1 Past Informed Trading and Past Dealer Inventory in US Equity Markets GERARD HOBERG* Yale University School of Management 135 Prospect Street Box 208200 New Haven, CT 06520-8200 Email: [email protected] Phone: 203-777-9596 Fax: 203-436-0630 * Yale University School of Management, New Haven. I wish to thank Ivo Welch, Matt Spiegel and Arturo Bris for excellent comments and advice. I would also like to thank participants at a seminar in Yale University. All errors are the author’s alone.2 Past Informed Trading and Past Dealer Inventory in US Equity Markets ABSTRACT Several studies document that past returns and past trading volume predict short-term stock returns. This study considers two new variables: the past number of trades signed as purchase and sale. These variables also predict weekly US stock returns and contain information that is distinct from past returns and past trading volume. Combined trading strategies outperform strategies based on past returns and past trading volume alone. A form of market inefficiency, lagged information dissemination, best explains the predictive power of the past number of signed trades. It cannot be explained by overreaction, biased self-attribution or systematic risk. I also find that the return predictability varies across exchanges and after tick size reductions. Several studies including Hasbrouck (1988, 1991a and 1991b) demonstrate that informed trading activity and dealer inventory costs can explain firm-by-firm spread width. Madhavan and Smidt (1991) and Hasbrouck and Sofianos (1993) show that informed trading activity and dealer inventory can also explain how individual trades impact the stock price. My study explores whether proxies for past informed trading activity and past dealer inventory constructed from order flow can predict future one-week stock returns. In a comprehensive dataset covering the major exchanges from 1993 to 2002, it considers 4 items: past informed trading proxies, past dealer inventory proxies, past returns and past trading volume. I find that past returns, past informed trading and past trading volume make distinct and economically relevant contributions to future returns. Past dealer inventory proxies make little or no contribution. It is a puzzle why variables constructed from observed order flow can predict future returns. The common theoretical perspective is to assume that they do not predict returns. For example, Kyle (1985) and many studies that follow assume that market makers set prices so that expected returns are zero given observed order flow. However, these same studies also conclude that informed traders place trades over time so that information dissemination is lagged. Intuitively, if the zero expected returns assumption is violated, agents could infer the direction of informed trading by observing order flow. Because information disseminates slowly, these inferences imply future informed trading in the same direction. In turn, these inferences can then predict future price changes. This form of market inefficiency can be large because an informed trader’s private information may include the existence of new product markets, knowledge of acquirers, or entry by rivals. My study shows that Kyle’s assumption is often violated and contributes to the literature of short-term return predictability. I find that inferences can be made about the actions of informed traders during the lagged process of price adjustment. In turn, agents can use these inferences to predict future returns. Moreover, these returns are permanent and cannot be explained by overreaction or biased self-attribution. The empirical results of this study, which are based on a comprehensive TAQ database that spans 1993 to 2002, can be summarized as follows:3 1. Results from statistical tests of return predictability: • Past returns, past trading volume, and past informed trading (signed buy and sell trades) are statistically significant predictors of future weekly returns. The significance of these predictors is most noteworthy for NASDAQ firms. • Past dealer inventory proxies make little or no contribution to the predictability of future weekly returns. • Past informed trading declines in importance after the 1997 (1/8th to 1/16th) and the 2000 (1/16th to 1 penny) tick size reductions, suggesting gains in market efficiency. • Future returns are more predictable for NASDAQ firms than for NYSE/AMEX firms. Future returns are also more predictable for small firms than for large firms. These results are generally consistent with market efficiency as both NASDAQ firms and small firms have larger transaction costs. 2. Summary of available trading profits and economic importance: • Past returns and past trading volume make economically large and distinct contributions to the profitability of trading strategies. Past trading volume grows in importance throughout the sample. • Past informed trading proxies make a modest but distinct contribution to trading profits. They are most important in 1993-1996 and for medium to small firms. • Past dealer inventory proxies do not make a distinct contribution to trading profits. • Trading strategies that invest in long positions earn profits that are consistently larger than strategies that invest in short positions. • Trading profits increase with diminishing returns for longer investment horizons. • Profits generally exceed the effective spread but are smaller than the quoted spread. • Trading profits are positive in bull markets (1998-1999) and in bear markets (2001-2002). 3. Inferences related to the sources of return predictability: • Future returns explained by past informed trading proxies are permanent. This is consistent with market inefficiency in the form of lagged and predictable information dissemination. My study makes three main contributions. First, I present a unified and comprehensive study of return predictability based on 4 items: past returns, past trading volume, past informed trading proxies and past dealer inventory proxies. This is the first study to consider past informed trading proxies (empirically important) and dealer inventory proxies (empirically unimportant). Second, this study presents a general method for constructing trading strategies when several variables are believed to predict future returns. Trading strategies that add past