A multi-task dynamic deep survival network for corporate bond default prediction

Existing default prediction models predominantly focus on estimating the probability of default (PD) within fixed time horizons, often neglecting the time-to-default (TTD). While survival analysis captures TTD distributions, it implicitly assumes that all samples will eventually experience the event (i.e., default), which is often unrealistic in bond markets. To address this limitation, we propose the Multi-Task Dynamic Deep Survival Network (MT-DDSN), a joint learning framework that simultaneously predicts both PD and TTD. Operating as a multi-task framework, MT-DDSN unifies the default classification and survival analysis tasks through coordinated optimisation of prediction, likelihood, and ranking losses, thereby mitigating the error accumulation risks inherent in mixture cure models. Additionally, a feature-sharing layer acts as an information-constraining channel, facilitating continuous interaction between PD and TTD tasks to enhance representational capacity and model robustness. Furthermore, MT-DDSN integrates Long Short-Term Memory (LSTM) networks with a temporal attention mechanism. This architecture dynamically reweights LSTM hidden states to capture long-term temporal dependencies in bond and firm-level features, effectively mapping the complex transition of default risk from latent accumulation to observable exposure. Empirical evaluations using bond data from Chinese listed firms demonstrate that MT-DDSN outperforms baseline models in predicting both PD and TTD across various maturity segments.