history of commodity futures trading in India; Sarkar (2006) for a comparison of commodity futures markets with other Indian derivatives markets; and Naik and Jain (2002) for a thorough description of the regional exchanges. Naik and Jain (2002) estimate a total of 3,105 customers in 12 exchanges in 2000. 6 Although global commodity prices soared, the average return in the sample is not statistically significantly
was a power broker in a weak government coalition during the period studied. From 2005, and continuing through 2008, the Futures Markets Commission (FMC) imposed measures to dampen speculation, including a series of increases in margins, introduction of new margins, and decreases in position limits and price limits. The government banned trading in black lentil, pigeon pea, rice and wheat in January and February 2007 during the sample period.
Details of the exchange data
The main data are the trading records of NCDEX from its inception on December 15, 2003 to the end of March 31, 2008. The dataset is a complete record of trading activity, consisting of details of the contract traded, transaction price, time (to the second), and account and member identifiers for 183 million trades. Each account can be classified as a client account or member proprietary account, and there are 159,273 client accounts and 772 member accounts (835 total members).7 Despite making up only a small fraction of total accounts, member accounts comprise 24.6% of the capital employed during the sample. I aggregate to the account-commodity-day level, which results in 30.1 million observations. For a subset of the accounts, the data include information on institutional type, location and date of birth. These data are collected by members on behalf of the exchange for trading surveillance. Institutional type and location are available for over 80% of the accounts and over 90% of the capital employed, while date of birth is available for only 53% of the accounts and 37% of the capital employed. I use the data to construct four sets of dummy variables. First, I categorize the institutional type of each account as household, private limited company, other business, or other, which includes accounts with no clear classification and those missing information.8 Institutional type may be recorded with error, as some companies may be misclassified as households.9 Second, I create indicators for settlement size: major city, large city, medium-sized city, or smaller settlement.10 Third, I construct an indicator for whether a commodity is “local” to a trader, defined as whether a 100km radius around the trader has more than twice the national average agricultural production greater than zero. Many of the commodities traded on NCDEX were not subject to major price increases, and Indian government intervention in spot markets prevents Indian prices from rising in step with global prices. 7 Information from the exchange shows there are 1.1 accounts per unique client. Unique client information was not available and the analysis is at the account level. 8 “Household” includes accounts identified as individuals or Hindu United Families, “other businesses” includes sole proprietorships, partnerships and co-operatives. 9 Almost one third of the 500 largest traders as measured by capital employed are categorized as households, and 27 accounts are classified as both households and members, an unlikely combination. Location and date of birth are believed to be accurate. 10 Major city means one of the 8 major Indian metropolises, large city mean over 1 million population, medium-sized city means population between 100,000 and 1 million. I use residential zip code for households and office zip code for companies.
density (see Hunt and Serban, 2009).11 Fourth, I separate traders based on their age on Jan 1, 2006 into three groups of approximately equal size: young (under 30), middle-aged (from 30 to 44), and old (45 and older). The exchange also provided a recording of the information steam sent to members from July 14, 2002 that contains the best bid and best ask. There are a small number of gaps in the data, but for most minutes of the day, there are bid and ask observations for each contract.12
Measurement of returns and capital
The main issue in measuring returns is how to capitalize positions. The standard methodology in the literature on commodity futures is to assume investors “fully collateralize” and hold capital equal to the notional value underlying the contract (see Bodie and Rosansky, 1980, Gorton and Rouwenhorst, 2006, and Erb and Harvey, 2006). I also make this assumption. To calculate capital in order to measure returns for traders, I consider within-day trading positions, as NCDEX requires that accounts hold margin to cover positions at all times. If, instead, I only considered end of day positions, then capital would be incorrect; day traders, for example, would not hold any capital and have infinite returns. I assume traders hold sufficient capital to cover their maximum absolute within-day position, which is Ndmax
= max |Nd,T | = max |Nd−1 + T ≤T¯
where Nd,T is the within-day position after the T th trade for a particular trader on day d, Nd−1 is the net position (negative if short) in number of contracts at the end of day d − 1, ndt is the volume of transaction t (negative if a sale) on day d and transactions run from 1 to T¯. I value the position at the final transaction price during the day of trade, and so capital for day d, for a particular contract and trader, is
Cd = Ndmax ∗ Pd where Pd is the last traded price.13 I apply this rule for each account and for each contract separately, and so do not reduce capital for spread positions. This measure of capital is close to 11 This measure is a generalization of the of the Coval and Moskowitz (2001) “local” metric to the case where the security has a characteristic distributed in two-dimensional space, as opposed to a discrete location. 12 The time is recorded using a different clock than that used by the trades data (different computers), so times in the two data sets do not match. 13 For ease of calculation I do not calculate the maximum price separately for each trader during the period that they held positions, as this would not affect results.
the end of day value of open interest, with the main difference being that the value of open interest does not consider within-day positions. As traders mark to market, they receive profits daily, and these profits are clearly defined. The profits and returns in this paper are excess profits and excess returns as traders also receive interest on their capital, which I do not include in these calculations. If Pdt is the price for transaction t for a trader, then her profit in a given contract during day d is
πd = Nd Pd − Nd−1 Pd−1 −
ndt ∗ Pdt
Namely, excess profit is the notional value of the position at the end of the day less the notional value of the position at the beginning of the day and net trading activity during the day. Excess return is simply
In order to the calculate returns for trading rules or run vector auto regressions, I create daily returns series. The analysis focuses on liquid commodities, as reasonable trading rules would presumably only consider these securities. I define a commodity as liquid on a given day if at least 40 of the previous 50 days had greater than 100 trades.14 This definition could reasonably have been applied by traders at the time, and it restricts the analysis to 29 commodities. I construct returns using the contract with the highest open interest at the end of the previous day. This choice of contract ensures that returns are calculated with a relatively liquid contract, that the benchmark contract does not change frequently, and that the contract with the greatest capital outstanding is chosen.15 Return is measured as the percentage change in price from the previous trading day, which assumes a fully collateralized long position in the contract. 14
I use this definition because my trading rules use prices from the last 50 days and I want these prices to come from days with sufficient trading. But some days, like Saturday, can have low trading. Hence I only require 40 of the last 50 day to have greater than 100 trades. I consider the number of trades so that the momentum strategies would be able to place trades. A different cutoff for liquidity would not change the analysis substantially as there are few episodes for any commodity when trading is between 50 and 1,000 trades per day. There are only four examples of commodities becoming liquid and then becoming illiquid in the sample. For two of these examples, pigeon pea and black lentil, the commodities became illiquid because trading was banned by the government. 15 An alternative option is to specify a particular contract, e.g. the contract with the 2nd nearest expiry date, but which contract is most traded varies over time and across commodities. Another alternative is the contract with the highest trading volume, but it would lead to frequent switching of contracts. Lastly, an open-interest weighted price would incorporate prices from less liquid contracts.
As shown in Figure 1, NCDEX grew rapidly for just over two years following its launch at the end of 2003 and then it shrank from the first half of 2006. We can see from the summary statistics for the exchange in Table 1 that capital, trading volume and contracts available were concentrated in agricultural commodities from the inception of the exchange. The exchange launched the majority of new contracts in its first two years. “2007+” indicates that the results combine data for 2007 and first quarter 2008. Figure 2 displays the percentage of capital by year for the ten commodities with the largest capital employed over the whole sample. It shows that by the end of 2004 the exchange had launched nine of these ten commodities and that the main commodities remained relatively stable, despite high growth, accounting for roughly 75% of total capital.16 Table 2 shows the full list of commodities which became liquid during the sample, categorized into three types: agricultural commodities with an international futures market (“international agri commodities”); agricultural commodities without an international futures market (“domestic agri commodities”) ; and metals and energy commodities, all with international futures markets. We can see that India produces over 50% of global supply for all but three of the domestic agri commodities.17 For 10 of these 13 commodities NCDEX is the main exchange globally. The domestic agri commodities are traded more heavily relative to capital and have a higher standard deviation of return than the international agri commodities. Table 3 presents summary statistics on traders categorized as small, medium or large, based on their average position size, averaged over all commodities, during the whole sample period. I use this metric for summary statistics because it is static for each trader, unlike the proxy for informedness that I use later, although the two are similar. We can see that small traders account for the majority of participants, almost 80%, but for only 8% of capital, and that large traders account for only 2% of participants but 50% of capital. Traders of all sizes have concentrated portfolios with roughly 80% of capital invested in the main commodity, transact between two and three commodities each month, and turn over their positions approximately 10 times per month. The high turnover suggests that hedging accounts for only a small percentage of market activity.18 There are also notable differences across the trader size categories. First, the average number 16
The ten largest commodities over the whole period made up 7 of the top 10 commodities in 2004, 8 of the top 10 in 2005, 8 of the top 10 in 2006, and 9 of the top 10 in 2007+. 17 Production is measured by the agricultural production of the underlying commodity, e.g. cotton for cotton oilcake and soybean for soy oil. 18 In a paper using comparable data for the US heating oil market, Ederington and Lee (2002) find that floor traders (speculators) turnover 19% of their positions each day, while refiners (who, in part, are hedgers) turnover 9% of their positions each day. Even the least frequent traders on NCDEX are trading much more frequently.
of months on the exchange is much less for small traders, 4.5 months compared to 8.1 months for large traders.19 Second, the average small trader holds long positions over twice as often as short positions (41.6% and 19.3% of the time) while the average large trader holds short or long positions almost equally. If we believe traders who hold short positions are the predominant hedgers, this may indicate that small traders are hedging less.20 Large traders are also much more likely to hold a contract to expiry which may indicate that they have spot commodity holdings. My discussions with brokers and the exchange provide little evidence that many farmers or manufacturers use NCDEX to hedge. However, there is no hard information on hedging activity. Agricultural trading firms are the predominant market participants that deal with the physical commodity and range from small businesses specializing in one commodity to large businesses trading in many commodities, such as the Indian subsidiary of US giant Cargill. It appears that a limited number of manufacturing businesses and sparingly few farmers hedged using futures during the sample period (a view also propounded by the of Consumer Affairs (2008)).21
The Performance of Momentum Trading Strategies
This section tests if an anomaly existed on NCDEX, and if so, how it evolved as the exchange matured. I analyze the profitability of momentum strategies that adopt the following rule, as used by Erb and Harvey (2006): buy all commodities that had positive returns and sell all commodities that had negative returns over a given period in the past, equal-weighting commodities. I call the given period in the past the “determining period”. The determining period can begin as as long as 50 trading days in the past (approximately two calendars months given a six day trading week) or as short as 2 trading days in the past. All transactions for the momentum strategies occur at the end of the day, and portfolios are rebalanced daily.22 I use the Erb and Harvey (2006) weighting as there are only a small number of securities and this rule maximizes diversification. The results hold for other weighting schemes used in the literature 19
This metric is calculated for all traders on the exchange including those traders who remain trading at the end of the sample and so is biased downward. 20 The normal backwardation theory of Keynes (1930) argues that shorts are more likely to be hedgers, based on the argument that shorts in agricultural futures markets would be farmers and investors would take long positions. See Gorton and Rouwenhorst (2006) for recent evidence in favor of normal backwardation in US markets. 21 The small agricultural production of the average farmer relative to futures contract size (the average farmers has a plot size of less than two hectares) and the complexity of trading requirements (such as daily mark-to-market) inhibit ordinary farmers from participating on NCDEX. sign(rc,d−j,d−k ) 22 , where If there are N liquid contracts on day d, then the weight of commodity c is wcd (j, k) = N 1 −1 d − k to d − j is the determining period. Weights can thus be N or N , and the sum of absolute weights equals one, PN PN c=1 |wcd (j, k)| = 1. Returns for the momentum portfolio are rmd = c=1 wcd rcd .
