Publicação
Adversarial generative forecasting of daily Fraud Amount for sparse transaction time series
| Resumo: | In collaboration with SIBS, this project forecasts daily accepted fraud amounts in e-commerce transactions to support proactive risk management. Utilizing a dataset ofover 166 million transactions (2023–2024), we engineered behavioral features to benchmark multiple machine learning models. XGBoost was the champion model, achievinga 16.56% MAPE, but struggling during volatility spikes. To improve reliability duringspike periods, we investigated three complementary strategies: GAN-based data augmentation to increase exposure to synthetic high-fraud scenarios, quantile-based forecasting to model the upper tail of the distribution, and transfer learning approaches that adapt large pre-trained time-series models to our use case. |
|---|---|
| Autores principais: | Mueller, Moritz |
| Assunto: | Fraud forecasting Time series XGBoost Generative adversarial neural network Transfer learning Value at risk Quantile regression Extreme value theory |
| Ano: | 2026 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | In collaboration with SIBS, this project forecasts daily accepted fraud amounts in e-commerce transactions to support proactive risk management. Utilizing a dataset ofover 166 million transactions (2023–2024), we engineered behavioral features to benchmark multiple machine learning models. XGBoost was the champion model, achievinga 16.56% MAPE, but struggling during volatility spikes. To improve reliability duringspike periods, we investigated three complementary strategies: GAN-based data augmentation to increase exposure to synthetic high-fraud scenarios, quantile-based forecasting to model the upper tail of the distribution, and transfer learning approaches that adapt large pre-trained time-series models to our use case. |
|---|