This study examines how generative adversarial network (GAN)-based synthetic-data generation can be used for money collection histories. Two provided synthetic historical collection datasets are used in this study. Customer records are organized as irregular histories: some customers have one observed collection state, while others have multiple collection states ordered mainly by delinquency days and related repayment, balance, and collection-action information. The workflow reviews data quality, removes leakage-prone fields, builds ordered sequence windows, and develops MoneyCollectionTimeGAN, a compact sequence-generation model that adapts selected TimeGAN ideas to generate collection behavior windows. Simpler prediction and forecasting checks are kept as supporting evidence. Related GAN methods for tabular data and minority-class augmentation provide background for the design choices. The generated histories are evaluated using reconstruction and temporal losses, separability AUC, statistical and temporal validation, proxy prediction tests, range-validity checks, and duplicate screening. Results show meaningful but partial temporal similarity, with generated sequences still distinguishable from the provided synthetic input process. The findings are reported as bounded synthetic-to-synthetic evidence rather than validation on live operational customer data.

Keywords

Generative adversarial networks; Money collection; Synthetic data generation; Collection analytics