Optimization routine vs. data format
Posted: Tue May 24, 2011 1:02 pmDear dynare users,
I've been browsing through this forum for a while now and finally I decided to post my first message due to a problem I've encountered.
I'm trying to estimate a log-linearized small-scale DSGE model. A bunch of my first attempts were pretty successful in terms of finding a stable posterior mode (using csminwel), MH convergence and overall shape of the posterior distributions. Nonetheless, some of the results were a little surprising like for instance the interest rate smoothing parameter equal ca. 0.09-0.1.
So I realized I might have provided the interest rate data in a wrong format -> literally in percentages, while my model defines the variable R_t as a discount factor for the value of future bond holdings, i.e. I have B_{t+1}/R_t in the budget constraint. Therefore I figured I should redo the estimation using R_new = 1+R_old/100
And from then on, the whole trouble began. Csminwel was no longer able to find the posterior mode ("Error using ==> chol; Matrix must be positive definite"). So I started using "mode_compute=6" but first of all it takes ages (on a professional and highly efficient computer), and secondly and most importantly, the quality of those estimates are horrible. Absolutely no MH convergence, crazy shapes of posterior distributions, etc. And I'm talking about 800 000 simulations in several blocks here! Interestingly, however, the posterior parameters themselves seem to be nice and consistent with the findings of many published models.
Could anybody advise me on how I should proceed in this situation? What format of the interest data to use? How to solve these problems? I'll add that my priors are rather moderately dispersed around the mean values taken from published articles and working papers, while all of my observable variables are taken in logs and HP-filtered. I will very much appreciate any help!
I've been browsing through this forum for a while now and finally I decided to post my first message due to a problem I've encountered.
I'm trying to estimate a log-linearized small-scale DSGE model. A bunch of my first attempts were pretty successful in terms of finding a stable posterior mode (using csminwel), MH convergence and overall shape of the posterior distributions. Nonetheless, some of the results were a little surprising like for instance the interest rate smoothing parameter equal ca. 0.09-0.1.
So I realized I might have provided the interest rate data in a wrong format -> literally in percentages, while my model defines the variable R_t as a discount factor for the value of future bond holdings, i.e. I have B_{t+1}/R_t in the budget constraint. Therefore I figured I should redo the estimation using R_new = 1+R_old/100
And from then on, the whole trouble began. Csminwel was no longer able to find the posterior mode ("Error using ==> chol; Matrix must be positive definite"). So I started using "mode_compute=6" but first of all it takes ages (on a professional and highly efficient computer), and secondly and most importantly, the quality of those estimates are horrible. Absolutely no MH convergence, crazy shapes of posterior distributions, etc. And I'm talking about 800 000 simulations in several blocks here! Interestingly, however, the posterior parameters themselves seem to be nice and consistent with the findings of many published models.
Could anybody advise me on how I should proceed in this situation? What format of the interest data to use? How to solve these problems? I'll add that my priors are rather moderately dispersed around the mean values taken from published articles and working papers, while all of my observable variables are taken in logs and HP-filtered. I will very much appreciate any help!