Trouble with Time Series ARDL Diagnostics
Hi everyone,
I am writing an undergrad Economics dissertation and am struggling massively with ARDL postestimation diagnostic tests - specifically the archlm and sktest for residuals, as well as the (Ramsey) ovtest for model misspecification.
I am exploring the long-run relationship between geopolitical risk (measured by Iacoviello's GPR index) and nominal EX vs. USD. I have around 270 observations for 4 different countries, namely UK, Japan, Canada and Denmark. Within each, i set up 5 different well-accepted monetary models for EX determination and augment them with the GPR index. All variables, bar interest rate and inflation differentials, are log-transformed (as is best-practice across the literature).
Using an ARDL framework and Pesaran's bounds test approach (my variables are a combo of I(0) and I(1)), I have found sufficient evidence of cointegration for at least 3 models in Denmark, Norway and Japan (nothing for UK). Where cointegration is found, I then proceed to estimate long-run elasticities via an ECM setup.
I have two main questions/issues if anybody could please help:
1. All diagnostics (B-G serial correlation, B-P heteroscedasticity, CUSUM, Jarque-Bera, Archlm, Ramsey RESET) are all passed satisfactorily for all my ECM models in Japan and Denmark (except a few cases where Jarque-Bera just fails - though I understand this is not too serious an issue). However, in Norway, across all models, the Jarque-Bera, archlm and Ramsey RESET tests fail, and the results get worse in the models with more control variables. In the most populated model, the Breusch-Godfrey serial correlation test even nearly fails (p=0.06) - which I understand to be crucial to reliability of ARDL estimation.
Can anybody please help by suggesting why this might be the case, and what I might do about it? I am very worried about getting this sorted before my deadline!
2. I am unable to find sufficient evidence of cointegration using the bounds test for any model for the UK. When I try to estimate the short-run effects using an ARDL model (lag-choice guided by AIC) the coefficient on GPR lacks significance.
Are there any methods anyone can think of that might improve the significance of my short-run estimate?
If not, can anybody suggest a way I might write up/describe such a result in my paper? I could obviously just drop the UK altogether, but that feels like it lacks academic integrity!
Thanks so much,
Jacob
I am writing an undergrad Economics dissertation and am struggling massively with ARDL postestimation diagnostic tests - specifically the archlm and sktest for residuals, as well as the (Ramsey) ovtest for model misspecification.
I am exploring the long-run relationship between geopolitical risk (measured by Iacoviello's GPR index) and nominal EX vs. USD. I have around 270 observations for 4 different countries, namely UK, Japan, Canada and Denmark. Within each, i set up 5 different well-accepted monetary models for EX determination and augment them with the GPR index. All variables, bar interest rate and inflation differentials, are log-transformed (as is best-practice across the literature).
Using an ARDL framework and Pesaran's bounds test approach (my variables are a combo of I(0) and I(1)), I have found sufficient evidence of cointegration for at least 3 models in Denmark, Norway and Japan (nothing for UK). Where cointegration is found, I then proceed to estimate long-run elasticities via an ECM setup.
I have two main questions/issues if anybody could please help:
1. All diagnostics (B-G serial correlation, B-P heteroscedasticity, CUSUM, Jarque-Bera, Archlm, Ramsey RESET) are all passed satisfactorily for all my ECM models in Japan and Denmark (except a few cases where Jarque-Bera just fails - though I understand this is not too serious an issue). However, in Norway, across all models, the Jarque-Bera, archlm and Ramsey RESET tests fail, and the results get worse in the models with more control variables. In the most populated model, the Breusch-Godfrey serial correlation test even nearly fails (p=0.06) - which I understand to be crucial to reliability of ARDL estimation.
Can anybody please help by suggesting why this might be the case, and what I might do about it? I am very worried about getting this sorted before my deadline!
2. I am unable to find sufficient evidence of cointegration using the bounds test for any model for the UK. When I try to estimate the short-run effects using an ARDL model (lag-choice guided by AIC) the coefficient on GPR lacks significance.
Are there any methods anyone can think of that might improve the significance of my short-run estimate?
If not, can anybody suggest a way I might write up/describe such a result in my paper? I could obviously just drop the UK altogether, but that feels like it lacks academic integrity!
Thanks so much,
Jacob