on momentum in commodity futures.23 The duration of the determining periods used here are shorter than those of most studies on momentum in the US in equity or commodity markets, which typically consider determining periods from one month up to one year in length.24 The motivation for considering shorter durations is that at the inception of the exchange few services existed that provided market news, such as Reuters Newswire. It is plausible that, over a period of days, the market inefficiently incorporated information such as excess rain in the fields or produce arriving at markets. In conversation, brokers have suggested that prices were slow to respond to information and price trends existed in the early days of the exchange. I first examine strategies for which holdings on a given day are based on a determining period that finishes at the end of the previous trading day. For example, my holdings today, before the close, from a strategy with a determining period that is 10 days long, are set by transactions that occurred at the end of the last trading day, because the strategy rebalances daily. At the end of the last trading day, I fixed long positions for commodities that had positive returns, and short positions for commodities that had negative returns over the previous 10 day period. I label this strategy a “1-11 days” strategy. I consider “1-2 days,” “1-11 days,” “1-26 days,” and “1-51 days” strategies. I then examine strategies for which holdings on a given day are based on a determining period that finishes before the end of the previous trading day. I consider “2-11 days,” “11-26 days,” and “26-51 days” strategies. I initially present results without controlling for risk and then control for risk using a three-factor model based on Miffre and Rallis (2007). Specifically my model is
rmd = α + βe (red − rf d ) + βb (rbd − rf d ) + βc (rcd − rf d ) + d where rm is the return of the momentum portfolio, rf is the return to 91-day Indian T-bills, and re , rb and rc are the returns to the S&P CNX NIFTY total returns index, NSE (National Stock Exchange) government securities composite total returns index, and a long-only equal-weighted 23
The two alternative weighting schemes analyzed are based on Miffre and Rallis (2007) and Wang and Yu (2004). The Miffre-Rallis weighting scheme, based on the approach of Moskowitz and Grinblatt (1999) and Jegadeesh and Titman (2001), ranks commodities into quintiles based on prior returns and goes long commodities in the top quintile and short commodities in the bottom quintile. The Wang-Yu weighting scheme, based on Lo and MacKinlay (1990) and Lehmann (1990), weights commodities proportional to the difference in the prior return to the market average return. Results available on request. 24 Studies on US commodity futures examine momentum for portfolios formed on the basis of the past month of returns up to the past year (Erb and Harvey, 2006; Miffre and Rallis, 2007). Studies on US equity markets also examine momentum in portfolios based on a longer period of return, for example Jegadeesh and Titman (2001) consider portfolios formed based on price changes over the previous six months.
investment in all liquid NCDEX commodities. I conduct the analysis of trading costs from the perspective of the members and do not consider brokerage fees. I initially assume that the momentum strategies buy and sell at the actual transaction prices recorded, and I increase transaction costs in the robustness checks. There are two additional considerations. First, since I measure return using transaction prices, bid-ask bounce will cause short-term negative autocorrelation of returns. However, the bias is small and goes against the finding of momentum. Second, the futures exchange imposes limits on the percentage price movement allowed from the previous day’s settlement price. Limit moves can be important when studying inefficiency, as they prevent prices from fully adjusting to new information (see Roll, 1984). Although, in practice, there is some trading after circuit breaks are triggered, I assume strategies cannot transact on a day with a limit move. Strategies can enter or exit positions only on the first day afterwards when there is no limit in place at the end of the trading day. Because there is some uncertainty about when limit moves occur in the data, I consider different rules to identify when limit moves occur in the robustness checks.
Table 4 displays the basic results for momentum strategies over the entire sample for the 1-2 days, 1-11 days, 1-26 days and 1-51 days strategies. The 1-2 days strategy earns annualized excess returns of 28.63%, which is highly statistically significant with a t-statistic of 6.8, and with a Sharpe ratio of 3.36. The 1-11 days strategy earns returns that are slightly larger, at 30.56% per annum, although the difference is not significant. While the returns for strategies based on longer periods are smaller, for 1-26 days and for 1-51 days the excess returns are 20.19% and 7.50% respectively. Thus, the results show that, over the sample period as a whole, there was short-duration momentum. The steeply declining returns for longer determining periods suggest that, in addition to shortterm momentum, there might have been medium-term reversal. I investigate this hypothesis in the first column of Table 5 by splitting the 2-51 days period into three sub-periods: 2-11 days, 11-26 days and 26-51 days. The excess return for the 2-11 days strategy is 23.28%, with a Sharpe ratio of 2.64; there is substantial momentum beyond one day. The 11-26 days strategy has statistically significant positive profitability, but the return is greatly reduced. The 26-51 days strategy shows medium-term return reversal; it earns negative returns, of -9.24% a year, which is significant at the 5% level. Because momentum strategy profitability is markedly high and highly statistically significant, I investigate how persistent this profitability was across the sample period. In the remaining columns of Table 5, I examine performance separately for each year. As new commodities are introduced by the exchange and subsequently become liquid, the set of commodities that the trading rule invests
in becomes larger. I find a clear and consistent pattern for the two shorter-duration strategies: returns are higher in the earlier years and fall steadily over time. In 2004 the excess returns are particularly high: the 1-2 days and 2-11 days strategies earn excess returns of 55.2% and 33.8% per annum. The returns are lower in 2005 and 2006, and then decrease again in 2007, when the 1-2 days and 2-11 days strategies earn excess returns of 11.3% and 15.7%. There are no obvious trends in the results for longer-duration strategies: the 11-26 days strategy tends to earn positive though statistically insignificant returns, and the 26-51 days strategy tends to earn negative and statistically insignificant returns. The steady decline in returns for the shorter duration strategies is compatible with adaptive markets, but would not be predicted by efficient markets, unless there was a steadily decline in the risk premium over time. The very high t-statistics and positive returns for these strategies, however, suggest that the strategies may not have been risky. In Table 6 I present the results after controlling for risk using the three-factor model. For the 1-2 days strategy, we see economically small negative coefficients on the equity and bond index returns, of -0.017 and -0.036. For the 2-11 days strategy we see a larger coefficient on the long-only commodity portfolio. Nonetheless, the α coefficients for both strategies are almost identical to the coefficients without the risk factors. For all four strategies the factor loadings are slight. The lack of partial correlation between commodity futures momentum returns and the equity and bond returns is similar to that found in the US by Miffre and Rallis (2007).25 In short, I find positive and highly statistically significant short-duration momentum which declines steadily through the first four years of the exchange, and which cannot be explained using conventional models of risk in commodity futures.
Robustness: Transaction Costs and Limit Moves
The basic results calculate returns using transaction prices from the exchange but do not assume any additional costs. Members face two types of transaction cost: the bid-ask spread and exchange fees. As the average bid-ask spread is 0.21% while exchanges fees are between 0.004% and 0.006%, I focus on the bid-ask spread. If traders use markets orders they pay costs of approximately one half the bid-ask spread, while if they trade with limit orders they gain approximately one half the bid-ask spread.26 By using transaction prices, the basic results implicitly assume the trader transacts with market orders half the time, and limit orders the other half of the time. While this assumption appears reasonable, as the exchange is an electronic limit order market and there are no market makers, traders using a strategy that mandates a transaction by a certain 25
In contrast to their findings, however, I find only a small relationship between momentum strategy returns and the long-only commodity index returns, which is perhaps unsurprising given that they analyze strategies with a much longer determining period. 26 On NCDEX most orders are initiated by marketable limit orders. When I use the term market orders I refer to orders that are technically market orders as well as marketable limit orders.
point in time may need to use a higher share of market orders than limit orders.27 I therefore consider two more scenarios: one in which the trader uses market orders 75% of the time and one in which the trader uses market orders 100% of the time (average transaction costs of 25% and 50% of the bid-ask spread, respectively).28 Since the momentum strategies transact at the end of the day, I measure the bid-ask spread at the last available quote. Average bid-ask spreads are approximately constant as the exchange matures.29 There are three main results from Table B.1. First, as spreads remain roughly constant over time, the marked decrease in profitability as the exchange matures remains. Second, while average returns necessarily reduce, the 1-2 days and 2-11 days strategies remain profitable in 2004 even with the highest level of transaction costs. Third, as the 1-2 days strategy churns roughly half its portfolio each day, this strategy is most affected by adding transaction costs, and is no longer the most profitable strategy. I next consider whether the limit moves might be causing the momentum result. I identify limit moves using past price changes and copies of exchange “circulars” (rules that it circulates to its members). As there is some uncertainty about whether the circulars contain every rule change, I adopt a “base case,” shown above, which uses the rules in the circulars and two further cases, an “unadjusted” case, which assumes no limit moves and a “> 4% removed” case, which assumes a limit move is in place for all price changes greater than 4%. I explain in more detail how I identify limit moves in Appendix B. If there is a limit move, portfolio weights for the commodity are not changed until the next day when there is not a limit move. I consider the “unadjusted” and “>4% removed” cases in Table B.2. Columns 2 and 3 show little effect from changing the rules for identifying limit moves. If limit moves constrain prices, the excess returns for short duration strategies should be higher in the unadjusted case, but although returns are affected, the magnitude of effect is miniscule. For example, the return for the 1-2 days strategy increases by 0.78%, a statistically insignificant change. Column 3 shows the results for the “> 4% removed” case, and although returns increase relative to the base case for some strategies, the changes are not statistically significantly different. Limit moves are not an issue.30 27
For a model where informed traders must transact by a particular point in time and use a mix of market and limit orders, see Pagnotta (2008). The models of Parlour (1998), Foucault (1999), Goettler, Parlour, and Rajan (2006), and Foucault, Kadan, and Kandel (2005) also allow for a mix of market and limit orders. For experimental evidence see Bloomfield, OHara, and Saar (2005) PN 28 For each day, the fraction of the portfolio changed is c=1 |wc,d+1 − wc,d |, and the total transaction cost for the two scenarios is 25% or 50% multiplied by the bid-ask spread multiplied by the change in weighting for the P commodity, summed over all commodities, or N c,d+1 − wc,d | = 1 where scd is the bid-ask spread and k is c=1 kscd |w P 0.25 or 0.5. Thus, return after transaction costs is rmd = N 1 wcd rcd − kscd |wc,d+1 − wc,d |. 29 In years 2004 through 2007, the spreads were 0.210%, 0.160%, 0.211% and 0.233% on average 30 Why might this be the case when Roll (1984) found strong effects of limit moves? In Roll’s market slightly over
Large Trader Returns, Share of Capital and Momentum In the previous section I found short-duration momentum and medium term reversal with mo-
mentum declining steadily from 2004 to 2007. The pattern of steady decline in returns, and their high statistical significance after adjusting for risk, is more consistent with the adaptive market hypothesis than the efficient markets hypothesis. However, evidence that informed traders profited from momentum, especially during 2004 and 2005, would further support the adaptive markets view. In this section I present such evidence and also shed light on the adaptive mechanisms that lead to market efficiency.
