LINEST & How Nonlinear is it?

[Jo]

Hi everyone,

You have a set of data in one row or one column.

Next, you want to do nonlinear regression analysis for forecasting
purposes and you use LINEST.

The question is:
Should someone keeps dancing back-and-forth inbetween different
nonlinear polynomials without knowing which one is more accurate? I
mean how to do you measure the inaccuracy level?

Any advice or feedback on this please?

Thanks,
Jo

 

[Bernard Liengme]

What is the source of the data? Is accuracy really an issue here?
Be aware that if you have n data points (say 4) then you will always get a
exact fit with an n-1 (third) order polynomial
The quantity to check is R2 - look in Help to see where LINEST reports this.
R2 value of 1 indicates a 'perfect' fit; Now I await Jerry Lewis's comment
<gr>
best wishes

 

[Jo]

What is the source of the data? Is accuracy really an issue here?
Be aware that if you have n data points (say 4) then you will always get a
exact fit with an n-1 (third) order polynomial
The quantity to check is R2 - look in Help to see where LINEST reports this.
R2 value of 1 indicates a 'perfect' fit; Now I await Jerry Lewis's comment
<gr>
best wishes
--
Bernard V Liengme
Microsoft Excel MVPwww.stfx.ca/people/bliengme
remove caps from email











- Show quoted text -

So, if I have 10 data points, I should consider 9th polynomial? How r2
is calculated? I read MS help but it is not that clear on it!

 

[Bernard Liengme]

If you have 10 data points a ninth-order polynomial will give R2=1
But that does not tell you very much - it is just a mathematical thing; you
need to consider the meaning of you data. Generally, unless you KNOW
beforehand that the data should fit a certain polynomial there is not much
to gain by just getting a fit to a high-order equation!
How is R2 computed? You should Google or read a stats book
best wishes

 

[Guest]

While R^2 would be 1, that fit will ususually be totally useless for
forecasting. Often a more useful measure is "adjusted R^2"
http://en.wikipedia.org/wiki/Coefficient_of_determination#Adjusted_R2
which includes a penalty for increasing model complexity without an adequate
corresponding increase in explanatory value.

Jerry