Nicholas,
For the last part of the lab where we do it on our own with the log transform -we are supposed to choose the best ARMA model, and that’s all? We don’t need to do the forecasting part -Is that correct?
No – you should do the forecasting part. Part of model identification is choosing a model that gives reasonable results. For example, the model identification step on the first part chose a Seasonal AR model, but it was clearly inadequate once forecasting was attempted. Part of showing that a model fits a time series well is showing that it gives realistic results when used for prediction. If it doesn’t give realistic results, then there is some aspect that it is not fitting.
So once we choose one of the ARMA models using the log transformed data, we then transform the data back using the exp function, and then we do the forecasting on the back transformed data?
Is that correct??
Thanks,
Dave
Nicholas,
For the last part of the lab where we do it on our own with the log transform -we are supposed to choose the best ARMA model, and that’s all? We don’t need to do the forecasting part -Is that correct?
Thanks,
Dave
No – you should do the forecasting part. Part of model identification is choosing a model that gives reasonable results. For example, the model identification step on the first part chose a Seasonal AR model, but it was clearly inadequate once forecasting was attempted. Part of showing that a model fits a time series well is showing that it gives realistic results when used for prediction. If it doesn’t give realistic results, then there is some aspect that it is not fitting.
So once we choose one of the ARMA models using the log transformed data, we then transform the data back using the exp function, and then we do the forecasting on the back transformed data?
Is that correct??
Thanks,
Dave
No, when you transform data you need to do ALL of the analysis with the transformed data. The only thing you back-transform for is plotting.
So, do your forecasting, and then backtransform when you plot.