[FairmontGS]

Originally Posted by russellw

Sorry I missed the question. Yes, the difference as explained by FoxtrotGolfXray 5.0 is the shift from 3rd to 4th order polynomials in NSW to reflect the start of a downward trend. Victoria is still on 3rd order and wouldn't change much even if I selected 4th,

The reason to change is that a polynomial trend works best with 'spiky' trends or where there are clearly defined 'shift's in the pattern - thus, as there is no such thing as first order, a pattern with one shift is 2nd order, 2 shifts is 3rd order etc. etc. NSW now has it's 3rd shift which is why the change to 4th order because the 3rd order trend would continue to rise even though the raw data isn't.

Here is NSW from yesterday using 3rd and 4th order polynomial trends:

image

image

Originally Posted by FoxtrotGolfXray 5.0

I see Russ hasn't had the opportunity to respond to your post, FGS, so I hope both you and he don't mind if I offer my opinion, based on my experience with using excel to do data analysis.

The trend line you're seeing is based upon the full data set, so the second chart (yesterday's trend) shows an extra day of data, which is continuing to decrease in magnitude. Therefore the trend line is going to predict what is happening based upon that continuing trend.

Also note that the two trends are also based on different level order polynomials; the trend line from Monday is based on a third order polynomial and yesterday's a fourth order. My take on that is the 3rd order doesn't as easily allow for a peak and decrease in the data as well as a fourth order. We can clearly see that the actual data is starting to flatten out the curve; I'd hazard a guess that the third order polynomial will continue to rise based on that revised data set as opposed to the 4th order which will cater for the easing of numbers.

I suspect Russ has moved to a 4th order poly to better 'fit' the trendline to the actual data. Remember the prediction of the trendline is limited to the capability of the equation used to best fit the actual data.

This page here provides you an idea as to what level capabilities each order poly has in best modelling the actual data set.

Hope that makes some sense?

Many thanks both. I guess I was working on the assumption that the trend was predictive and didn't need a shift of gear to the next phase but I understand better now.