Recent position size as a proxy for informedness
Unusually, this data allows for the identification of individual traders. I can construct metrics for informedness based on their trading positions, and use these metrics in addition to informedness measures based on the demographic data. As the average trader remains on the exchange for only a few months and new traders arrive continually, I define a rolling measure of size. Recent position size is the average capital employed by the trader in that commodity over the previous 24 trading days (approximately four trading weeks). The average is only taken over days that the trader had a position, and if the trader did not trade, then the capital employed on the current day is substituted as the size measure. I categorize investors as small, medium or large if their recent position size is less than Rs1m ($22.9k), between Rs1m and Rs10m ($22.9k to $229k), or greater than Rs10m ($229k) respectively. These cutoffs points are arbitrary but are chosen at round numbers so that the medium and large groups have a similar share of total capital, and the small group has the majority of investors. Investors can change size category over time and across commodities, as they are categorized daily and separately for each commodity. There are two rationales for recent position size as a proxy for informedness. On the one hand, from a dynamic perspective, in a market which evolves from inefficiency towards efficiency, traders that have successful strategies can be expected to grow in size and vice-versa (Luo, 1998). Moreover, across commodities, traders can be expected to tilt their portfolios towards those securities for which they are more informed (Cohen, Polk, and Silli, 2008). On the other hand, from a static perspective, we may expect small traders to be less informed. In the US equity market, the evidence appears to suggest that individual investors make gross losses (Odean, 1999) while institutional investors make (small) gross gains (Cohen, Polk, and Silli, 2008). On NCDEX there are sizeable numbers of 10% of the trading days resulted in a limit move and there are many instances of consecutive limit moves. In contrast for the Indian commodities, since production is more widespread than for oranges, price movements are less abrupt, and there are 245 limit moves, or 1.0% of trading days, and few days with consecutive limit moves. In short, limit moves are less common and limit moves predict further limit moves less than in the OJ market. A second reason is that the strategies in this paper invest in all commodities, so effects of any individual limit move will be small.
retail traders, who may be less commodity-specific information than agricultural trading firms or commodity brokers. Many papers use similar metrics in the literature and find a positive relationship between size and return.31 First, a series of papers use trade size as a proxy for type of investor. Lee and Radhakhrishna (2000) allocate trade sizes above a certain cutoff to institutions and below another cutoff to individuals, and several other authors use a similar methodology (see Malmendier and Shanthikumar, 2007 and Hvidkjaer, 2006). As institutions have incentives to disguise their trades (see Chakravarty, 2001), Campbell, Ramadorai, and Schwartz (2009) develop a more sophisticated methodology using information from the distribution of trade size. Second, many studies conclude that institutions possess superior information to individuals (e.g. Sias and Starks, 1997, Sias, Starks, and Titman, 2006). Third, literature in the forex market finds large traders possess superior information (Evans and Lyons, 2007, Peiers, 1997 and Moore and Payne, 2009). Most directly relevant to this paper, Menkhoff and Schmeling (2009) construct a measure for trader size using exchange data containing trader identifiers and find large traders earn higher returns. While their metric is similar to mine, these authors analyze only nine days of data and their size measure is based on trading volume over the whole sample. As their metric is not constructed prior to the period during which performance is evaluated, causation may run in the opposite direction, namely, traders that perform well may increase in size. A rolling measure of recent position size is preferable to a number of alternative proxies for informedness. First, informedness metrics using trading over the whole sample are inappropriate because measures should be formed before the period over which performance is measured, to prevent reverse causation. Second, one could use demographic data, in particular institutional type, to create a measure of informedness. However, some companies may be poorly informed, and institutional type appears to misclassify some companies as individuals in these data, as discussed earlier. I compare recent position size to institutional type in the results below. Third, one could categorize investors using past returns. Recent traders, however, will have a much wider distribution of returns due to the small sample length than traders with more experience. So more recent traders will tend to be classified as particularly well or poorly informed. Fourth, one could use size at the beginning of trading, but this might be unduly influenced by the initial attributes of the trader rather than informedness. Fifth, one could use different cutoffs or rankings to determine size. In fact, the results do hold following such changes as discussed in robustness checks below. 31 Research on the mutual fund industry finds that fund size is negatively related to performance. Chen, Hong, Huang, and Kubik (2004) suggest that this result can be explained by the price impact of large funds and organizational diseconomies of scale.
Large trader returns and share of capital
Table 7 shows the results of a regression of annualized excess return at the trader-commodity-day level on indicators for recent position size and five sets of indicators for other trader characteristics, clustering the standard errors by commodity group and trade date (Thompson, 2009). Column 1 shows results from an equally-weighted regression, and as the missing category is for small traders, the coefficient on the constant shows that small traders lose 11.9% per annum. Medium traders earn 5.3% per annum more than small traders, and the difference is highly statistically significant, though these traders still lose on average. Large traders earn 14.4% more than small traders and gain 2.5% on average. To analyze the relationship between size and returns in greater detail, I regress annualized excess daily return on a greater number of size category indicators and show the results in Figure 3. Return is close to linear in log size for the majority of the range of log size, and for the parts of the range where this does not hold, very small and very large traders, there are relatively few observations. The histogram shows that the majority of capital is held by large traders. I capitalweight observations which means that, because total returns must sum to zero, the sum-product of returns and the % of capital is zero. I find a robust positive relationship between recent position size and returns, as one would expect if recent position size were a good proxy for informedness. Column 2 of Table 7 shows the results from regressing return on the five sets of indicators for other trader characteristics discussed earlier.32 The coefficients on all indicators are substantially smaller than those on the size indicators though some are significant. First, members perform significantly better than clients, but the difference is only 2.0% per annum. Second, businesses perform between 3.5% and 5.9% better than households or unclassified accounts. Third, traders identified as old (> 45 years) earn 0.6% more than traders identified as young or middle-aged. Fourth, traders in larger settlements earn 6.7% greater returns, with traders in major cities earning the highest returns. Fifth, a trader being “local” to a commodity is associated with increased return of 4.1%, similarly to Hunt and Serban (2009). Column 3 reveals that the coefficient on large and medium size indicators remain sizeable, even after controlling for the other trader characteristics. Column 4 changes to capital-weighting to assess the aggregate portfolio return for small, medium, and large traders, because the coefficients of equal-weighted regressions may be skewed by a great number of traders with very small amounts of capital. But this change has no qualitative effect. In Table B.3 I present the results of a regression of returns on proxies for informedness based on alternative definitions of size.33 There remains a robust relationship between size and return using 32
The omitted categories are client, unclassified accounts, young traders, traders in settlements with less than 100,000 inhabitants, and non-local traders respectively. 33 Each proxy is calculated from average size over the previous 24 trading days, and I use log size because the
these alternative specifications. Although not shown, alternative definitions of size also produce similar results, e.g. defining large traders as the top 5% by position size. While I consider risk further in the next subsection, three facts suggest risk does not explain these results. First, as we saw in Table 3, the average trader turns over their position every three days, suggesting that a low percentage of positions are for hedging purposes. Second, large traders may actually be more likely to hedge, as they are more likely to hold contracts to expiry or hold short positions, which are indicators of spot risk exposure. Third, the magnitude and consistency of returns suggest risk is an unlikely explanation as returns do not appear risky. Panel A of Table 8 shows how returns for traders of different sizes change as the exchange matures. Return is total profit across all the traders divided by total capital, equivalent to the capital-weighted average return; thus, the coefficients are similar to those in Column 4 of Table 7. I include only liquid commodities. Positive returns for large traders decrease monotonically with time, as does the magnitude of negative returns for medium and small traders.34 Large traders earn sizeable returns in 2004 of 23.0% per annum, which decrease in 2005 to 7.7%, and by 2007 to 2.6%. Medium traders experience similar declines in their losses. In contrast, though small traders’ losses also decrease, they still experience considerable losses in 2007 of 11.8%. I find that the pattern of large trader returns is similar to the pattern of momentum strategy returns, high at the outset of the exchange followed by a steady decline. The greater decline in losses for medium traders, as compared with small traders, is consistent with medium traders learning more quickly about market inefficiencies. Panel B of Table 8 presents the corresponding breakdown of the share of capital. As the regression in Panel A is capital-weighted, the sum product of returns and capital share is equal to zero. As the exchange matures, the share of capital held by large traders increases substantially, while that of small and medium traders falls. The share of small traders decreases from 14.3% of total capital in 2004 to 7.3% in 2007. Figure 4 shows capital share by quarter. Part of the reduction in small and medium trader capital appears to be that small and medium traders exit the market more rapidly, as shown in Table 3, although size is defined differently in this table.
Decomposition of large trader returns
While the previous subsection shows large trader returns decline as the exchange matures, it does not demonstrate that these returns are related to the momentum strategy returns of the last section. In this subsection I focus on that relationship. relationship between size and return is approximately log-linear, as shown in Figure 3. 34 Net profits across the three categories must sum to zero over time, as exchange trading is zero sum.
My methodology aims to decompose large trader returns into a component correlated with momentum strategies and an orthogonal component by regressing large trader returns on momentum strategy returns. Essentially I treat the momentum strategy returns as factors and estimate the extent to which large trader returns load on these factors. First, I separate large trader returns into “inter-day returns” and “within-day returns.” Since the momentum strategies that I study transact daily based on returns over some days, and do not use within-day price information, we should not expect large traders’ returns from within-day trading to correlate with the returns from these strategies. In fact, as expected, these returns series do not correlate. Before decomposing large trader returns into momentum components, I separate large trader profits into “inter-day profits” and “within-day profits.” If Nd is the net position of a trader in a given contract at the end of day d, and Pd is the price of the contract at the end of day d, then inter-day profits are the profits earned from holding the net position from the end of the previous day unchanged for a whole day: πdinter = Nd−1 (Pd − Pd−1 ) And if ndt is the number of contracts bought in transaction t. Then within-day profit is all remaining profit: πdwithin =
ndt (Pd − Pdt )
Where Nd = Nd−1 +
and Pdt is the price of transaction t. To calculate returns, I divide
profits by total capital. Next, I regress large trader inter-day daily returns on the returns from the 1-2 days, 2-11 days, 11-26 days, and 26-51 days strategies from Section 3 and a constant. If large trader profits correlate with episodes of market inefficiency, I expect large trader inter-day returns to positively correlate with returns from the 1-2 days and 2-11 days strategies. There may also be a positive correlation with the 11-26 days strategy and a negative correlation with the 26-51 days strategy. So if j indexes momentum strategies and rjd is the return for strategy j on day d
βj rjd + d
The main point in question is whether the component of large trader returns that correlates with 21
the momentum strategy returns is sizeable. I thus define “momentum return” to be the average return of this component - the sum of the β coefficients times the average return for each strategy
momentum return =
One can then examine whether the momentum return is a large fraction of the average large trader inter-day return. I define % return explained as % return explained =
momentum return = average inter-day return
j=1 βj r j
If momentum strategies are strongly correlated with large trader returns, then not only should the R2 of the regression be high, but the momentum return and % return explained should be high. Lastly, I assess whether the decline in large trader returns is explained by the decline in momentum strategy returns. Large trader returns can decline because these traders hold positions that correlate with momentum strategies less or because momentum strategy returns decline. To determine the extent to which the latter holds I calculate the return large traders would have made if they kept the 2004 loading on the momentum factors in later years
momentum return (2004 weight) =
Panel A of Table 9 shows that average large trader returns are mostly due to inter-day returns, and that the reduction in large trader returns over time is primarily due to inter-day returns.35 In 2004 inter-day returns are 16.2% per annum and account for over 80% of total returns, and in 2005 they fall to 4.4% with further reductions in 2006 and 2007. The first column of Panel B in Table 9 decomposes large trader inter-day returns into components that correlate with the returns of momentum strategies. It shows that the combination of simple momentum strategies explains a sizeable component of the large trader returns over the whole sample. Large trader inter-day returns correlate strongly with returns from the 1-2 days and 2-11 days strategies, with highly statistically significant coefficients of 0.05 and 0.07.36 The coefficient on the 11-26 days strategy is positive as expected but not significant, and the coefficient on the 35
The coefficients for large trader returns in Table 8 and this table differ because days are equal-weighted here. As large trader returns are calculated from the net profits and total capital of all large traders, and some large traders will be short while others are long, it is not surprising that the coefficients are small. 36
26-51 days strategy is negative as expected and significant. The R2 of the regression is 0.15 and the % return explained is 66%. A substantial fraction of large trader returns comes from a component that correlates with momentum strategy returns, which shows that large traders do benefit from the anomaly, as hypothesized, and so supports the adaptive markets view. I next decompose large trader returns for each year to ascertain their relationship with momentum strategy returns as the exchange matures. The second column of Panel B in Table 9 shows that there was a particularly high correlation between momentum strategy returns and large trader returns in 2004: momentum return is 9.6%, % return explained is 59%, and the adjusted R2 is 0.27. The results for 2005 are similar, though the coefficient on the 1-2 days strategy decreases substantially. But in 2006 and 2007 the magnitude of the coefficients on momentum strategy returns and the R2 decrease. Over 50% of the average large trader inter-day return is accounted for by a combination of these simple momentum strategies in all years excluding 2007, with the greatest fraction coming from the 1-2 days and 2-11 days strategies. It appears that large traders load on the anomaly less as it declines, which is consistent with the inefficiency declining either because small and medium traders exit the market or because medium traders learn about the inefficiency. At the bottom of Table 9, I analyze how much of the decrease in large trader inter-day returns is caused by a decrease in momentum strategy returns and how much is caused by a reduction in the use of momentum strategies: by calculating returns if the momentum strategy weights in 2004 had remained. I find that, if weights were unchanged, the 2007 large trader momentum return would be 3.55% (down from 9.62% in 2004), whereas it is actually 0.39%. So, the decline in large trader return comes from both a decline in trading positions that correlate with momentum positions and a decline in the return to momentum (as well as a decline in return uncorrelated with momentum).
