Han Ozsoylev, Johan Walden, M. Deniz Yavuz, and Recep Bildik
Investor Networks in the Stock Market
Review of Financial Studies | Volume 27, Issue 5 (May 2014), 1323-1366

Heterogeneous and decentralized spread of information is known to influence the trading behavior of investors, see, e.g., Shiller and Pound (JEBO 1989), Ivkovic and Weisbenner (RFS 2007), and Hong, Kubik, and Stein (JF 2004). Such so-called information diffusion may help explain several fundamental stylized facts of stock markets. First, investors are known to hold vastly different portfolios. Such heterogeneous behavior is what one would expect when there is decentralized information diffusion in the market, but is contrary to the prediction of classical models that everyone should hold the market portfolio. Second, stock markets are known to experience large movements that are unrelated to public news, as documented in Cutler, Poterba, and Summers (JPM 1989) and Fair (JB 2002), suggesting other channels than public information channels through which information is incorporated into asset prices. Third, the dynamics of asset returns and trading volume are known to be very rich, see Karpoff (JFQA 1987), Gallant, Rossi and Tauchen (RFS 1992), Bollerslev and Jubinski (JBES 1999), Lobato and Velasco (JBES 2000), and Gabaix, et al. (NATURE 2003). Returns and volume in many markets are “heavy-tailed" (meaning that the risk for extreme events is much higher than in standard models), time varying, have “long memory" (meaning that shocks are very persistent), and are related to each other in a complex manner. Lumpy, heterogeneous information diffusion provides a potential explanation for such rich behavior.

Model and Results

We use a network approach to model information diffusion in a stock market, and study how the network position of an investor affects performance and trading behavior with respect to information events in the market. Loosely speaking, an information network describes how diverse information spreads over time among a population of investors, as shown in Figure 1. A fundamental property of these models is that more central agents in the network tend to access and trade on information earlier than less central agents, as well as make higher profits, see Ozsoylev and Walden (JET 2011) and Walden (WP 2014). Centrality here captures the important property that it is not only how many people you know that is important, but also who you know. An investor's centrality therefore does not only depend on how many neighbors that investor has but also on the centrality of those neighbors, in an endogenous fashion. In Figure 1, for example, the implication is that investor 1 is more central than investor 2, although investor 2 actually has more neighbors than investor 1.

1: [Information network]
The figure shows an information network of 21 investors in a market. Investors are represented by circles and links between investors by lines. In the Figure, investor 4 receives a valuable piece of information at time t, trades, and then shares it with his neighbors at t + 1, who then trade and share the information with their neighbors at t + 2. Information thus diffuses through the network in a well-defined manner.

Investors who are centrally placed in the network tend to receive information signals early, whereas investors who are in the periphery tend to receive them late. As a result, the trading behavior and profitability of individual investors are influenced by their position in the network, and the dynamics of aggregate asset prices depend on the network's structure.

We develop a method to approximate the market's information network, using observable data. The general idea is that information links may be identified from realized trades, since investors who are directly linked in the network tend to trade in the same direction in the same stock at a similar point in time. Using this approach, we identify an Empirical Investor Network (EIN), and in simulations show that the true information network is indeed well estimated by the EIN. We calculate the EIN using account level data that cover all trades on the Istanbul Stock Exchange in 2005. We find that our method is quite robust to omitting a significant fraction of investors from the sample, suggesting that our method can also be used with nonexhaustive datasets.

We then study the relationship between investor centrality and returns, and find substantial support for a positive relationship. A one-standard deviation increase in centrality, all else equal, leads to a 0.7%-1.8% increase in returns for an investor over a 30-day period, depending on the specification. These results are obtained after controlling for other variables, such as trading volume. We also document that centrality is directly related to acting early on information. We identify several idiosyncratic information events that were associated with large stock price movements, and find that central investors in the network tended to trade—in the right direction—before peripheral investors. Our results suggest that information diffusion is an important determinant of investors' trading behavior and profitability.

Concluding Remarks

Our study reinforces a view of the stock market as a place where information is gradually incorporated into asset prices. Information networks provide an intermediate information channel, in between the public arena, where news events and prices themselves make some information available to all investors, and the completely local arena of private signals and inside information.

Our results suggest a decentralized diffusion mechanism through localized channels, e.g., social networks, in line with several recent studies that focus on specific investor groups. Our knowledge is still limited, however. Which factors determine a market's information network? Geographical location? Social networks? Other channels? Given datasets with more detailed information about investors in the market, further research may shed light on this important question.