Modified nearest neighbour algorithm for time series forecasting.

Rahat Abbas; Wajid Aziz; Muhammad Arif

Time series forecasting is the one of the important field of computer sciences in which existing values of outcomes of a system is taken and future values are forecasted. Nearest neighbour is an algorithm used for time series forecasting in. This algorithm the neighbouring data points of a series is compared with the previous data points of the series and Euclidean distance is calculated for all the trials. The best fit pattern in the data point of the series is chosen and its output is considered the forecasting of current pattern of time series. The series is split into segments of some fixed length. This fixed length is called the window size. The efficiency of forecasting depends very much on the window size. If it is too small or too big the forecasting error becomes higher. The window sized is often taking by hit and trail. In the current study the optical window size is suggested using autocorrelation function of the series. It is found that a maximum of autocorrelation function determines the optimal window size for nearest neighbour algorithm. The concept is tested on three benchmark data sets i.e. sunspot time series, Mackay-Glass time series and Santa Fee laser time series. The limitation of nearest neighbor algorithm is that if a pattern does not resemble well with any pattern in the past the forecasting error is much larger. In the current study the correction factor based of energy of the signals is introduced in standard nearest neighbour algorithm to overcome this problem. The correction factor improved forecasting results.