Portfolio Momentum Strategies and Financial Databases: An investigation of Intra-day Pattern in Price Momentum

How do different databases define a firm? What are the rules for listing/de-listing firms across databases? In this paper we show that the listing/de-listing criteria across the CRSP and TAQ databases are different enough to impact the magnitude and direction of statistical profits of “momentum” portfolios. We show that while standard momentum returns established in the literature can be replicated using CRSP returns, TAQ data leads to entirely different results due to listing/de-listing discrepancies. We construct a two-stage filtering process to eliminate new/de-listing discrepancies between CRSP and TAQ to show a convergence in results. We then document that intra-day momentum profits exhibit an inverted U-shape. Given that intra-day prices follow a U-shape, we then show that an investor can “time” the market and enhance the profits of momentum portfolios if (s)he buys at noon prices and sells at the close. Introduction There is a growing body of literature on the predictability of stock prices based on information contained in past returns. Many studies show that portfolios of common stocks formed according to the simple rules of “relative strength” or “momentum” investment strategies show substantial statistical profits. This was first documented by Jegadeesh and Titman (henceforward JT) in 1993. Momentum strategies take advantage of return persistence and buy winners and sell losers. The motivation behind these portfolio-holding strategies is a belief that investors have common psychological biases and recent papers have built models of biased marginal investors to explain such return predictability. Much of the theoretical and empirical literature has uncritically accepted the proposition that statistical profits are real profits, that is, constructing portfolios by applying computer algorithms to financial databases and computing statistical profits replicates the actual profits of investors following these strategies. However, some recent work seem to suggest otherwise. For example, it has been documented that transactions cost of these strategies are substantial and essentially eliminate the profits. Korajczyk and Sadka (2003) find that price impact severely limits the amount of money that can be invested in a standard momentum strategy. They suggest that a momentum strategy that uses transactions cost as a price factor will still be successful. Grundy and Martin (2001) calculate that one round-trip transactions cost of 1.5% eliminates all momentum profits. Lesmond, Shill and Zhou (2003) find that stocks that generate momentum profits are precisely those stocks with high trading costs. Each of these models assume a model of trading costs that relative strength advocates have reason to criticize. Korajczyk and Sadka build a linear model of price impact, Lesmond et al. use a model that is contrary to microstructure theory and Grundy and Martin use simulation in their study. They do not measure transactions cost directly. To contribute to this branch of the literature that questions the relationship between statistical and real profits, in this paper we test whether financial databases themselves, along with the timing of the prices selected for computing returns, are an important determinant of statistical momentum profits. To date, no study has yet tested whether momentum profits actually exist or are they an artifact of the dataset used? A related question is if their existence is established, how (if at all) has their magnitude or significance changed? This is a pertinent question in light of the fact that most studies that document predictability of returns use significantly overlapping sample periods and there have been few out-of-sample tests. Table I lists some of the relevant research in this area; the time period, data source and exchange listings of firms that are included in the sample. We find that most studies have significant overlap in the time period and all of them use CRSP returns. It is not inconceivable that their similar conclusions are being driven by the common sample years or the same data source. Momentum in stock prices has been found to be correlated to certain variables – earnings, systematic risk, volume, etc. Yet, to date, no theoretical explanation accounts for why investors should systematically behave in a way consistent with the existence of intermediate term persistence of winners (and losers). As Chan, Jegadeesh and Lakonishok (1996) point out: “In the absence of an explanation, the evidence on momentum stands out as 1 An exception is JT 1999, which examines a sub-sample period 1990-98 for the 6-month hold strategy. a major unresolved puzzle ... The lack of an explanation suggests that there is a good chance that a momentum strategy will not work out-of-sample and is merely a statistical fluke”. In this paper, we seek to address the point made by Chan et al. We find that while the JT results can be replicated using CRSP (close-to-close) returns, comparable TAQ returns lead to an entirely different conclusion. This discrepancy arises due to the way the two databases handle new and de-listing of securities (firms). We construct a two-stage filtering process to eliminate new/de-listing discrepancies between CRSP and TAQ to show a convergence in results. We then show that intra-day momentum profits exhibit an inverted Ushape. Given that intra-day prices follow a U-shape, we then show that an investor can “time” the market and enhance the profits of momentum portfolios if (s)he buys at noon prices and sells at the close. This is consistent with the view that the price impact found by Korajczyk and Sadka, and Lesmond et al. is an important part of momentum profits. Both facts suggest that real momentum portfolios will not experience the returns shown by statistical momentum portfolios. The rest of the paper is organized as follows. In section 1 we discuss the data sources and the samples used. We also report the initial results of the JT momentum strategies using samples selected from the CRSP and TAQ databases. Section 2 discusses the filtering procedures used to make the two data sources comparable, and reports the results on portfolios formed from the filtered samples. Section 3 examines the effect of closing prices versus intra-day prices. Specifically, momentum strategies are implemented by using (returns calculated from) intra-day prices from the TAQ database and then compared with similar strategies that use close-to-close returns. We then construct a simple market-timing strategy that exploits the predictable variation in intra-day prices. Section 4concludes. 2 Chan, Jegadeesh and Lakonishok (1996), pp. 1682. 1. Sample and Methodology In this paper we use two different data sources to calculate returns to momentum portfolios. First we use CRSP data on monthly returns of stocks that are listed on the NYSE. We also use the TAQ database to obtain real time trade prices of NYSE-listed firms. Using these prices, we construct the returns series for all common stocks that are listed on the NYSE. To make these returns comparable with CRSP returns, we use the last recorded price for a security for a particular day as its closing price. This price, which may have been recorded in the TAQ database after 4:00 p.m. (NYSE closes at 4:00 p.m.) corresponds to a trade that could have occurred anytime up to 4:00 p.m. on the NYSE or 4:30 p.m. on a regional exchange. In case there is no volume on a certain day, we use the previous trade price as the closing price on the no-volume day. This is procedurally different from what CRSP does. CRSP uses the average of the bid and ask quotes for a security as its price on a day when there is no trade. Our sample period is April 1993 to December 1996. Penny stocks, American Depository Receipts (ADRs), Real Estate Investment Trusts (REITs) and such others are filtered out from the sample. Portfolios are constructed and held in the same way as JT 1993. At the beginning of each month all stocks are independently ranked on the basis of their past J-monthly holding returns (J=3/6/9/12). The stocks are then grouped into ten equally weighted portfolios based on the ranking. The portfolio with the highest (middle/bottom) returns is called the Top 3 To construct the returns series for securities from the TAQ database, we have to adjust for dividends and stock splits (including reverse splits) by using the Dividend File in the TAQ database. 4 We use NYSE firms and not Nasdaq because Nasdaq firms are smaller and are difficult to trade in momentumbased strategies. The double counting of dealer trades is another problem with Nasdaq firms. (Mid/Bot) portfolio. Each portfolio is held for the next K (K=3/6/9/12) months. Like JT, the power of the tests is increased by constructing overlapping portfolios. The JT trading strategy buys the equally weighted decile of stocks with the highest past returns and sells the equally weighted decile of stocks with the lowest past returns. The reward to this strategy is the difference in returns between the Top and Bot portfolios. Previous research has shown that some of the profits to contrarian strategies are due to methodological drawbacks in the calculation of returns. For example, DeBondt and Thaler (1985) cumulate short term (monthly) returns to loser an winner stocks over long periods and their measure of the arbitrage portfolio’s profitability is the difference in the cumulative raw returns (CRRs) of winner and loser portfolios. By cumulating short term returns, these strategies not only cumulate “true” short term returns but also the upward bias in each of the single period returns. Moreover, if the loser firms are low-priced relative to the winners, the return to the arbitrage portfolio will have a spurious upward drift that is unrelated to market overreaction. In this paper, we use holding period returns (HPRs) to measure portfolio performance. The buy and hold strategy (implied by the HPRs) is the appropriate measure to use and it has the added advantage of reducing transactions costs. HPRs are computed using the formula