Customer requests matter: Early stage software effort estimation using k-grams

Estimating the software development effort associated with a customer request, immediately after the request has been made, is quite challenging for project managers. Studies on software effort estimation often utilize expert knowledge to build statistical or machine learning models, although it is subjective and human-dependent. Rich text, natural language based descriptions provided by the customers are mostly not incorporated into the models during estimation. The aim of this study is to propose an early stage software effort estimation model that can predict the effort of the related problem immediately after a customer request or bug fix problem is appeared. Our model utilizes the textual descriptions of customer requests collected in the requirement management tool, and other features stored with the customer requests. Our results show that the use of textual data helps to make better predictions for an effort estimation system. Besides the effect of textual data, we asked whether there is a significant difference between the performance of an effort estimation system that can be used for all customers (Unified Model) and Customer Specific Models which is trained using only the related customer's requests. The Pred(25), Standard Accuracy (SA) and Gibbs' A were used during the evaluation. The Unified Model for early stage effort estimation achieves 43% pred(25) and 51% SA values. Customer Specific Models, on the other hand, depending on the customer, obtain 39%-58% Pred(25), and 35%-58% SA values. The results show that there is no significant advantage between the Unified Model and Customer Specific Models, and which model to use should be determined according to the customer.