SERVING THE QUANTITATIVE FINANCE COMMUNITY

 
User avatar
edult
Topic Author
Posts: 49
Joined: October 31st, 2007, 8:24 pm

factor selection in factor models

June 23rd, 2009, 3:34 am

Is there any standard in the selection/number of factors in factor modeling? I have seen different approaches in several papers. Once the selection is not for principal components, but for factors like industry returns things are a little bit different. Some industries have more stocks, some are correlated with each other for a certain period of time, etc. Do you know any good reference on the subject? I have already checked Grinold & Kahn, not very helpful. The following paper seems to be recent, but uses principal components...http://www.econ.nyu.edu/user/baij/econometrica02.pdf
 
User avatar
ACD
Posts: 107
Joined: April 19th, 2004, 8:09 am

factor selection in factor models

June 23rd, 2009, 6:01 am

Try the Bayesian Information Criterion ( BIC ) Edit: I should probably add this technique will have problems if presented with strongly co-linear factors (slight changes in data can change the outcome - any selection technique will have this problem). In which case you might want to consider revising the available factors or using a Bayesian estimation approach using known descriptive information about the stocks to bias stocks toward factors that they are more intuitively associated with in candidate models where colinear factors are present.
Last edited by ACD on June 22nd, 2009, 10:00 pm, edited 1 time in total.
 
User avatar
edult
Topic Author
Posts: 49
Joined: October 31st, 2007, 8:24 pm

factor selection in factor models

June 24th, 2009, 1:23 am

What about comparing BIC calculated at one period with respect to the BIC in another period? Since the model will be built for a time period, it not only needs to give good cross-sectional fit, but also throughout time... I dont think you can average BIC or AIC, but you might be able to average R2
 
User avatar
ACD
Posts: 107
Joined: April 19th, 2004, 8:09 am

factor selection in factor models

June 24th, 2009, 6:20 am

You can convert a set of candidate models to posterior probabilities for assesment purposes, using this they are comparable across different time periods. Don't have a reference to hand but the formula to get the probabilities for a set of candidate models on a estimation is: exp(-0.5*BIC) / SUM(exp(-0.5*BIC))Where the sum is across all models and the numerator is the model you are interested in.
ABOUT WILMOTT

PW by JB

Wilmott.com has been "Serving the Quantitative Finance Community" since 2001. Continued...


Twitter LinkedIn Instagram

JOBS BOARD

JOBS BOARD

Looking for a quant job, risk, algo trading,...? Browse jobs here...


GZIP: On