An Ensemble Transformer for Financial Product Purchase Prediction

Analyzing large-scale online behavioral data (e.g., clicking sequences) is the primary way to understand customer purchase intention in the realm of Fintech. Despite the development of sequential learning-based models, challenges in modeling this unique data type have limited their effectiveness. For example, the diverse activities in financial clicking sequence make it difficult to learn the short- and long-range latent dependencies among clicks. Transformers with self-attention seem to be a natural and reasonable choice; however, some unique characteristics of financial clicking sequences that may influence customer purchase intention prediction are still overlooked by Transformers. To address these issues, we propose an ensemble Transformer method based on a unified framework that incorporates clicking activities, dependencies within and across sequences. The method is tested using a real-world dataset and a field experiment, demonstrating its effectiveness compared to several cutting-edge benchmarks. The practical implications of the study in the Fintech industry are also discussed.