Tab-TRM: Tiny Recursive Model for Insurance Pricing on Tabular Data

Battle Count: 108

The paper is primarily focused on insurance pricing and actuarial modeling rather than quantitative trading. However, the recursive latent reasoning architecture and state-space interpretation could potentially be adapted for financial risk modeling, credit scoring, or other tabular prediction tasks in finance. The Poisson deviance framework and GLM connections are more relevant to insurance than trading. The recursive refinement paradigm shares conceptual similarities with iterative portfolio rebalancing but is not directly applicable to trading strategies.

The architecture is conceptually elegant but involves several components: feature tokenization (piecewise-linear encodings for continuous, entity embeddings for categorical), recursive core with inner/outer loops, residual connections, LayerNorm, and a decoder with exponential activation. The networks themselves are tiny (zero hidden layers optimal), but the recursive structure, hyper-parameter tuning via Optuna, and nagging ensembles add practical complexity. The algorithm is well-specified with pseudo-code, making implementation feasible for experienced ML practitioners.

The paper provides detailed algorithm pseudo-code (Algorithm 1), complete hyper-parameter search space (Table 4), specific architecture details, and uses publicly available benchmark data (French MTPL from Dutang et al., 2024). The training setup, optimization procedure (AdamW, Optuna), and evaluation protocol are fully described. However, no direct code repository link is provided for Tab-TRM implementation.