Predicting the Duration of User Stories, Machine learning for agile planning

Abstract

Effective effort estimation in project planning is vital, because it helps organizations to build product plans which they can stick to, have shorter turn-around time and better cost discipline. In this thesis, a series of supervised machine learning models were studied, analyzed and implemented to solve the problem of predicting effort estimate in Agile Scrum. The main approaches used were, Term Frequency - inverse document frequency (TF-idf), fastText, Neural Networks (Recurrent Neural Network (RNN), Long Short Term Memory (LSTM)) and Bidirectional Encoder Representations from Transformers (BERT). The models were fitted with two publicly available datasets. The fastText (with pre-trained model) significantly performed better in predicting the story-points of user-stories. The second-best performing model was bidirectional- LSTM. distilBERT performs poorly among all the models analyzed. This study can pave a way for organizations to benefit from these machine learning models and pre- dict accurately the project deadlines and schedules. This could help organizations to get a head of their competitors and have happy customers.