Enhancing Survival Analysis of Recurrent Events with Transformers: A Financial Prediction Approach

Recent deep learning models like WTT-RNN, Deep Recurrent Survival Analysis (DRSA) and CmpXRnnSurv-AE have advanced survival analysis but struggle with long-range dependencies and time-varying relationships in financial data, especially for irregular events like credit defaults. To overcome this, we propose a Transformer-based model for survival analysis of recurrent events, focusing on variable gap times between events. Using self-attention mechanisms, the model captures complex patterns in event sequences without predefined temporal assumptions. Time-encoding techniques account for dynamic gap times, and the model is trained with Cox partial likelihood and a custom survival loss to handle censored data. Applied to the Credit Card Payment Default Prediction dataset, our model outperforms WTT-RNN, Deep Recurrent Survival Analysis (DRSA) and CmpXRnnSurv-AE in predictive accuracy, validated by the Concordance Index (C-index), making it a promising solution for survival analysis in financial settings.

کلیدواژه ها:

Recurrent Event Data ، Transformer Neural Network ، Gap Times