A VAR of large trader flows and commodity returns
The previous subsection shows that a considerable fraction of large trader average inter-day returns is accounted for by a component that correlates with momentum strategy returns. An alternative approach, which can confirm and clarify the earlier results, is to examine the relationship between large traders’ net position changes and commodity returns. We expect that positive returns predict that large traders have long positions and that positive returns predict that returns will be positive in the short-run (momentum) and negative in the medium-run (reversal). We can also look for evidence of whether flows lead returns, consistent with private information flows causing momentum, or returns lead flows.
I estimate a VAR of large traders’ net position changes and commodity returns with the independent variables lagged up to 50 days. To reduce the number of estimated parameters, I average the lagged variables over four different periods, 1-2 days, 2-11 days, 11-26 days and 26-51 days. To construct large trader daily flow, net position change, I calculate the sum of net changes in number of contracts across all maturities and then divide it by the total open interest across all maturities at end of the previous day. Return is the return for a long position in the highest open interest contract in the commodity. I regress net position change and commodity return in the current period on the lagged averages of each variable over the four different periods r
rcd = α + npccd = αn +
4 X k=1 4 X
βkn rck,d−1 +
δkr npcck,d−1 + rcd
k=1 4 X
δkn npcck,d−1 + rcd
where k indexes the 4 periods over which averages are taken, and c indicates that the regressions are at the commodity level. I pool the data from different commodities and use ordinary least squares with standard errors clustered by commodity group and trading day.37 Campbell, Ramadorai and Schwartz (2009) use a similar specification but with exponentially weighted moving averages with half-lifes of 1, 10 and 25 days. My results are very similar using their specification, but I adopt this specification so that results are comparable to those from earlier subsections. 4.4.2
Columns 1 and 3 of Table 10 display the results of the vector autoregression, with Column 1 showing flows regressed on lagged returns and lagged flows. The coefficients on lagged returns are negative for all lags and strongly significant for all lags greater than one day. Large traders make sizeable position changes the same day that large price changes occur; so, the negative coefficients on returns in the flow equation represent the unwinding of these initial positions. This interpretation is consistent with the results of the previous section, as almost half of the large trader inter-day return that is correlated with momentum strategies was due to the 1-2 days strategy. Therefore, 37
Given the lack of effect of limit moves in Section 3, I do not control for limit moves here. To control for limit moves in the VAR, I could use the methodology of Roll (1984), which creates two sets of independent variables for returns, one for when there are consecutive circuit breaks, and another one when there are not consecutive circuit breaks. All the results hold, and there is only a small change in coefficients.
large traders must have taken sizeable positions the same day as the price change.38 To understand the contemporaneous relationship between flows and returns further, I estimate a structural VAR with contemporaneous returns included in the large trader flow equation (the treatment of vector autoregressions here is similar to that of Griffin and Topaloglu, 2003). While I cannot directly test a causal relationship between contemporaneous returns and large trader flows, further examination offers insight into the relationship. Column 2 of Table 10 shows that there is a strong contemporaneous correlation between returns and large trader net position changes, which is compatible with the interpretation just discussed. A 1 standard deviation increase in return (1.39%) is associated with a 0.32 standard deviation increase in net position in the same day (0.82%). The inclusion of contemporaneous returns increases the adjusted R2 from 0.0126 to 0.113. Column 3 shows short duration momentum and longer term reversal, consistent with earlier results. There is a positive correlation of present return and return from the previous day, even though bid-ask bounce will induce negative autocorrelation and bias against this result. The coefficients on lagged flows are also of note, though the implications are less clear.
Evidence for Gradual Information Diffusion The previous section shows that a considerable fraction of large trader inter-day returns is ex-
plained by momentum, and that the majority is due to short-term strategies, but the cause of momentum and large trader profits on NCDEX remains unclear. In this section I outline competing theories for momentum and a particular candidate theory, gradual information diffusion, and two predictions that come from the model, which are then tested.
A candidate theory and predictions
Two categories of theory explain momentum. On the one hand, momentum may be caused by public information. Barberis, Shleifer, and Vishny (1998) present a theory of momentum based on under-reaction and over-reaction to public information caused by the psychological biases of conservatism and representativeness. Grinblatt and Han (2001) present a theory where momentum is caused by the systematic reaction of traders to past price changes through their response to gains or losses (the disposition effect). On the other hand, momentum may be caused by private information. In the gradual information flow theory of Hong and Stein (1999), momentum is caused by information diffusing slowly through the population, coupled with imperfect filtering of the information of others from prices. In the overconfidence model of Daniel, Hirshleifer, and 38 Moreover, these initial positions must be sufficiently sizeable so that the unwinding of positions in the subsequent 2 to 11 days, shown in Column 1, does not lead to traders reversing their initial positions.
Subrahmanyam (1998) traders become overconfident if private information they have is confirmed by public information flows but not under-confident if the information they have is disconfirmed, leading prices to overshoot before falling back to the correct level. Given the lack of information services in India, the primary candidate for explaining momentum on NCDEX is the gradual information diffusion model of Hong and Stein (2007). In their model, the value of an asset is the sum of two shocks. In the first period, two symmetrically informed people each see a shock - a different shock each - and form expectations of value based on that shock. They do not infer the shock that the other person has seen from the asset price. In the second period each person sees the other shock, and the price adjusts to its true value. When the two shocks have the same sign, the returns in each period have the same sign: momentum occurs. To adapt the model to the current setting, it is necessary to make two additions: a less-informed third person who does not view the original shocks in the first period (but does in the second) and a new commonly-observed component of value. As with the original model, lack of inference from prices of the shocks observed by others leads to momentum. But as the third person is less informed and believes the information he observes, he makes losses. The two (partially) informed traders gain. Private information therefore explains both momentum and the gains for informed traders. How can the market become more efficient? First, informed traders, as they earn profits from trading on their private information signals, can increase their capital deployment against these signals. Second, the less informed investor, as he makes losses from trading on his private information signals, can decrease his capital deployment by reducing trading or by exiting the market. Third, any of the three investors can become more informed. For example, investors may put more weight on the market price they observe in forming beliefs about the value of the asset. These three mechanisms were considered earlier in the paper. Two predictions are specific to private information models of momentum. The first prediction is that traders are informed only on some commodities, those for which they have private information. In contrast, if traders were informed on public information and traded directly on the realization of past prices using technical models, they should be informed across many commodities. The model predicts, therefore, that a proxy for informedness on other commodities does not predict performance on a particular commodity, after controlling for informedness on that commodity. The second prediction is that commodities with a greater fraction of commonly-observed information have lower momentum and wealth transfers. I test if there is more momentum for commodities for which there is less public information by conducting a test similar to that used in equity markets. For US stocks, Hong, Lim, and Stein (2000), Zhang (2006), and Hwang (2008) use low analyst coverage as an indicator for a poor public information. This test distinguishes between momentum
models that depend on private information or public information, but it does not distinguish between the Hong and Stein (1999) gradual information flow model or the Daniel, Hirshleifer, and Subrahmanyam (1998) overconfidence model. I discuss further suggestive evidence for gradual information diffusion below.
Commodity-specific versus aggregate proxies for informedness
To investigate whether information is commodity-specific, in Column 1 of Table 11, I regress return at the trader-commodity-day level on a continuous measure for size, log size, trader characteristic dummies, and a constant, consistent with the regressions in Table 7. Column 1 reconfirms the earlier results: size is strongly related to return. In Column 2, I change to an aggregate position size metric, which is the same as the commodityspecific size measure but based on all the trader’s positions. Aggregate size is also strongly related to return, which is perhaps unsurprising given that traders are concentrated in their portfolios. Column 3 tests whether trader position size in other commodities is related to trading performance in a specific commodity. The coefficient on log aggregate position size shows that this is not the case: after controlling for commodity-specific size, there is no relationship between aggregate trader size and return. This result, coupled with the fact that traders are concentrated in their portfolios, suggests informed traders are not technical traders who detect trends in prices and then trade upon them.
Domestic versus international commodities
No international futures market as a measure of high private information
Using whether a commodity has an international futures market or not as a measure of public information (in lieu of high analyst coverage, commonly used in research on stocks), I conduct a test for whether momentum is higher for commodities with less public information. Table 2 shows which commodities do and do not have international futures markets. When NCDEX was launched, there was limited information available on commodities with no international futures market, such as chickpea or cluster bean seed, and this information was provided by a handful of private companies that specialized in gathering agricultural data and conducting research (Agriwatch or Commodity India are two examples). Coverage of commodities in the Indian news was relatively limited, and none of the major papers had a separate commodities section. In contrast, for agricultural commodities that have an international futures market, such as soybean or wheat, much more information was available. The US Department of Agriculture world agricultural supply and demand estimates covers all but one of these commodities, for example,
but none of the domestic agri commodities. And there is extensive coverage of the international agri commodities in the global financial press, newswires and specialist information providers. After 2003, there was substantial development in information provision. Companies were launched to provide information on commodities, and the existing companies experienced rapid growth. The newswires - Reuters, Bloomberg and an Indian domestic firm called Newswire 18 - covered Indian commodities from late 2004, and over time they expanded their service into more commodities and more in depth reporting. Many major newspapers started a commodities section, such as the Hindu Business Line in 2005. Companies involved in the commodity futures markets also developed their in-house data gathering and analysis. For example, Anand Rathi, a broker, gathers information from partnerships it has developed in the main crop growing areas and produces its own supply estimates through the growing season. So, the relative paucity of information for domestic agri commodities diminished. If momentum is caused by private information, and large traders trade on private information, there are four predictions. First, momentum should be higher for the non-international agri commodities in 2004, but the difference should decrease over time. Second, the same patterns should also be observable in large trader returns. 5.3.2
Table 12 presents returns for momentum strategies for international commodities and domestic agri commodities. As predicted, return is much higher for domestic commodities in 2004. Returns for the 1-2 days strategy were 105.8% for the domestic commodities and 37.1% for the international commodities. By 2007 these returns decreased to 7.3% and 13.7% respectively. The results are qualitatively similar for the 2-11 days strategy, but the return is relatively low for the international commodities, between 13% and 20% throughout the sample, and does not obviously decrease over time. Returns decreased more for the domestic commodities as predicted. Table 13 presents the results for large trader returns. In 2004 large traders earned 29.2% per annum in non-international commodities compared with 13.5% in international commodities. By 2007 these returns decreased to 3.4% and 2.2% respectively. The returns also decreased more for the domestic commodities.
Discussion The results in Section 4 provide evidence for the adaptive markets view. The alternative per-
spective, the efficient markets view, is that the discovery of short-duration momentum on NCDEX must be due to data mining or an inadequate model of risk.
