On Sun, Aug 24, 2014 at 6:38 PM, David Roberson <[email protected]> wrote:
You also probably realize that a polynomial fit to a high power order > yields coefficients that vary depending upon the order of the polynomial > chosen. Many combinations of coefficients will fit the input/output data > over a restricted range. The problem shows up once you use those > different coefficients to project the curve forwards into unknown future > points. > > We are now clearly in witness to an example of the type of problem that I > am speaking of. ... > I think the bad fit to the data you identify could just as likely be an underfit than an overfit; i.e., they have adequately modeled the first-order phenomenon (an increase in temperature) but failed to take into account one or more second-order cyclical trends. Eric
