Serving the Quantitative Finance Community

 
User avatar
statarb63
Topic Author
Posts: 0
Joined: March 2nd, 2013, 6:51 pm

multivariate models for returns prediction

July 29th, 2013, 9:16 pm

Dear all,I'd like to clarify a few basic concepts in multivariate time series models. Let's say the aim is to forecast returns themselves, if possible by using information contained in the correlation between assets. For example if I know that two assets are correlated, and one starts trending, then I think this should help me forecast the other.1) In what ways can I take into account correlation in both time and cross-section? For example, if I use a Kalman filter, should the dependence between assets be modeled in the covariance matrices (in that case, should it be in the measurement or hidden states or both covariances?), or through the measurements and state transition matrices? The former case seems more classic, but the latter seems like another way to include dependence between series.2) When should I use Kalman filter rather than a vector AR model for this application?3) Would the dynamic models above (KF or VAR) be expected to be better than some kind of lead/lag regression (not necessarily linear regression but a static model, although one could re-estimate the weights often or something)?
Last edited by statarb63 on July 28th, 2013, 10:00 pm, edited 1 time in total.