Testing Model Robustness & Repredicting Values
Comparing Actual Data to Repredicted values
How?
- Record the actual data values that you are intending to remove.
- Then delete out some data values at the end of the time series.
- How?
- Filter data
- Select cases,
- Select Row number, proceed (and enter in row numbers)
You will need to check the row numbers before you start
Then proceed.
Redo the predictions (estimates) as normal
Compare the Estimates and the Actual data values (with considerations of the 95% prediction intervals)
Discuss this (in context)
Actual Data
iNZight Predictions
The iNZight predictions(Aug 5100, Sept 5600, & Oct 7100 visitors) are all greater than the actual data (Aug 3949, Sept 4331, & Oct 6446 visitors from Germany).
Only the October actual data value (6446 visitors) was within the 95% prediction interval for the estimates (6070 to 8190 visitors). The values for August & September were both below the lower limit of the 95% prediction interval.
This indicates that the iNZight Holt-Winters LOWESS technique is not that robust in this situation. The removal of the last three data values caused the trend line to change to increase after 2012, thus producing higher estimates than the actual data, as can be seen in the graph below:
Example: Tourist numbers to NZ from Germany
Excel Predictions
iNZight Predictions
Testing Robustness
(Comparing predicted vs actual values - A video by Priscilla Allan)