Improved software cost estimation. A robust and interpretable modeling method and a comprehensive empirical investigation

Delivering software on time, within budget, and to an agreed level of quality is crucial for the business reputation and competitiveness of many organisations. Accurate estimates are essential for a better project planning, tracking, and control and pave the way for successful product delivery. This calls for support in software cost estimation, since many software companies are working within tight schedules and finish their projects behind schedule and budget, if they finish them at all. This is because the estimation of cost for software projects leads to many practical measurement and modelling difficulties. To overcome limitations, this research proposes an enhancement of the method Optimised Set Reduction (OSR) to solve estimation problems. The goal is to provide an estimation method that overcomes the main drawbacks of other methods and that produces models with an accuracy comparable to other commonly used models. The OSR method is originally defined to solve classification problems and is based on machine learning and robust statistics (Briand et al., 1992). It exploits the rigor of statistical modelling and still generates easy-to-interpret, rule-based models.