Essays on Information Diffusion and Stock Markets

Thesis (Ph.D.)--University of Washington, 2017-06My dissertation is a compilation of three separate research studies that explore how information diffuses in financial markets. The first chapter examines how non-uniform information diffusion through distinct networks segments U.S. financial markets. Using changes in newspaper ownership networks, I document that a network link between different geographic areas leads to increased comovement of turnover and returns between stocks headquartered in those areas. Consistent with delayed content sharing within a network, the largest increase in comovement is observed using weekly data. I show that the network-driven comovement is not driven by fundamentals and is weaker for large firms with high institutional ownership and decreases over time. I also document that a network link causes price levels of linked stocks to become more similar. My findings show that segmented information networks lead to segmented financial markets with implications for market efficiency, home bias, and the effects of changes in the U.S. media landscape on financial markets. The second chapter shows that investors do not fully monitor the information about directors available in the past prices of firms within the network the directors oversee. A long-short portfolio using this information yields an annual alpha of 6.6%. This predictability is limited to when firms share a director and is not driven by industry or previously identified economic links between firms. The predictability is largest in the long end, when small firms predict big firms, and when information on shared directors is costlier to obtain. Trading by the shared directors is a key mechanism: filtering on their trades increases the annual alpha to 15%. The third chapter studies the econometric properties of a commonly used network-based measure of information diffusion between economically linked firms. Previous studies use this measure to document failures of market efficiency with price discovery requiring up to a year. The measure is constructed as the long-short alpha of portfolios formed sorting on the preceding returns of firms economically linked to portfolio firms. We show that correlated alphas between linked firms bias these measures. Existing studies have monthly biases as large as a factor of two. This bias creates predictability even after price discovery completes. Subtracting the predicted return from the sorting firms' returns removes this bias. Eliminating this bias reveals a more efficient market than previously documented: price discovery takes one month