Trend and initial conditions

• Standard practice: ``Eliminate'' the trend and mean, Then pretend the detrended data is raw data.
  • Can seriously distort fit comparisons between DSGE's that assume stationarity and BVAR's (or DSGE's that model growth).
  • Fit is no longer connected to recursively updated forecasting performance.
• Better practice
  • Ideally, the DSGE models trend explicitly.
  • Second-best: The estimation of detrending parameters is done jointly with other parameters, rather than as preprocessing step.