Browse Prior Art Database

Method to predict future shape of a time series within a fixed window using non-local information to a specified level of accuracy

IP.com Disclosure Number: IPCOM000185473D
Original Publication Date: 2009-Jul-27
Included in the Prior Art Database: 2009-Jul-27
Document File: 2 page(s) / 80K

Publishing Venue

IBM

Abstract

Disclose is a method to predict the approximate shape of the curve of a time series in a window in the future. This differs from conventional time series prediction, in which the idea is to predict an individual sample from past samples.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 66% of the total text.

Page 1 of 2

Method to predict future shape of a time series within a fixed window using non -local information to a specified level of accuracy

        Authors
Laurent S Mignet, Jaya Deva1

Fig 1 illustrates the concept. The problem involves predicting the shape of the region from time tk1 to time tk2, by using a fixed number of samples prior to tk1.

predict the shape of the region y

Given a set of equispaced samples y-k1, yk2-, … ykM,

k(M+1), … yk(M+L). The

assumption is that the number of past samples used (L) is sufficient to allow such a prediction to

be performed. The proposed approach is as follows.

1. The region from tk1 to tk2 is a function of time. The region to be predicted is approximated by a regressor, such as a support vector machine regressor (SVR). The inputs to the SVR are time (abscissa) within the window (i.e. starting from 0) and the value of the function at the corresponding time instant within the window.

2. The SVR approximates the shape to a desired level of accuracy (). The regressor is obtained as a weighted sum of L kernel entries. If a sparse representation is obtained, many of the weights would be zero.

3. The co-efficients of the SVR regressor now describe the shape of the function within the k-th window.

4. A second learning problem is now solved: the prediction of the weights as a function of M past samples. The training data for...