A Longitudinal Attribute-Conditioned Neural Network for Modeling Health-State Transition Probabilities in Temporally Irregular Data: The LANTERN Framework

Battle Count: 75

The paper is primarily focused on actuarial science and long-term care insurance modeling. While it involves probabilistic forecasting, calibration, and risk stratification concepts that have parallels in quantitative finance, the domain (health-state transitions, disability insurance) is quite distinct from trading. The multi-state transition matrix approach and calibration methodology could loosely inform regime-switching or state-transition models in trading, but direct applicability is minimal.

The architecture combines multiple components: GRU recurrent memory, Time2Vec temporal encoding, multi-head attention for attribute conditioning, and a hierarchical output layer separating mortality from alive-state transitions. Training requires careful handling of irregular time intervals, class imbalance, and proper aggregation into transition matrices. The model has moderate parameter count (latent dim 128, Time2Vec dim 8, 4 attention heads) but the overall pipeline from data preparation through aggregation to actuarial projection adds complexity.

Code is publicly available on GitHub (https://github.com/BrightManu-lang/lantern-health-prediction). Data is from the publicly available RAND HRS Longitudinal File. Fixed random seed (42) is used. Training hyperparameters are fully specified. However, some architectural details (exact layer sizes, initialization) may require code inspection for full reproduction.