The data-mining explanation, however, would not predict that the anomaly decreases steadily as the exchange becomes more mature. Neither would it predict that large traders make substantial average returns, a large fraction of which is attributable to momentum. The inadequate-model-ofrisk explanation implies that the premium decreases over time either because large traders are more willing to bear risk or because risk decreases. I find, however, that controlling for risk factors does not affect returns for momentum strategies. Moreover, the returns do not appear to be risky as large traders make persistently high returns. Thus the view that large traders earn returns because they are exploiting an inefficiency, and that the inefficiency decreases because of the wealth transfers from small and medium traders to large traders, is more consistent with the evidence. The results also shed some light on the mechanisms which led NCDEX towards efficiency. I find informed traders exploited the inefficiency from the outset of the exchange. But I find little support (see Table 9) that informed investors traded on the inefficiency with greater vigor over time and that this led towards efficient pricing. Instead, I find that less informed traders exited the market at a faster rate than informed traders, and that some less informed traders (medium traders) may have increased their informedness as the market matured, because their losses fall rapidly (unlike those of small traders). The magnitude of the transfers has potential implications for regulatory authorities. The returns assume full collateralization, which is a conservative capitalization method, and in practice the returns for investors are likely to be much larger. Many small traders probably had their capital completely eroded. In dollar terms the net gain to large traders over the sample was $224.7m and the net losses to small and medium traders were $89.7m and $135.0m, large sums in the context of India. Might this require a regulatory response? One potential response is enhanced investor education programs, as practiced in the US, combined with legislation to create a self-regulatory industry association to provide education. The Commodity Futures Trading Commission and the National Futures Association (the US industry association) both provide extensive information to the public.39 An alternative response is to mandate that members provide information to each new client on their existing client accounts, including the typical account performance, balance, and lifespan. The results in Section 5 provide evidence on the nature of informedness and the cause of momentum. Specifically, the results suggest that large traders have commodity-specific private information, and this private information leads to both momentum and large trader profits. The results on aggregate versus commodity-specific size, and the finding of high portfolio concentration, suggest that large traders are informed about only a few commodities. The results on international and domestic commodities are compatible with private information leading to both momentum and 39 The information provided includes explanations of trading goals, risk and contractual obligations, and checklists for retail investors to complete before they begin trading.
large trader returns. As mentioned earlier, a variety of services have developed that provide information to exchange participants for the commodity markets. As private information is profitable, traders have an incentive to increase their information gathering. It is possible that medium traders may have become more informed because of the improvement in the general provision of information. If large trader returns are strongly associated with momentum strategy returns, but these investors are not technical traders (taking positions based on past prices), then in what sense are they “exploiting” the inefficiency? If they trade on private information, and market prices are slow to respond, as in the gradual information flow model, then trading on private information is particularly profitable precisely because of the limited response of market prices to trading. The inefficiency means that information is especially valuable. An alternative hypothesis to gradual information diffusion is that informed traders are overconfident and their beliefs excessively respond to price changes that confirm their private information (Daniel, Hirshleifer, and Subrahmanyam, 1998). In this model momentum occurs because prices overshoot, and so prices must reverse to the original level. The evidence, however, shows that there is only modest price reversal on NCDEX and no greater price reversal in the early years of the exchange. Thus the results better fit a simple model of gradual information diffusion.
Conclusion This paper studies the disappearance of an asset pricing anomaly using detailed data on the trad-
ing positions of investors from a new commodity futures exchange in India, NCDEX. The evidence, which consists of three main results, supports the interpretation that the anomaly represents a real inefficiency and that the market converges towards efficiency because informed traders exploit the trading opportunity. First, following the launch of the exchange, momentum strategies that use prices from the previous two weeks earn substantial returns, which steadily decline over a four year period. Second, large traders, who are believed to be more informed, earn returns that are strongly correlated with the returns to momentum strategies at the expense of the majority of investors who make considerable losses. These returns also decline through the sample. Third, the decline in momentum coincides with faster exit by small and medium traders from the exchange, and it coincides with reduced losses for medium traders relative to small traders, which is consistent with greater medium trader learning about the inefficiency. Large traders also trade momentum less aggressively. These findings also provide some preliminary evidence on the mechanisms that cause the transition to market efficiency. In particular they suggest learning by medium traders and differential drop out by small and medium traders (the uninformed), rather than increased capital deployment by large traders (the informed), caused the trend towards efficiency. 30
This paper also suggests that gradual diffusion of information caused short duration momentum profits on this exchange. I find that investor returns in one commodity are not predicted by recent position size in other commodities, as would be expected if large traders were technical traders investing based on analysis of past prices. This finding, coupled with the high concentration of trader portfolios, suggests rather that large traders receive commodity-specific private information. In addition I find that momentum and large trader returns are greater for agricultural commodities with no international futures market. As these commodities have less public information provision in the early stages of the market, this finding is consistent with large traders receiving news before other traders at the outset of the exchange and with gradual diffusion of this information leading both to momentum and to large trader profits. Given the short duration and high statistical significance of the profitable momentum strategies, it is perhaps unsurprising the opportunities did not persist. The four year duration for which excess profits were available is of similar length to the eighteen month to four year duration for which arbitrage profits were available for US equity index futures contracts in the early 1980s (Chung, 1991; Thomas, 2002). The relatively speedy disappearance of these anomalies contrasts with the persistence of trading profits for other anomalies, such as the post-earnings announcement drift (Ball and Brown, 1968; Bernard and Thomas, 1989; Campbell, Ramadorai, and Schwartz, 2009). Further research is required to uncover exactly why some anomalies persist and other rapidly disappear to determine which adaptive mechanisms are primarily responsible for leading markets towards efficiency.
References Ball, R., and P. Brown (1968): “An Empirical Evaluation of Accounting Income Numbers,” Accounting Research, 6, 159–178. Barber, B., and T. Odean (2000): “Trading Is Hazardous to Your Wealth: The Common Stock Investment Performance of Individual Investors,” University of California Davis Manuscript. Barber, B. M., Y.-T. Lee, Y.-J. Liu, and T. Odean (2008a): “Day Trading in Equilibrium,” U.C. Davis Working Paper. (2008b): “Just How Much Do Individual Investors Lose by Trading,” Review of Financial Studies. Barberis, N., A. Shleifer, and R. Vishny (1998): “A Model of Investor Sentiment,,” Journal of Financial Economics, 49, 307–343. Bernard, V. L., and J. Thomas (1989): “Post-earnings Announcement Drift: Delayed Price Response or Risk Premium?,” Journal of Accounting Research, 27, 1–48. Bloomfield, R., M. OHara, and G. Saar (2005): “The Make or Take Decision in an Electronic Market: Evidence on the Evolution of Liquidity,” Journal of Financial Economics, 75, 165–199. Bodie, Z., and V. I. Rosansky (1980): “Risk and Return in Commodity Futures,” Financial Analysts Journal, 36, 27–39. Campbell, J., T. Ramadorai, and A. Schwartz (2009): “Caught on Tape: Institutional Trading, Stock Returns, and Earnings Announcements,” Journal of Financial Economics. Chakravarty, S. (2001): “Stealth-trading: Which traders trades move stock prices?,” Journal of Financial Economics, 61, 289–307. Chang, E. C., M. Pinegar, and J. Ravichandran (1993): “International Evidence on the Robustness of the Day-of-the-week Effect,” Journal of Financial and Quantitative Analysis, 28, 57–74. Chen, J., H. Hong, M. Huang, and J. Kubik (2004): “Does Fund Size Erode Mutual Fund Performance? The Role of Liquidity and Organization,” American Economic Review, 94. Chordia, T., R. Roll, and A. Subrahmanyam (2005): “Evidence on the Speed of Convergence to Market Efficiency,” Journal of Financial Economics, 76, 271–292.
Chung, Y. P. (1991): “A Transactions Data Test of Stock Index Futures Market Efficiency and Index Arbitrage Profitability,” Journal of Finance, 46, 1791–1809. Cohen, R., C. Polk, and B. Silli (2008): “Best Ideas,” London School of Economics Working Paper. Coval, J. D., and J. Moskowitz, Tobias (2001): “The Geography of Investment: Informed Trading and Asset Prices,” Journal of Political Economy, 109, 811–841. Daniel, K., D. Hirshleifer, and A. Subrahmanyam (1998): “Investor Psychology and Security Market Under- and Overreactions,” The Journal of Finance, 53, 1839–1885. Dimson, E., and P. Marsh (1999): “Murphys Law and Market Anomalies,” Journal of Portfolio Management, 25, 5369. Ederington, L., and J. H. Lee (2002): “Who Trades Futures and How: Evidence from the Heating Oil Futures Market,” Journal of Business, 72, 353–373. Erb, C., and C. R. Harvey (2006): “The Tactical and Strategic Value of Commodity Futures,” Financial Analysts Journal, 61, 69–97. Evans, M. D., and R. K. Lyons (2007): “Exchange Rate Fundamentals and Order Flow,” Working Paper, Haas School of Business, U. C. Berkeley. Fama, E. (1965): “The Behavior of Stock Market Prices,” Journal of Business, January 1965, 34–105. Fama, E., and K. R. French (1996): “Multifactor Explanations of Asset Pricing Anomalies,” The Journal of Finance, 51, 55–84. Fama, E. F. (1998): “Market Efficiency, Long-Term Returns, and Behavioral Finance,” Journal of Financial Economics, 49, 283–306. Foucault, T. (1999): “Order Flow Composition and Trading Costs in a Dynamic Limit Order Market,” Journal of Financial Markets, 2, 99–134. Foucault, T., O. Kadan, and E. Kandel (2005): “Limit Order Book as a Market for Liquidity,” Review of Financial Studies. Friedman, M. (1953): “The Case for Flexible Exchange Rates,” Essays in Positive Economics, University of Chicago Press.
Goettler, R. L., C. A. Parlour, and U. Rajan (2006): “Microstructure Noise and Asset Pricing,” mimeo. Gorton, G., and K. G. Rouwenhorst (2006): “Facts and Fantasies about Commodity Futures,” Financial Analysts Journal, 62, 47–68. Griffin, J.M. adn Harris, J., and S. Topaloglu (2003): “The Dynamics of Institutional and Individual Trading,” Journal of Finance, 58, 22852320. Grinblatt, M., and B. Han (2001): “The Disposition Effect and Momentum,” Yale International Center for Finance Working Paper Series. Grinblatt, M., and M. Keloharju (2000): “The Investment Behavior and Performance of Various Investor-Types: a Study of Finlands Unique Data Set,” Journal of Financial Economics, 55, 4367. Grinblatt, M., S. Titman, and R. Wermers (1995): “Momentum Investment Strategies, Portfolio Performance and Herding: A Study of Mutual Fund Behavior,” American Economic Review, 85, 10881105. Hong, H., T. Lim, and J. C. Stein (2000): “Bad News Travels Slowly: Size, Analyst Coverage, and the Profitability of Momentum Strategies,” The Journal of Finance, 55, 265–295. Hong, H., and J. C. Stein (1999): “A Unified Theory of Underreaction, Momentum Trading and Overreaction in Asset Markets,” Journal of Finance, 54, 21432184. (2007): “Disagreement and the Stock Market,” Journal of Economic Perspectives. Hudson, R., K. Keasey, and K. Littler (2002): “Why Investors Should be Cautious of the Academic Approach to Testing for Stock Market Anomalies,” Applied Financial Economics, 12, 681–686. Hunt, S., and L. Serban (2009): “Informed Trading when Information is Publicly Available: Evidence from Agricultural Commodity Futures,” Mimeo, Harvard University. Hvidkjaer, S. (2006): “A Trade-based Analysis of Momentum,” Review of Financial Studies, 19, 457491. Hwang, B. (2008): “Distinguishing Behavioral Model of Momentum,” Mimeo, Goizueta Business School. Jegadeesh, N., and S. Titman (2001): “Profitability of Momentum Strategies: An Evaluation of Alternative Explanations,” Journal of Finance, 56, 699–720. 34
Keim, D. B., and W. Ziemba (2000): “Security Market Imperfections: An Overview,” D. B. Keim and W. T.Ziemba (eds.), Security Market Imperfections in Worldwide Equity Markets, Cambridge: Cambridge University Press, pp. xv-xxvii. Keynes, J. M. (1930): “A treatise on money,” AMS Press, New York. Khwaja, A., and A. Mian (2005): “Unchecked Intermediaries: Price Manipulation in an Emerging Stock Market,” Journal of Financial Economics, 78, 203–241. Kolamkar, D. S. (2003): “Regulation and Policy issues for Commodity Derivatives in India,” In Thomas, S., editor, Derivatives Markets in India, OUP. Lee, C. M. C., and B. Radhakhrishna (2000): “Inferring Investor Behavior: Evidence from TORQ Data,” Journal of Financial Markets, 3, 83–111. Lehmann, B. (1990): “Fads, Martingales, and Market Efficiency,” Quarterly Journal of Economics, 105, 1–28. Linnainmaa,
“Who Makes the Limit Order Book?
