1. Statistically, the slope coefficient in a simple regression comes with a standard error. The beta estimate for a typical US company has a standard error that is about 0.20. One way to read this number, is when you are told that the beta for a company (from a regression) is 1.10, the true beta could be anywhere from 0.70 to 1.50 (plus or minus two standard errors).
2. Outside the US, regression betas often look much more precise but only because the indices used tend to be narrow local indices (DAX, Bovespa, Sensex). As anyone who has toyed with the parameters of the beta regression (2 vs 5 years, daily vs weekly, different market indices) knows, you can arrive at very different betas for the same company, based on your choices. None of them is the right beta, and they may all be coming from the same distribution, but it is a wide one.
3. By definition, a regression beta has to be backward looking, since you need past returns. To the extent that companies change their business mixes (by expanding, divesting and acquiring businesses) or their financial leverage (debt ratios) over time, the regression beta may not be a good measure of the beta for the future.
4. If the only way you can estimate betas is with a regression, you will be stymied right from the start, if you are analyzing the division of a company, a private business or a company that went public recently.
5. By making beta a statistical number, we are missing the fundamentals that drive beta. Every company has an intrinsic or true beta that comes from choices it has made and understanding how these choices can cause your beta to change is central to a better beta estimate.
So, I would take any beta reported for a company by a service or an analyst with a grain of salt. It probably came from a regression and should not define your thinking about the firm's risk. In my next two posts, I will offer my analysis of the determinants of betas and an alternative to regression betas.