There is a lot awry.
- 32 observations is not sufficient. [Update: see the clarification here [url]http://www.dynare.org/phpBB3/viewtopic.php?f=1&t=4723&start=0]
- Hp filtering removes the mean, but your observable variable is not mean 0
- Your model and your data are/were not in logs. Thus, they are not percentages but levels. This means that the scaling matters.
- Two-sided HP-filtering is non-causal and strongly discouraged for Bayesian estimation
Points two and three essentially mean that you need a correctly specified observation equation.