REGRESSION MODELS FOR ESTIMATING THE DURATION OF BANKING SOFTWARE DEVELOPMENT

Abstract

The paper addresses the problem of improving the estimation accuracy of software development duration in a tightly integrated environment of banking information systems, which is critically important for effective project planning, optimal resource allocation and ensuring business success. It is shown that existing approaches to estimating software development time, in particular the COCOMO and ISBSG models, do not provide sufficient accuracy for the specific conditions of the banking domain, which is characterized by high architectural complexity, integration of numerous external services, regulatory requirements and increased security demands. Based on data from previous studies and analysis results, the use of nonlinear regression models with prior data normalization for estimating the duration of banking software development based on project effort is justified. An approach is proposed that involves constructing separate models for single-team and multi-team projects, which allows the model to account for additional time costs due to inter-team interaction, work synchronization and organizational complexity. The empirical basis of the study consists of data from 482 real banking software development projects for the period 2014–2024, developed in Java and JavaScript. A methodology for constructing nonlinear regression models is implemented, which includes data normalization using decimal logarithm transformations, the bivariate Box-Cox transformation and the univariate Johnson transformation, detection and removal of outliers based on the prediction ellipse, construction of linear regression for normalized data, derivation of nonlinear models through inverse transformations, construction of confidence intervals and prediction intervals and identification of outliers using prediction intervals. The quality of the models is evaluated using the R², MMRE and PRED(0.25) metrics. The obtained results indicate that models constructed using logarithmic transformation are unsuitable due to a significant number of outliers. In contrast, models based on the Box-Cox and Johnson transformations demonstrate acceptable accuracy: R² values exceed 0.85, MMRE does not exceed 0.25 and PRED(0.25) is above 0.75 in most cases. The best results are achieved by models using the bivariate Box-Cox transformation, which are characterized by a smaller number of outliers and better agreement with empirical data. The proposed models can be used to improve the accuracy of predicting the duration of banking software development in Java and JavaScript based on project effort. At the same time, the limitations of the approach are identified, in particular the dependence of the models on the specifics of the application domain and the limitation of the range of project effort values.