The scientific basis for prediction research

Abstract

In recent years there has been a huge growth in using statistical and machine learning methods to find useful prediction systems for software engineers. Of particular interest is predicting project effort and duration and defect behaviour. Unfortunately though results are often promising no single technique dominates and there are clearly complex interactions between technique, training methods and the problem domain. Since we lack deep theory our research is of necessity experimental. Minimally, as scientists, we need reproducible studies. We also need comparable studies. I will show through a meta-analysis of many primary studies that we are not presently in that situation and so the scientific basis for our collective research remains in doubt. By way of remedy I will argue that we need to address these issues of reporting protocols and expertise plus ensure blind analysis is routine.