Some Indicators For Testing The Accuracy Of A Prediction Model And Codes In Matlab

Abstract

The accurate software cost estimation of the software project is an important issuein the software engineering community. To date there are build different algorithmic models of effort (especially when there are some independent variables). After that,the work is focused upon selection of the most efficient prediction model. Thegoodness of the model is expressed on high percentage of spreading of dependent variables and lessening the number of independent variables. The accuracy of the model measure via F statistics, T statistics, or another indicator which are functionof the model's relative errors. Usually, the efficient model used in practice is whichever give the less mean error without testing if this model is in fact statistically significant. This can lead to unstable (erroneous) results (conclusions).There are statistics used in practice which are conditioned from the data's probability distribution; when this is unknown, the test of hypothesis yields problems. In this paper, the application of non parametric criterion, like Wilcoxoncriterion, seem to be more reasonable. In this work is used this technique in order totest the significance between two prediction models: linear regression and log-linearregression. A program is written, tested and executed in MATLAB 6.5 the calculations, implementation and testing of those algorithms are performed withmathematics package MATLAB 6.5.