Long time lurker, first time poster, thanks in advance for your help
I am trying to estimate output gaps through a Kalman Filter. The model is fairly simple, it includes a Philips curve signal equation (with some controls), another signal equation for the output gap, state equations for the unobserved variables and the dynamics of them.
The SSpace formulation is the following:
Code: Select all
@SIGNAL INFLATION_RATE = C(1) + C(2)*INFLATION_RATE(-1) + C(3)*Y_GAP + C(4)*USD + C(5)*WTI + [VAR = EXP(C(6))]
@SIGNAL Y = Y_GAP + YSTAR
@STATE YSTAR = YSTAR(-1) + [VAR = 0.1]
@STATE Y_GAP = Y_GAP(-1) + Y_GAP1(-1) + [VAR = EXP(C(7))]
@STATE Y_GAP1 = Y_GAP(-1)
PARAM C(1) .0 C(2) .0 C(3) .0 C(4) .0 C(5) .0 C(6) .0 C(7) .0
Once I run the model the smoothed series somewhat agree with what we expect to see for output gap, however, I noticed that the smoothed series explodes towards the end (the last few years). I was wondering if any of you have any opinion or insight on why this may be happening?
At the same time, if I truncate the estimation to a smaller sample (let’s sat to 2010, series in blue below), the model also explodes towards the end of the series.
I suspect that the explosion could be related to either:
A)The AR specification for the output gap (tried to include higher AR regressors but the series explodes more), with a lower AR number of regressor the model "breaks" (in the sense that the smoothed variables just flat line)
B)Someone suggested that the problem may be a timing issue of the shocks and how they are interacting with the state vars. ---> Not sure how to "test" this, I added a different set of control variables (instead of USD and WTI; food energy inflation and import prices index), with very similar results (in the sense of an exploding smoothed series towards the end).
I am aware that SSpace models are very flexible and high sensitive to their specification, but I cannot seem to understand completely why the behavior of the output gap explodes towards the end of the series. Any input from all of you is highly valuable, happy to share further details if needed.
Thanks in advance
----
Skyboard
PS: Using Eviews 10, estimating the sspace model by OPG-BHHH Marquardt (Covariance method: Ordinary-OPG)