Comparison of FDA based Time Series Prediction Methods
Tuomas Kärnä, Amaury Lendasse · 2007
Abstract. Functional Data Analysis (FDA) provides an important addition to traditional data analysis methods. In FDA a function is fitted to the data and the fitting coefficients are examined instead of the original data. The function fitting, however, is not a straight forward task. The choice of the function space is often crucial and the fitting may involve unknown parameters that need to be determined. This paper presents a comparison of different FDA based methods for time series prediction. The experimented function types are B-splines, Wavelets and Gaussian kernels. In all cases k Nearest Neighbor (k-NN) model is used for the prediction. Furthermore, some input selection methods are experimented to improve the k-NN performance. 1