Implications for Contrarian
Strategies, Attention-Grabbing Hypothesis, and the Disposition Effect,” Available at SSRN: http://ssrn.com/abstract=474222. Lo, A. (2004): “The Adaptive Markets Hypothesis: Market Efficiency from an Evolutionary Perspective,” Journal of Portfolio Management, 30, 15–29. Lo, A. W., and A. C. MacKinlay (1990): “Data-Snooping Biases in Tests of Financial Asset Pricing Models,” Review of Financial Studies, 3, 431–467. Luo, G. (1998): “Market Efficiency and Natural Selection in a Commodity Futures Market,” Review of Financial Studies, 11, 647–674. Malmendier, U., and D. Shanthikumar (2007): “Are Small Investors Naive about Incentives?,” Journal of Financial Economics, 85, 457489. Marquering, W., J. Nisser, and T. Valla (2006): “Disappearing Anomalies: A Dynamic Anlysis of the Persistence of Anomalies,” Applied Financial Economics, 16, 291–302. Mehdian, S., and M. Perry (2002): “Anomalies in US Equity Markets: A Re-examination of the January Effect,” Applied Financial Economics, 12, 141–145. 35
Menkhoff, L., and M. Schmeling (2009): “Whose Trades Convey Information? Evidence From a Cross-Section of Traders,” Journal of Financial Markets, forthcoming. Miffre, J., and G. Rallis (2007): “Momentum strategies in commodity futures markets,” Journal of Banking and Finance, 31, 1863–1886. Moore, M., and R. Payne (2009): “Size, Specialism and the Nature of Informational Advantage in Inter-dealer Foreign Exchange Trading,” Working Paper, Warwick Business School. Morgan, C. (2001): “Commodity futures markets in LDCs:a review and prospects,” Progress in Development Studies, 1, 139–150. Morgan, C. W., A. J. Rayner, and C. Vaillant (1999): “Agricultural Futures Markets in LDCs: A Policy Respone to Price Volatility,” Journal of International Development, 11, 893–910. Moskowitz, T., and M. Grinblatt (1999): “Do Industries Explain Momentum?,” Journal of Finance, 54, 12491290. Naik, G., and S. K. Jain (2002): “Indian Agricultural Commodity Futures Markets: A Performance Survey,” Economic and Political Weekly, 36. Odean, T. (1999): “Do Investors Trade too Much?,” American Economic Review, 89, 1279–1298. of Consumer Affairs, M. (2008): “Report of the Expert Committee to Study the Impact of Futures Trading on Agricultural Commodity Prices,” Ministry of Consumer Affairs, Food and Public Distribution, Government of India. Pagnotta, E. (2008): “Trading strategies at optimal frequencies: theory and evidence,” New York University, Working Paper. Parlour, C. (1998): “Price Dynamics in Limit Order Markets,” The Review of Financial Studies, 11, 789–816. Peiers, B. (1997): “Informed Traders, Intervention, and Price Leadership: A Deeper View of the Microstructure of the Foreign Exchange Market,” Journal of Finance, 52, 1589–1614. Roll, R. (1984): “Orange Juice and Weather,” American Economic Review, 74, 861–880. Sarkar, A. (2006): “Indian Derivatives Markets,” The Oxford Companion to Economics in India. Schwert, G. W. (2003): “Anomalies and Market Efficiency,” Handbook of the Economics of Finance, Chapter 15, G. M. Constantinides and R. Stulz (Eds), Elsevier Science, Amsterdam.
Sias, R., and L. T. Starks (1997): “Return Autocorrelation and Institutional Investors,” Journal of Financial Economics, 46, 103–131. Sias, R., L. T. Starks, and S. Titman (2006): “Changes in Institutional Ownership and Stock Returns: Assessment and Methodology,” Journal of Business, 79, 2869–2910. Sullivan, R., A. Timmermann, and H. White (2001): “Dangers of Data-Driven Inference: the Case of Calendar Effects in Stock Returns,” Journal of Econometrics, 105, 249–286. Thomas, S. (2002): “The Saga of the First Stock Index Futures Contract: Benchmarks, Models, and Learning,” Journal of Money, Credit and Banking, 34, 767–808. Thomas, S. (2003): “Agricultural commodity markets in India:Policy issues for growth,” Forwards Markets Commission Report. Thompson, S. (2009): “Simple Formulas for Standard Errors that Cluster by Both Firm and Time,” Available at SSRN: http://ssrn.com/abstract=914002. UNCTAD (2006): “Survey of the World’s Commodity Exchanges 2005,” The World’s Commodity Exchanges: Past - Present - Future. Wang, C., and M. Yu (2004): “Trading Activity and Price Reversals in Futures Markets,” Journal of Banking and Finance, 28, 13371361. Zhang (2006): “Information Uncertainty and Stock Returns,” The Journal of Finance, 61, 105– 137.
Table 1: Exchange Summary Statistics This table provides summary statistics for December 15, 2003 to March 31, 2008. “2004” includes 12 days from 2003 and “2007+” includes Q1 2008. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. This capital measure is slightly higher than the value of open interest. Number of Positions is the number of trader-commodity-day observations. The calculation of all other measures should be clear. For ease of comparison “2007+” does not include Q1 2008 for Number of Traders and Number of Positions. The exchange rate used is Rs43.7 / $.
Agri Metals Energy
1.54 0.25 0.01
1.97 0.32 0.01
1.93 0.17 0.01
(avg daily $bn)
Agri Metals Energy
1.16 0.16 0.01
1.51 0.34 0.01
1.05 0.14 0.01
Agri Metals Energy
36 5 2
37 8 2
32 7 4
Capital (avg daily $bn)
# Traders # Positions (m)
Table 2: Commodity Summary Statistics This table groups commodities according to whether an international market for futures exists and whether the commodity is agricultural. It includes commodities which are liquid during the sample, defined as having greater than 100 trades in 40 of the previous 50 days. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. This capital measure is slightly higher than the value of open interest. The Standard Deviation of Return is calculated from the price change for the highest value open interest contract in the commodity and is expressed in percent. % of World Production is the total of world production for the raw commodity, for example soybean production is used for soy oil. Main Indian Exchange indicates whether NCDEX has the largest trading volume in India during April 2005 to April 2006. Launch Date is the first day the contract was available.
Trade Value Capital % Total % Total
Return Std Dev
Main % World Indian Prodn Exch.
No International Market black lentil castor chickpea chilli cluster bean gum cluster bean seed cumin jjute lentil mint oil pepper pigeon pea turmeric Total/Average
7.2 02 0.2 18.1 1.8 2.0 25.5 4.5 0.1 0.0 0.3 5.5 1.4 1.3 67.9
6.1 07 0.7 12.2 1.8 3.0 12.6 4.8 0.3 0.0 0.6 5.5 1.8 1.8 51.2
1.9 10 1.0 1.4 2.1 1.5 1.9 2.0 1.1 1.1 2.2 1.8 1.6 1.2 1.6
65.0 66 4 66.4 64.2 49.0 80.0 80.0 77.0 65.3 29.3 87.0 24.0 80.0 80.0 65.2
0.0 0.7 0.3 0.6 0.2 2.9 0.0 6.0 3.3 2.1 1.2 17 4 17.4
0.1 1.2 0.4 1.3 0.5 5.7 0.1 8.9 9.0 6.0 3.4 36 5 36.5
1.3 1.2 1.1 1.0 1.1 0.8 1.9 0.8 1.0 0.8 0.9 11 1.1
0.9 14.0 14.0 12.0 2.1 14.1 7.5 3.9 3.9 12.0 11.6 87 8.7
0.3 0.2 6.8 7.1 0.4 14.7
0.3 0.1 4.8 5.8 1.2 12.2
1.5 1.6 0.9 1.6 0.9 1.3
29‐Jul‐04 24‐Jul‐04 13‐Apr‐04 12‐Mar‐05 4‐Feb‐05 27‐Jul‐04 13‐Apr‐04 27‐Jul‐04 22‐Oct‐05 30‐Aug‐05 13‐Apr‐04 9‐Apr‐05 28‐Jul‐04
International Market: Agricultural barley cotton oilcake cotton raw jaggery maize mustard potato soy oil soybean sugar wheat Total/Average
12‐Dec‐06 6‐Apr‐05 5‐Oct‐05 6‐Jan‐05 6‐Jan‐05 15‐Dec‐03 8‐Jul‐06 15‐Dec‐03 15‐Dec‐03 28‐Jul‐04 7‐Jul‐04
International Market: Metals and Energy crude oil copper gold silver steel Total/Average
No No No No No
16‐Sep‐05 7‐Jun‐05 15‐Dec‐03 15‐Dec‐03 15 Dec 03 12‐Mar‐05
Table 3: Trader Summary Statistics This table groups traders according to their average position size across all commodity-day observations. Position size is the average capital the trader held in the commodity in the previous 24 trading days on days they participated in trade. In this table, traders are categorized once using information over the whole sample for the purpose of summary statistics, while in other tables traders are recategorized daily for each commodity. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. Total months in market is calculated for all traders, even if they have not left the exchange by the end of the sample. Long Position is the percentage of days where trader end-of-day net position in terms of number of contracts, summed over all maturities in a commodity, is positive. Short Position and No End-of-day Position are similar. Switch or close out position indicates when a trader changes from being short to long at end of day or vice-versa, or moves from having a position to no position at end of day. Switch direction of position indicates when a trader changes their position from being short to long at end of day, or vice-versa. t-statistics are omitted, but almost all comparisons across groups are statistically significant.
Trader size category
< Rs 1m < $ 23k
Rs 1m to Rs 10m $23k to $ 230k
> Rs 10m > $ 230k
11 24.4 5.3 11.1
135 207.7 8.0 12.3
1,007 871.6 11.2 9.6
4.5 4.4 2.2
7.1 6.1 3.1
8.1 5.0 2.5
% Capital in top 1 commodity, monthly average % Capital in top 3 commodities, monthly average
Long position, % of days Short position, % of days No end-of-day position, % of days
41.6% 19.3% 39.0%
39.1% 27.5% 33.3%
34.8% 34.5% 30.7%
Switch or close out position, % of days Switch direction of position, % of days
Size range (Average position size) # accounts % of total capital Average Average Average Average
capital ($ ’000) trades per month trade size ($ ’000) monthly position turnover
Total months in market Total number of commodities traded Average commodities traded per month
# times hold contract to expiry
Table 4: Returns of Momentum Strategies This table reports average annualized returns for portfolios formed using momentum strategies. The strategies buy commodity futures that had positive returns over an earlier period (the determining period) and sell commodity futures that had negative returns. Commodities are equal-weighted, portfolios are rebalanced daily, and all transactions occur at the end of the day at the last traded price. The table presents results for strategies based on a determining period that finishes at the end of the previous day. The strategies use determining periods of length 1 day, 10 days, 25 days or 50 days, and to mark the beginning and the end of the determining period, the strategies are labeled as 1-2 days, 1-11 days, 1-26 days and 1-51 days. The strategies only apply to commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades. The returns are adjusted for limit moves by prohibiting trade on the day of a limit move. Observations are equal-weighted. Years
Strategies Based on Returns From 1-2 days
return t-stat Sharpe ratio
0.286 (6.79) 3.36
return t-stat Sharpe ratio
0.306 (6.77) 3.36
return t-stat Sharpe ratio
0.202 (4.69) 2.33
return t-stat Sharpe ratio
0.075 (1.86) 0.924
There are 1253 daily observations for each portfolio Absolute t statistics in parentheses Robust standard errors + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001
Table 5: Returns of Momentum Strategies, by Year This table reports average returns for portfolios formed using momentum strategies. The strategies buy commodity futures that had positive returns over an earlier period (the determining period) and sell commodity futures that had negative returns. Commodities are equal-weighted, portfolios are rebalanced daily, and all transactions occur at the end of the day at the last traded price. The table presents results for strategies based on a determining period that began x days ago and finished y day ago at the end of day and are labeled as “x-y days”. The strategies considered are 1-2 days, 2-11 days, 11-26 days and 26-51 days. The strategies only apply to commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades. The returns are adjusted for limit moves by prohibiting trade on the day of a limit move. Observations are equal-weighted.
