<![CDATA[Fintech software development is characterized by rapid product iteration and stringent regulatory requirements, resulting in changing requirements as an interrelated set rather than individual elements. In this research, Sequential Requirements Prediction is proposed as a decision support task in Requirements Engineering, where the time-ordered prefix of completed backlog items is used to predict the next likely canonical requirement type as Top-k ranked output. To mitigate noise and inconsistency in backlog data, an LLM-aided semantic normalization step maps diverse requirement descriptions to a closed set of fintech requirement types. The research compares an interpretable rule-based Markov-1 predictor with Transformer-based sequential predictors under a case-level time-aware split. The proposed method is evaluated on a synthetic fintech-like backlog dataset consisting of 900 cases, 5,252 events, and 18 canonical requirement types. The best-performing model, Transformer + normalization + augmentation (M4), achieved Recall@5 = 0.638889, MRR@5 = 0.536806, and NDCG@5 = 0.566667. These results surpassed the rule-based predictor and non-normalized Transformer model. In addition, augmentation further improved Recall@5 from 0.493056 to 0.527778 in the rare-type subset. These findings suggest the methodological promise of the proposed framework for sequence-aware and compliance-conscious backlog analytics in synthetic fintech-like settings.]]>
Copyrights © 2026
