Kaishev, V. K., Dimitrova, D. S., Haberman, S. & Verrall, R. J. (2004). Automatic, computer aided geometric design of free-knot, regression splines (Report No. Statistical Research Paper No. 24). London, UK: Faculty of Actuarial Science & Insurance, City University London.
Download (3MB) | Preview
A new algorithm for Computer Aided Geometric Design of least squares (LS) splines with variable knots, named GeDS, is presented. It is based on interpreting functional spline regression as a parametric B-spline curve, and on using the shape preserving property of its control polygon. The GeDS algorithm includes two major stages. For the first stage, an automatic adaptive, knot location algorithm is developed. By adding knots, one at a time, it sequentially "breaks" a straight line segment into pieces in order to construct a linear LS B-spline fit, which captures the "shape" of the data. A stopping rule is applied which avoids both over and under fitting and selects the number of knots for the second stage of GeDS, in which smoother, higher order (quadratic, cubic, etc.) fits are generated. The knots appropriate for the second stage are determined, according to a new knot location method, called the averaging method. It approximately preserves the linear precision property of B-spline curves and allows the attachment of smooth higher order LS B-spline fits to a control polygon, so that the shape of the linear polygon of stage one is followed. The GeDS method produces simultaneously linear, quadratic, cubic (and possibly higher order) spline fits with one and the same number of B-spline regression functions. The GeDS algorithm is very fast, since no deterministic or stochastic knot insertion/deletion and relocation search strategies are involved, neither in the first nor the second stage. Extensive numerical examples are provided, illustrating the performance of GeDS and the quality of the resulting LS spline fits. The GeDS procedure is compared with other existing variable knot spline methods and smoothing techniques, such as SARS, HAS, MDL, AGS methods and is shown to produce models with fewer parameters but with similar goodness of fit characteristics, and visual quality.
|Item Type:||Monograph (Working Paper)|
|Uncontrolled Keywords:||spline regression, B-splines, Greville abscissas, CAGD, free-knot splines, control polygon|
|Subjects:||H Social Sciences > HG Finance|
|Divisions:||Cass Business School > Faculty of Actuarial Science & Insurance > Faculty of Actuarial Science & Insurance Statistical Research Reports|
Actions (login required)
Downloads per month over past year