This may be seen as a follow up question for the previous discussion on time-series vs cross-sectional factor models: Which approach to estimating fundamental factor models is better, cross-sectional (unobservable) factors or time-series (observable) factors?

Assume that we use a cross-sectional factor model (e.g. BARRA model).

Using cross-sectional regressions, we estimate the pure factor returns for each time period (by regressing stock returns on firm characteristics, such as P/E).

So we obtain time series of pure factor returns.

Then, is it appropriate to estimate a time-series regression where individual stock returns (that are also used in developing the cross sectional model) are regressed on pure factor returns (that are estimated using cross sectional regressions)?

And if yes: 1) What are the econometric implications of such an approach? Since the explanatory variables are also estimates, we may have an errors-in-variables problem. 2) How the betas estimated in time-series regressions compare with the original factor exposures (i.e. firm characteristics)?

Are there any research papers on these issues?


2 Answers 2


What you're describing sounds like the reverse of a Fama-Macbeth regression. The original Fama-Macbeth approach estimated rolling time series regressions to get CAPM betas and then doing a cross-sectional regression to estimate the overall sensitivity of returns to beta.

If I were to write down what the model looks like, I think you're talking about something like below $$y_{it}=\alpha_{i}+G_{it}F_{t}\beta_{i}+\varepsilon_{i} $$ where $G$ are firm characteristics, $F$ are the cross-sectional factor estimates and $\beta$ are the time series betas. One thing that stands out is that $F$ and $\beta$ are only identified because you're talking a two-step approach.

Offhand, I know of no papers that take the approach you describe. I would be more worried about some of the issues with Fama-Macbeth, like getting the standard errors right. For the errors in variables issue, I can't say for sure, but my hunch is that most finance data is wide so the standard error on cross-sectional regressions should be low enough that it isn't a huge issue. Further, I would expect the betas to be strongly correlated with the average level of the firm characteristics over time.

A better question is why you would want to take this approach. Using firm characteristics means that the exposure of each stock to the factor changes over time. It is more responsive than the time series beta will be.

It makes more sense to me to do the cross-sectional regressions $$ y_{i}=\alpha+G_{i}F+\varepsilon_{i} $$ and get the residuals $$ \varepsilon_{i,t}=y_{i,t}-\alpha_{i}-G_{i,t}F_{i} $$ and then do a time series regression of the residuals against the factors to see if there is any lingering exposure $$ \varepsilon_{t}=\alpha_{t}+\beta F_{t}+\eta_{t} $$


Intuitively, I think you will end up with average beta over the period. This is like doing panel regression by first running cross sectional regression and then running time series regression.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.