I am using VECM for establishing the relationship between ECB and exchange rate, interest rate, GDP, inflation, trade, and risk. The lag came 1 as per SIC and HQ and applying VECM at lag 1, R is coming 20% and adjusted R is 13 % but when I am taking lag = 2 then R is coming 45% and adjusted is 25%. at lag 3 model is not coming significant. can I take lag 2 on the basis of higher R? in both cases there is no serial correlation and heteroskedasticity but is not normal. the sample size is 64 quarters.
How is your conceptual model looking? What are the depending and independent variables?
The dependent variable is external commercial borrowing and the independent variable are differential interest rate, exchange rate, inflation, GDP, trade (import+export), and risk of the country.
Would be great to see graphs of time series. If the distribution of residuals is not normal, perhaps you can use data transformations. As for the criteria, SIC is a good indicator. In addition, I would recommend trying out-of-sample MAE or AvgRelMAE described here:
Chapter Forecast Evaluation Techniques for I4.0 Systems
Generally, visual analysis of your data may help double-check if modelling is correct and is based on plausible assumptions:
Poster A Visual Framework for Longitudinal and Panel Studies (with ...
i have done log transformation now model is coming normally distributed
now the model is robust as there is no serial correlation, no heteroscedasticity, and also normally distributed only R square value is coming 20% and adjusted r is 10%. does a model with a low r square is significant?
Glad the method worked. What are R2 and SIC values for alternative models?
I think lag 1 is OK, taking lag 2 only on the basis of R2 is not advisable as this may lead to over-fitting. Generally, R2 is not a good indicator for choosing the model specification.
@ Andrey Davydenko
Does r square matter in VECM model ? Because it is coming only 23 %. Although variables in the long run are coming significant.
Shilpa Gupta No, I would not recommend selecting models based on R2. It does not give you a clear interpretation of modelling performance. Perhaps, you will be interested in these works where more appropriate metrics are described:
A brief description of data formats and error metrics for comparing alternative models:
Chapter Forecast Evaluation Techniques for I4.0 Systems
A brief intro to error metrics:
Chapter Forecast Error Measures: Critical Review and Practical Recommendations
A more detailed version of the above chapter:
Chapter Measuring Forecasting Accuracy: Problems and Recommendations...
Most of the methods presented in the above chapters are described in the settings of time series setup, but easily adjusted to panel data (see the first ref)
since VECM models long-run relationships shouldn't we use a larger no of lags?
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression