q**j
发帖数: 10612 | 1 打算得到比较smooth (stable)的mean and covariance estimates。现在mean用了baye
sian model，已经相对比较smooth了。但是Bayesian的covariance也不smooth，而且牵
涉到wishart distribution，在mle的时候非常volatile。请问这方面的大侠有什么指教
么？
当然还是要efficient的estimator。多谢了。 |
y*****y
发帖数: 98 | 2 your question is unclear. please describe your problem and your approach in
a logic manner. |
q**j
发帖数: 10612 | 3 sorry. There are T years and N observation with K independent variables in e
ach year. I can get a set of K estimates each period. My problem now is the
get a smooth estimate for the covariance of the K parameters, ie. a K by K m
atrix.
Sample covariance is not stable. Most Bayesian model innovate on the mean, w
ith less attention on covariance. When they actually focus on covariance, th
ey use Wishart distribution, which is very hard to optimize due to the gamma
function, when I try to find the optimal degree of freedom for Wishart.
This is my current problem and my experience. Any suggestins would be highly
appreciated.
in
【在 y*****y 的大作中提到】
: your question is unclear. please describe your problem and your approach in
: a logic manner.
|
y*****y
发帖数: 98 | 4 the K variables are independent, why do you need a covariance matrix?
anyway. the covariance estimation could be very hard due to large K and
small N. the Bayesian approach is to put a prior structure (conjugate
Wishart or objective prior etc.). but i don't think that the Wishart prior
is difficult if there is no constraint. everything is conjugate. estimation
should be pretty straightforward.
by the way, you have time associated. the model should be temporal. |
q**j
发帖数: 10612 | 5 the model is for srue temporal. i have aobut 400 time period to estimate a
12 by 12 matrix, and i want the estimate to be smooth. if this is too large,
i can reduce the matrix to 8 by 8. any good suggestions? i am open to any
suggestions, bayesian model or not.
estimation
【在 y*****y 的大作中提到】
: the K variables are independent, why do you need a covariance matrix?
: anyway. the covariance estimation could be very hard due to large K and
: small N. the Bayesian approach is to put a prior structure (conjugate
: Wishart or objective prior etc.). but i don't think that the Wishart prior
: is difficult if there is no constraint. everything is conjugate. estimation
: should be pretty straightforward.
: by the way, you have time associated. the model should be temporal.
|
