CAPM model as a regression
The CAPM model states that the returns of a stock are-
$ r_s=r_f+\beta (r_m-r_f)+\varepsilon_s $
The $ \beta $ defined above is then calculated as $ \frac{cov(r_s,r_m)}{var(r_m)} $ . My question is regarding this formula. This is the regression coefficient if the intercept is NOT set to 0, which is not the case in CAPM since the intercept is set to a constant $ r_f $ , which is essentially regressing $ r_s-r_f $ against $ r_m-r_f $ while setting the intercept to 0. That should yield $ \beta=\frac{E[(r_s-r_f)(r_m-r_f)]}{E[(r_m-r_f)^{2}]} $ , which is not equal to the canonical form.
Please let me know what is the issue here.
If you really believed the CAPM’s prediction that $ \alpha=0 $ , then imposing $ \alpha=0 $ in your estimation would indeed lead to your 2nd formula.
The problems?
- The CAPM doesn’t work so imposing a false restriction during estimation is problematic.
- More generally, taking factor models extremely seriously and imposing $ \alpha=0 $ in estimation to gain efficiency loses you some robustness because factor models are almost certainly at least somewhat misspecified.
Empirical researchers generally don’t restrict a constant to zero during estimation.
Model 1 (without a constant):
Let’s assume we have the following regression model (without a constant):
$$ r_{st} - r_{ft} = \beta_1 \left( r_{mt} - r_{ft} \right) + \epsilon_t $$ Assuming the orthogonality condition $ \operatorname{E}\left[\epsilon_t \left( r_{mt} - r_{ft}\right)\right] = 0 $ , then $ \beta_1 $ would be given by:
$$ \beta_1 = \frac{\operatorname{E}\left[\left( r_{st} - r_{ft} \right)\left(r_{mt} - r_{ft} \right) \right] }{\operatorname{E}\left[\left(r_{mt} - r_{ft}\right)^2\right]} $$ If you really take the CAPM theory seriously, then there is something principled to imposing the restriction $ \alpha= 0 $ in estimation (which is what we did above). Quoting Cochrane (2004) with regards to more general factor models with normally distributed errors, “The maximum likelihood estimate of $ \beta $ is the OLS regression without a constant.” As Cochrane describes though, researchers don’t generally estimate without a constant because it sacrifices some robustness.
Model 2 (add a constant):
$$ r_{st} - r_{ft} = \alpha_2 + \beta_2 \left( r_{mt} - r_{ft} \right) + \epsilon_t $$ Now with $ \alpha_2 $ there and assuming the orthogonality conditions $ \operatorname{E}[\epsilon_t] = 0 $ and $ \operatorname{E}\left[\epsilon_t \left( r_{mt} - r_{ft}\right)\right] = 0 $ , you get:
$$ \beta_2 = \frac{\operatorname{Cov}\left( r_{st} - r_{ft} , r_{mt} - r_{ft} \right) }{\operatorname{Var}\left( r_{mt} - r_{ft} \right)} $$ Model 1 is a special case of Model 2 where $ \alpha $ is restricted to 0.
Model 3 (if the risk free rate weren’t random):
If the risk free rate isn’t random then it drops out:
$$ \beta_3 = \frac{\operatorname{Cov}\left( r_{st}, r_{mt} \right) }{\operatorname{Var}\left( r_{mt} \right)} $$ In periods like the present where the risk free rate is constantly about 0, maybe this bogus assumption is innocuous. I think it’s hand-wavy, intro MBA type stuff though.
A comment on the CAPM
Be aware that the CAPM is a zombie theory: long ago shot dead in academia because it doesn’t work, the CAPM continues to skulk the earth. Quoting Fama and French (2004), “… the empirical record of the model is poor—poor enough to invalidate the way it is used in applications.”
References
Cochrane, John. 2005. Asset Pricing, p. 273