Variables in financial time series

Question

I'm currently trying to do some forecasting with financial time series in R. I've started to do a linear regression where the dependent variable is the excess return calculated for 1, 12, 24, 36, and 48 months. I've calculated ln(r1/r0) for 1-month return, and ln(r13/r1) for the 12-month return. My question is: should I also calculate the predictors (e.g., dividend yield) in that way? So return ln(r13/r1) combined with dividend yield ln(dy13/dy1), or just the dividend yield in the 13th month combined with the return ln(r13/r1)?

Answer 1

Your question is very general. Assuming you've got some hypothesis you'd like to test with a linear model, you can create any number of predictors and then test them. So the answer is, you can do both!

However, as your number of predictors grows, the number of different models you can test increases by a factorial, which soon becomes a giant number. If you find yourself in this scenario (which you will sooner or later), I would highly recommend researching and reading up on the Lasso regression. The r package glmnet will handle the complexities of this. Essentially, it is designed to test all predictors for you and drop those which don't have much impact.