Abstract
This study builds on recent research by Clarke, Ferris, Jayaraman, and Lee [Citation2006]. Based on a sample of U.S. bankruptcies between 1995 and 2001, Clarke et al. failed to find evidence of a positive bias (or overoptimism) in analyst recommendations. We extend Clarke et al.'s findings by including the Global Financial Crisis (GFC) period and using an international sample of large corporate bankruptcies (the combined total assets of our firm failure sample exceed US$1.5 trillion). We also extend their research by examining the explanatory and predictive power of earnings growth forecasts, analyst downgrades/upgrades, and analyst coverage in a firm failure model. Overall, we find significantly more evidence of analyst overoptimism than did Clarke and colleagues.
Notes
1. For example, an article in the Washington Post (November 3, 2004), “Bear Stearns Analysts Had WorldCom Doubts,” cited court documents and internal email communications between Bear Stearns analysts and investment bankers revealing that Bear Stearns was aware of Worldcom's financial difficulties prior to its collapse, notwithstanding its public support for the stock.
2. The firm settlements were reached with the SEC, the National Association of Securities Dealers, the New York Stock Exchange, the New York Attorney General, and other state regulators.
3. Hubbard and Stephenson [1997] document the poor returns from investing in bankrupt firms (see also Clark and Ofek [1994] for similar results on distressed restructures).
4. See Bloomberg BusinessWeek article “Bear Stearns’ Big Bailout,” March 14, 2008.
5. Frino, Jones, and Wong [2007] discuss some differences in international bankruptcy law, particularly between the United States and Australia.
6. Arguably, the rationale for using matched pair designs is when data collection is prohibitively expensive and/or where the data are not readily available. With the widespread availability of international databases (such as IBES), it is now easier for researchers to extract larger samples at little or no additional cost.
7. We also tested the interaction of the U.S. firm dummy variable with downgrades in analyst recommendations but found that this variable was highly correlated with the interaction variable (SELL*US_FIRMit) specified in Model 2. Further, this interaction effect was not found to be significant in the logit model, possibly because it was highly correlated with other variables in the model.
8. The t-values are calculated by dividing the hazard ratio by the standard error.