# Forecasting noisy data with polynomials

What's the best way to incorporate polynomials or sinusoidal equations into a linear model to predict a future amplitude in this sort of data? You can see that if those polynomials continued to go on, they wouldnt be anywhere close to fitting the final location. Im trying to use the greeen shaded area to predict the non shaded area.
Those light grey lines are polynomials like this: $$y_1 = a_1(x)^2 + b_1(x) + c_1\\$$ that are fitted to different sections of the data. Do I just use the a b and c variables from each polynomial, or do extrapolate out and use some other value?

• This is a very broad question. Can you give any details on what kind of data this is? May 8 '14 at 3:03
• Home price data. I'm trying to look outside of conventional methods to improve predictions. May 8 '14 at 3:05
• Can you post the dataset? Basically I think that polynomials are no good for prediction/extrapolation of data. They tend to variate a lot outside of dataset which you are fitting. There are other technique I could propose, but posting dataset will be helpful.
– jojek
May 8 '14 at 7:27

Starting with your last question: to extrapolate you do just use your fitted coefficients and extend them to a greater $x$ range.
However, I think this approach is inherently flawed for data such as yours over any significant time period, especially when using a polynomial model. The problem is that any polynomial has a limited number of turning points. Once you pass these the data will tend to head off to $\pm \infty$ which is unlikely to model anything well.