A COMPARISON OF FORECASTING EFFICIENCY OF MULTIPLE REGRESSION ANALYSIS MODEL, WHEN DATA HAVE DIFFERENT VARIATION LEVELS.

Abstract

The objective of this research was to compare the forecasting performance of two estimation methods of multiple regression model with outliers, namely the Least Trimmed Square (LTS) and the Artificial Neural Networks (ANN). The coefficient of variation, C.V., represented the characteristic of the data which was categorized into 3 levels, i.e. the rather low level (C.V. <5%), the moderate level (5% < C.V. < 15%), and the rather high level (C.V. > 15%).The samples were the Natural Rubber Price in Thailand in 2005(C.V.=2.11%), Oxygen Consumption of Wireworm Larva in 1961 (C.V.=7.56%), and the saltiness-value levels of the sea water in the North Carolina’s Pamico straight region in 1980 (C.V.= 19.01%). The Mean Square Error (MSE) was used to measure the performance of the two methods.

From data analysis, it was found that the Natural Rubber Price in Thailand data (the rather low variation, C.V.=2.11%) and the saltinessvalue levels of the sea water data (the quite high variation, C.V.=19.01%) fitted by the ANN method had better performance than the LTS Method with the relative efficiency(R.E. : the ratio of MSE( )/MSE( )) of 1.353 and 1.563, respectively.

However the Oxygen Consumption of Wireworm Larva data (the moderate level variation,

C.V. = 7.56%) the ANN method had better performance than the LTS method with the R.E. about 6.

In summary, this study revealed that there was a trivial different efficiency between the LTS method and the ANN method when the data had the rather low and rather high variation. While the Oxygen Consumption of Wireworm Larva, which had the moderate level of variation, the ANN method gave much better performance than the LTS method.