The point about 6th order fits is not that Excel's differ from those
calculated by other packages,
Although for some difficult problems they do! Historically in old XL
versions the polynomial fit used inside the old graph charting
software had significantly better numerical stability than LINEST. And
I honestly would not trust LINEST with anything beyond a cubic
polynomial in any version of Excel. I have seen it fail too often.
but that these fits are not based on any
models of physical behavior, but rather on conforming "nicely" to the
measured points. If you are interpolating by using the fit, you will get
reasonable predictions. If you extrapolate too far beyond the range of
observations, you could be very far from the true behavior.
It also depends how many measured points you have and their
distribution. If there are too few points then you can get an
excellent least squares fit *at the specified data points* and wild
oscilations inbetween.
It depends very much on what you are doing. Usually it is better to
use a physical model of the problem in hand rather than fitting a
generic high order polynomial. In some circumstance you know on purely
physical grounds that the calibration should for example contain only
odd or even powers of the dependent variable.
Regards,
Martin Brown