Strategies Based on Returns From 1-2 days
return t-stat Sharpe ratio
0.286 (6.79) 3.36
0.552 (4.08) 4.49
0.273 (4.04) 4.04
0.294 (3.63) 3.63
0.113 (1.88) 1.69
return t-stat Sharpe ratio
0.233 (5.33) 2.64
0.338 (2.48) 2.73
0.246 (3.55) 3.56
0.227 (2.73) 2.73
0.157 (2.31) 2.07
return t-stat Sharpe ratio
0.095 (2.23) 1.11
0.130 (0.92) 1.01
-0.002 (0.04) (0.04)
0.167 (2.17) 2.17
0.093 (1.37) 1.23
return t-stat Sharpe ratio
-0.092 (2.34) (1.16)
-0.063 (0.54) (0.59)
-0.054 (0.85) (0.85)
-0.192 (2.44) (2.44)
-0.062 (1.00) (0.89)
Absolute t statistics in parentheses Robust standard errors + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001
Table 6: Returns of Momentum Strategies: Risk This table reports analysis on the returns of portfolios formed using momentum strategies using a multi-factor risk model. The regression is rmd = α + βe (red − rf d ) + βb (rbd − rf d ) + βc (rcd − rf d ) + d . rm is the return of the momentum portfolio, re , rb and rc are the returns to the S&P CNX NIFTY total returns index, the NSE (National Stock Exchange) government securities composite total returns index, and a long-only equal-weighted investment in all liquid NCDEX commodities. rf is the return to 91-day Indian T-bills. The momentum strategies buy commodity futures that had positive returns over an earlier period (the determining period) and sell commodity futures that had negative returns. Commodities are equal-weighted, portfolios are rebalanced daily, and all transactions occur at end of day at the last traded price. The table presents results for strategies based on a determining period that began x days ago and finished y day ago at the end of day and are labeled as “x-y days.” The strategies considered are 1-2 days, 2-11 days, 11-26 days and 26-51 days. The strategies only apply to commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades. The returns are adjusted for limit moves by prohibiting trade on the day of a limit move. Observations are equal-weighted.
Absolute t statistics in parentheses. Robust standard errors There are 1253 daily observations for each portfolio + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001
Table 7: Returns by Trader Size This table reports regressions of annualized return measured at the trader-commodity-day level on indicators for trader size and five other sets of indicators for trader characteristics. Position size is the average capital the trader held in the commodity in the previous 24 trading days on days they participated in trade. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. If Position Size is greater than Rs 10m ($230k), then a trader is Large, if between Rs 1m ($23k) and Rs 10m ($230k), then Medium, else Small. See Section 2.2 for a description of each variable. The omitted category for each set of indicators is: Small for size, Client for member/client, Unclassified for institutional type, Young for age, Small settlement for settlement size, and Not local for local. Observations are equal-weighted or capital-weighted as indicated. Standard errors clustered by commodity group (53 clusters) and trade date (1,295 clusters).
Large (> Rs10m) Medium (Rs1m to 10m)
Absolute t statistics in parentheses. 30,113,829 Observations + p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
Table 8: Returns and Share of Capital by Trader Size, by Year Panel A reports the average returns for the aggregate portfolios of Large, Medium and Small traders. Position size is the average capital the trader held in the commodity in the previous 24 trading days on days they participated in trade.
Capital is calculated assuming full collateralization of positions and that each trader has sufficient
collateral for the maximum trading position within the trading day. If Position Size is greater than Rs 10m ($230k), then a trader is Large, if between Rs 1m ($23k) and Rs 10m ($230k), then Medium, else Small. Return for each portfolio is the total profit of all traders in the category divided by total capital. Only commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades, are included. Observations are capital-weighted. Panel B reports the share of total capital for Large, Medium and Small traders.
As exchange trading is
zero-sum and returns are before transaction costs, the sum product of capital share and return over the three trader categories is equal to zero for each period.
Panel A: Return
Panel B: Capital Share
Absolute t statistics in parentheses. Robust standard errors +
p < 0.1,
p < 0.05,
p < 0.01,
p < 0.001
Table 9: Decomposition of Large Trader Returns into Momentum Strategies Panel A reports the average returns for the aggregate portfolio of Large traders split into inter-day return and within-day return. Inter-day profits are the profits earned from holding the net position from the end of the previous day for a whole day. Within-day profit accounts for all remaining profit. Only commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades, are included. Observations are equal-weighted. Panel B reports a regression of large trader inter-day return on the returns from the 1-2 days, 2-11 days, 11-26 P days and 26-51 days strategies shown in Table 5, plus a constant. The regression is rdlarge = α + 4j=1 βj rjd + d . Momentum return measures the average return for the component of large trader inter-day return that is correlated P with the momentum strategies and is defined as 4j=1 βˆj rj . % return explained is the fraction of average large trader inter-day return that is accounted for by the momentum return and is defined as
momentum return average inter-day return
j=1 βj r j r large
momentum return (2004 weight) calculates momentum return assuming the correlation of large trader inter-day P return remained the same as in 2004, 4j=1 βˆj2004 rj .
0.0581∗∗∗ (4.147) 0.0226∗∗∗ (11.134)
0.162∗∗∗ (3.416) 0.0386∗∗∗ (4.468)
0.0440∗∗ (3.215) 0.0254∗∗∗ (8.731)
0.0270 (1.606) 0.0225∗∗∗ (9.245)
0.0149 (1.527) 0.00980∗∗∗ (9.953)
Panel A: Average Return
Panel B: Decomposition of Inter-day Return 1-2 days
Momentum Return Momentum Return (2004 weight) % Return Explained
Observations Adjusted R2
Absolute t statistics in parentheses. Robust standard errors + p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
Table 10: Vector Auto Regression of Returns and Large Trader Net Position Changes This table reports a vector autoregression of large trader flows and commodity returns. Large trader flow, net position change, is constructed by summing net changes, in number of contracts across all maturities, and dividing by the open interest at the end of the previous day. Return is for the highest open interest contract in the commodity. Lagged variables are constructed using averages over four different periods, 1-2 days, 2-11 days, 11-26 days and 26-51 days. For example, the 2-11 days return, r 1 2, is the average daily return at end of day from 11 days ago to 2 days ago, where the close price is measured as the last traded price. The equations estimated are npccd = P P P P αn + 4k=1 βkn rck,d−1 + 4k=1 δkn npcck,d−1 + rcd and rcd = αr + 4k=1 βkr rck,d−1 + 4k=1 δkr npcck,d−1 + rcd where k indexes the four periods over which averages are taken. In Column 2 contemporaneous returns are included in the flow equation. Data from different commodities is pooled and ordinary least squares is used with standard errors clustered by commodity group and trading day. Only commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades, are included. Observations are equal-weighted.
Returns 0.588∗∗∗ (5.039)
r 2 11
r 11 26
r 26 51
npc 1 2
npc 2 11
npc 11 26
npc 26 51
Net Position Changes
Absolute t statistics in parentheses. 23,752 Observations + p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Standard errors clustered by commodity group and trade date
Table 11: Aggregate and Commodity-Level Size Measures This table reports regressions of annualized return measured at the trader-commodity-day level on a measure for trader size and five other sets of indicators for trader characteristics (coefficients not shown). Position size is the average capital the trader held in the commodity in the previous 24 trading days on days they participated in trade. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. The size variable used in Column 1 is Log Position Size, following the results in Figure 3 that return is approximately linear in Log Position Size. The variable Aggregate Position Size in Column 2 is the same as Position Size but is calculated over all commodities. Observations in each regression are capital-weighted. Standard errors clustered by commodity group (53 clusters) and trade date (1,295 clusters)
log position size log aggregate position size Other Trader Characteristics Adjusted R2
0.0559∗∗∗ (3.74) 0.0170∗∗ (2.93)
Absolute t statistics in parentheses. 30,113,829 observations +
p < 0.1,
p < 0.05,
p < 0.01,
p < 0.001
Table 12: Momentum Strategy Returns by Commodity Type This table reports average returns for portfolios formed using momentum strategies. The strategies in Panel A invest only in commodities without an international futures market, and those in Panel B invest only in commodities with an international futures market. The strategies buy commodity futures that had positive returns over an earlier period (the determining period) and sell commodity futures that had negative returns. Commodities are equal-weighted, portfolios are rebalanced daily, and all transactions occur at end of day at the last traded price. The table presents results for strategies based on a determining period that began x days ago and finished y day ago at the end of day, which are labeled as “x-y days.” The strategies considered are 1-2 days and 2-11 days. The strategies only apply to commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades. The returns are adjusted for limit moves by prohibiting trade on the day of a limit move. Observations are equal-weighted.
Panel A: Without International Market 1-2 days
return t-stat Sharpe ratio
0.347 (4.84) 2.50
1.058 (3.32) 4.69
0.261 (2.28) 2.28
0.420 (3.08) 3.08
0.073 (0.77) 0.69
return t-stat Sharpe ratio
0.307 (4.09) 2.11
0.852 (2.52) 3.56
0.316 (2.75) 2.76
0.252 (1.75) 1.75
0.126 (1.23) 1.11
Panel B: With International Market 1-2 days
return t-stat Sharpe ratio
0.224 (5.57) 2.76
0.371 (3.09) 3.40
return t-stat Sharpe ratio
0.174 (4.39) 2.18
0.276 (4.72) 4.73
0.161 (2.04) 2.04
0.137 (1.98) 1.77
0.137 (1.19) 1.30
0.173 (3.06) 3.07
0.191 (2.55) 2.55
0.185 (2.55) 2.29
Absolute t statistics in parentheses. Robust standard errors + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001
Table 13: Returns for Trader Size Categories by Commodity Type This table reports the average returns for the aggregate portfolios of Large, Medium and Small traders. Panel A includes only commodities without an international futures market, and Panel B only those commodities with an international futures market. Position size is the average capital the trader held in the commodity in the previous 24 trading days, on days they participated in trade. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. If Position Size is greater than Rs 10m ($230k), then a trader is Large, if between Rs 1m ($23k) and Rs 10m ($230k), then Medium, else Small. Return for each portfolio is the total profit of all traders in the category divided by total capital. Only commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades, are included. Observations are capital-weighted.
Panel A: Without International Market
Panel B: With International Market
Absolute t statistics in parentheses. Robust standard errors + p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
Figure 1: Capital over Time This figure shows total Capital from Q1 2004 to Q1 2008. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. This
average daily capital, $ bn 1 1.5 2
capital measure is slightly higher than the value of open interest.
Figure 2: Share of Capital by Commodity over Time This figure shows the share of Capital for the top 10 commodities by share of capital. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. This capital measure is slightly higher than the value of open interest.
