I just conducted a Fama-Macbeth regression where in the first step I calculated a time-series regression for each individual stock to get three betas (for mkt-rf, smb, hml) for each stock. Then I ran a cross-sectional regression with the returns of all companies at each timepoint as my dependent variable and the all the estimated betas of all companies as the independent variables. This gave me a time-series of the different risk premia of which I took a simple average.

Unfortunately, the results are very weird from my perspective. As I cannot find a mistake in my regression, I'd like to ask if anyone of you can somehow explain the results:

The risk premium of Mkt-Rf is -0.11% per month (but not statistically significant), SMB is 0,62% (significant) and HML -0,14% (not significant). Additionally I get a (in my opinion very high) alpha of 1.05% which is statistically significant.

While this is already very strange as I expected HML and especially Mkt-Rf to be positive, at least they're not significant.

But even worse, I tried to double check my results.

  1. I checked four subperiods as e.g. the Value-Premium/HML wasn't performing since the financial crisis 2007. Not even the Pre-crisis era from 2000-2007 in which HML performed really well.

  2. Instead of the three-factor model I caluclated CAPM and a two-factor model just with SMB. Against all my expectations the Alpha was lower in the One-factor-modell than in the other two although two factors where added which should increase the explanatory power and decrease the alpha.

Has anyone of you a logical explanation for the results? Otherwise I probably should check my calculcations again.


  • 1
    $\begingroup$ Which country and time horizon are you looking at? Further, did you apply any data filter like excluding financial companies and/or tiny, illiquid penny stocks? $\endgroup$ Dec 11, 2019 at 12:43
  • 1
    $\begingroup$ Your approach seems correct. What is your sample in terms of stocks and time-series? $\endgroup$
    – phdstudent
    Dec 11, 2019 at 12:44
  • $\begingroup$ The sample consists of 335 companies of the S&P500 since 2000 up to today. The data set was provided to us, the missing companies are those which don't have a full record since 2000 and delisted companies. No filter was used. I know that this may reflect a sampling bias but before (in a seperate task) the Fama-Macbeth regression I formed portfolios based on Book/Market Equity and found the value effect for this sample (at least in some subperiods) $\endgroup$
    – user43224
    Dec 11, 2019 at 12:53
  • $\begingroup$ I should add, according to the task I was given I did not sort the companies into portfolios or did some other correction. This obviously results in an error-in-variables problem as of the later used betas quite a lot are not signficant (in fact, 57% of the SMB-betas and 32% of the HML-betas but only 1,7% of the Mkt-betas). But especially because most betas of the Mkt-rf risk premium are significant at least that premium should be positive in my opinion. $\endgroup$
    – user43224
    Dec 11, 2019 at 13:01
  • $\begingroup$ Well on one hand we know the CAPM does not hold empirically. On the other hand there is so much noise in your time-series (individual stocks), that it wouldn't surprise me that negative risk premia is happening. Try a standard portfolio formation and see what happens. That helps understand if something is wrong with your code. $\endgroup$
    – phdstudent
    Dec 11, 2019 at 17:15


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.