Some background

I run a factor analysis of a time series $Y$ using a standard OLS model with n+1 independent variables $(F,X_1...X_n)$, where $F$ is the main factor (from an explanatory power perspective).

$X_i$ and $F$ are highly correlated ($\rho>0.8$) but on certain days $t$, $X_i(t)$ will diverge significantly from its estimated value using the coefficient of its regression against $F$ : $X_i=\beta_i F$. In other words the distribution of $\epsilon_i(t)=X_i(t)-\beta_i F(t)$ has fat tails.

The reason why those highly collinear factors are included in the regression is precisely the presence of outliers which have a significant impact on the dependent variable $Y$.

My approach

To get rid of the multicollinearity problem yet keep the outliers, I create new factors $X_i'$ which are constructed like this:

  • $X_i'(t)=\epsilon_i(t)$ if $\epsilon_i(t)$ is in the top or bottom $n$ percentiles of $\epsilon_i$
  • $X_i'(t)=0$ otherwise

so if $n=5$ for example, $X_i'$ will be zero 90% of the time.

I then regress $Y$ against $(F, X_1'...X_n')$

My questions

  • Is there a flaw in the method?
  • Are there alternatives that would help me capture that effect?

(I know that I could orthogonalise each $X_i$ vs $(F,X_0...X_{i-1})$ but the results are the same as and as unstable as regressing against the original factors).

Note: in practice, $Y$ represents the returns of a stock, $F$ represents the returns of the market and the $X_i...X_n$ represents other factors such as industry indices.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.