100% Others Cumin
Pepper Silver 60% Black Lentil Mustard 40%
Sugar Soy Oil Soybean
Chickpea Cluster Bean Seed 0% 2004
Figure 3: Average Returns by Size Category The top figure shows the coefficients and confidence interval from a regression of annualized return at the trader-commodity-day level on 10 indicators for trader size. Position size is the average capital the trader held in the commodity in the previous 24 trading days on days they participated in trade. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. Each indicator covers one order of magnitude using the exponential base, that is if Natural Log Position Size is between 11 and 12, then the size indicator for 11 is equal to 1, else 0. Observations are capital-weighted. There are 30,113,829 observations. Standard errors clustered by commodity group (53 clusters)
Position Size and Excess Return
and trade date (1,295 clusters)
The bottom figure shows the share of capital for each size category.
Rs 1m $ 23k Annualized 20% Return
Rs 10m $ 230k
‐10% ‐20% 20%
% of Capital
10% 0% 11 12 13 14 15 16 17 18 19 20
Natural Log Position Size
Figure 4: Capital Share by Size Category The figures show the share of capital for Large, Medium and Small traders from Q1 2004 to Q1 2008. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. This capital measure is slightly higher than the value of open interest.
capital share, % .1 .15 .2
capital share, % .2 .3 .4
capital share, % .3 .4 .5
NCDEX operates similarly to commodity futures exchanges in established markets, including the measures it uses to mitigate counter party risk and regulate the market. I review the main features here: margins and mark-to-market, delivery, order matching, the member and client structure of the exchange, and position limits. Contracts traded are obligations to sell or purchase a specific quantity of commodity at a specific time and location, and traders take short or long positions. Traders pay margin when they enter into a contract, and pay or receive mark-to-market payments as the price of the contract moves for or against them. Mark-to-market payments are made daily using the settlement price, which is a value-weighted average of end-of-day trading prices for liquid contracts. Initially the majority of the contracts were cash-settled, but from December 2006, most contracts moved to compulsory delivery. Position limits (maximum position sizes) for clients and members limit the price influence of any single market participant, and price limits prevent prices from moving more than a prescribed amount on any given day. The exchange fixes margins, position limits and price limits both by itself and under order from its regulator, the Futures Market Commission. The trading system is electronic and most traders never handle the physical commodity; a small fraction of positions end up in delivery at expiry and most positions are squared off. The exchange operates as a limit order market, with fee-paying members who submit orders either on their own behalf or on the behalf of their clients if they have a brokerage business. Some members provide their clients with immediate trading access through their proprietary front end systems.
To identify limit moves, I compiled data on the settlement price for each contract for each day, and circulars to members of the exchange that list initial and subsequent contract specifications, which include limit move details. Limit move rules consist of an initial price limit and a final price limit, specified in percent. If the absolute percentage difference between the traded price and the previous day’s settlement price is less than the initial price limit, trading continues freely. On hitting the initial price limit, a 15 minute cooling off period is triggered, unless this occurs in the last half an hour of trade.40 Following the cooling off period, trading continues within a price band defined by the final price limit. If the final price limit is hit, trading can continue but only within the price band defined 40
Before April 27, 2006, trading was suspended during this period. After April 27, 2006, trading was permitted during the cooling off period within a price band defined by the previous day’s settlement price and the initial price limit. If the initial price limit is hit during the final 30 minutes of trade, there is no cooling off period. In the earlier period trading was suspended, and in the later period trading continued within the price band.
by the previous day’s settlement price and the final price limit. In practice, trading does continue after price limits are hit, but at a reduced level, as some day traders want to book profits and other participants want to exit positions. Price limits were initially set at 10% for the initial price limit and 20% for the final price limit for all contracts. From January 5, 2005, price limits were set differently for different commodities. Price limits decreased dramatically through the sample. Including the amendments on the dates mentioned, there were 97 changes to contract price limits for the 29 main contracts in the sample.41 I use the panel of price limits to determine limit moves for each contract. As I do not know exactly when limit moves occurred, I must identify them from price changes. I initially define a limit move as any price change greater than 0.98 times the final price limit or between 0.98 and 1.02 times the initial price limit. I find 205 limit moves during the sample period. However, I find price changes greater than the final price limit and also find limit moves for black lentil, cluster bean seed and cluster bean gum during August and September 2005 that occurred at 6%, when my panel of price limits lists the final price limit at 20%. I conclude that the publicly available circulars do not record all the changes in the price limits and that the price limits for black lentil and the cluster bean contracts must have changed before October 17, 2005. Since I do not find increased density at the major price limit percentages, e.g. 6%, after removing identified limit moves, I do not believe this is a widespread issue, though I control for it in the following way. For my base case, I define a limit move as any change greater than or equal to 0.98 times the final price limit, between 0.98 and 1.02 times the initial price limit, and as any price changes between 3.9% and 4.1% and 5.9% and 6.1%. This increases the number of limit moves to 245, or 1.0% of all liquid commodity-day observations. To make sure I capture all possible limit moves, I construct a “> 4% removed” definition of limit moves, where a limit move is considered to be any price change greater than 3.9%. This definition results in 589 limit moves, or 2.5% of all liquid commodity-day observations.42 For the momentum strategies, I assume that the strategy is not able to transact if there is a limit move. I allow the strategies to change positions on the next day when there is no limit move at the end of the day. For the vector auto regressions, as reported in my base case, I do not control 41
With the launch of the jaggery and the maize contracts on January 5, 2005, price limits began to differ by contract. From the April 26, 2005, price limits for individual contracts were changed. A major change occurred on October 17, 2005 when limits were reduced, in some cases to as low as 6% for the final price limit. After complaints from exchange participants, this change was revised on November 20, 2005 with most contracts receiving increased limits. Further changes across multiple contracts occurred on May 9, 2006 and May 23, 2006, when the majority of agricultural contracts were given initial and final price limits of 4% and 6% respectively. On February 18, 2008, all agricultural contracts were given a final price limit of 4%. 42 As a comparison, Roll (1984) finds in his study of orange juice futures that slightly over 10% of observations are limit moves. The reason for the difference between the finding in India and OJ futures in the US is that the price limits of OJ futures were more restrictive, and the main determinant of OJ prices is freezing weather in a specific location, which is readily apparent to the market and leads to marked shifts in prices.
for limit moves. When I do control for limit moves (unreported), the results are almost identical to those reported. I explain how to control for limit moves for VARs in footnote 37.
Table B.1: Returns of Momentum Strategies: Transaction Costs This table reports average returns for portfolios formed using momentum strategies after adding additional transaction costs. Panel A presents results with an additional transaction cost of 25% of the bid-ask spread, and Panel B presents results with an additional transaction cost of 50% of the bid-ask spread. The strategies buy commodity futures that had positive returns over an earlier period (the determining period) and sell commodity futures that had negative returns. Commodities are equal-weighted, portfolios are rebalanced daily, and all transactions occur at end of day at the last traded price. The table presents results for strategies based on a determining period that began x days ago and finished y day ago at the end of day, which are labeled as “x-y days.” The strategies considered are 1-2 days and 2-11 days. The strategies only apply to commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades. The returns are adjusted for limit moves by prohibiting trade on the day of a limit move. Observations are equal-weighted.
Panel A: + 25% Bid-Ask Spread 1-2 days
return t-stat Sharpe ratio
0.133 (3.13) 1.55
0.414 (3.05) 3.35
0.167 (2.46) 2.47
0.129 (1.59) 1.59
return t-stat Sharpe ratio
0.189 (4.32) 2.14
0.303 (2.23) 2.45
0.216 (3.12) 3.12
0.173 (2.07) 2.07
0.276 (2.01) 2.20
0.059 (0.88) 0.88
0.268 (1.97) 2.17
0.186 (2.68) 2.69
-0.079 (1.32) (1.19) *
0.103 (1.52) 1.36
Panel B: + 50% Bid-Ask Spread 1-2 days
return t-stat Sharpe ratio
-0.021 (0.50) (0.25)
return t-stat Sharpe ratio
0.144 (3.30) 1.64
-0.037 (0.45) (0.45)
-0.271 (4.47) (4.01)
0.118 (1.41) 1.41
0.050 (0.73) 0.65
Table B.2: Returns of Momentum Strategies: Limit Moves This table reports average returns for portfolios formed using momentum strategies. The strategies buy commodity futures that had positive returns over an earlier period (the determining period) and sell commodity futures that had negative returns. Commodities are equal-weighted, portfolios are rebalanced daily, and all transactions occur at end of day at the last traded price. The table presents results for strategies based on a determining period that began x days ago and finished y day ago at the end of day, which are labeled as “x-y days.” In Panel A the strategies considered are 1-2 days, 1-11 days, 1-26 days and 1-51 days. In Panel B the strategies considered are 2-11 days, 11-26 days and 26-51 days. The strategies only apply to commodities which are liquid, defined as having 40 of the last 50 days with greater than 100 trades. The returns are adjusted for limit moves by prohibiting trade on the day of a limit move. The Base Case identifies limit moves using information from the exchange on the rule and prices. There is some uncertainty in this identification, so two further cases are considered. The Unadjusted Case assumes that there are no limit moves and allows transactions every day. The > 4% Removed case assumes any price changes greater than 4% are limit moves. Observations are equal-weighted. Base Case
> 4% Removed
Panel A: Strategies Based on Recent Returns 1-2 days
return t-stat Sharpe ratio
0.286 (6.79) 3.36
0.294 (7.06) 3.50
0.290 (6.88) 3.41
return t-stat Sharpe ratio
0.306 (6.77) 3.36
0.307 (6.87) 3.41
0.291 (6.47) 3.21
return t-stat Sharpe ratio
0.202 (4.69) 2.33
0.205 (4.79) 2.37
0.205 (4.60) 2.28
return t-stat Sharpe ratio
0.075 (1.86) 0.92
0.073 (1.84) 0.91
0.077 (1.88) 0.93
Panel B: Strategies Based on Lagged Returns 2-11 days
return t-stat Sharpe ratio
0.233 (5.33) 2.64
0.238 (5.48) 2.72
0.237 (5.34) 2.65
return t-stat Sharpe ratio
0.095 (2.23) 1.11
0.095 (2.21) 1.10
0.101 (2.33) 1.15
return t-stat Sharpe ratio
-0.092 (2.34) -1.16
-0.093 (2.38) -1.18
-0.095 (2.34) -1.16
There are 1253 daily observations for each portfolio Absolute t statistics in parentheses. Robust standard errors + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001
Table B.3: Alternative Size Measures This table reports regressions of annualized return measured at the trader-commodity-day level on a measure for trader size and five other sets of indicators for trader characteristics (coefficients not shown). Position size is the average capital the trader held in the commodity in the previous 24 trading days, on days they participated in trade. Capital is calculated assuming full collateralization of positions and that each trader has sufficient collateral for the maximum trading position within the trading day. The size variable used in Column 1 is Log Position Size, which is renamed Log Average Capital, following the results in Figure 3 that return is approximately linear in Log Position Size. Columns 2 through 7 use alternative size metrics, which are all defined over the previous 24 trading days. Log Average Trades, and Log Average Trade Value are defined similarly to Log Average Capital. Log Max Capital is the log of the day with the most capital in the previous 24 trading days. The variables Log Total Capital, Log Total Trades and Log Total Trade Value sum the respectively variables over the previous 24 trading days before taking the logarithm. Observations in each regression are capital-weighted. Standard errors clustered by commodity group (53 clusters) and trade date (1,295 clusters)
(1) log average capital
log average trades log average trade value
0.0033 (0.71) 0.0399∗∗∗ (4.27)
log max capital
log total capital
log total trades log total trade value
Other Trader Characteristics Adjusted R2
0.0011 (0.37) Yes 0.000187
Absolute t statistics in parentheses. 30,113,829 observations + p < 0.1, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001