Seyedreza Tavakoli | Searcharxiv

arXiv2026

RiskReason: A Benchmark for Verifiable Financial Risk Reasoning

Robust financial-risk inference requires more than mathematical correctness. Models must accurately apply conditional rules, maintain evolving financial states, and make decisions that respect operational constraints. This work introduces RiskReason, a benchmark for verifiable financial-risk reasoning that spans diverse risk-management mechanisms. The benchmark includes Standard tasks, which assess individual financial mechanisms, and Long-horizon tasks, which combine multiple mechanisms through extended dependencies. Each task provides reference outcomes, formal specifications, reasoning annotations, and dependency representations to support systematic analysis of model performance. Eleven model configurations are evaluated using strict single-attempt accuracy, complemented by structural and qualitative failure analyses. The strongest configurations achieve 80.0--90.2\% overall accuracy, with several maintaining high performance on composed financial scenarios. However, analysis of failures reveals incorrect risk decisions despite accurate intermediate calculations, inconsistent propagation of financial states, and remediation actions that violate binding constraints. In contrast, some strict evaluation failures result from incomplete reporting rather than incorrect financial conclusions. These findings indicate that aggregate accuracy alone is insufficient to characterize financial-risk reasoning reliability and motivate evaluations that distinguish numerical correctness, state consistency, constraint satisfaction, and decision validity.

q-fin.RM↗

arXiv2026

Certifying Reward Alignment in Portfolio Reinforcement Learning

Financial reinforcement learning often translates an investment objective into an additive risk-adjusted reward. An agent can optimise that reward accurately yet implement a different financial preference. We develop a decision audit for terminal mean-variance equilibrium that separates reward misspecification from learning error. In a finite exogenous-state market, we bound the gain from changing one action using actor, reward and continuation-critic residuals. The certificate ignores state-constant critic offsets, accommodates convex action constraints, and has an independent-sample extension that does not require the equilibrium policy. A continuation-aware actor-critic method estimates the future-policy correction from samples. Across nine controlled markets and twenty seeds, it reduces policy distance by and one-step deviation gain by relative to a matched myopic-reward network. A separate PPO comparison tests the same reward intervention within a standard optimiser. Historical tests cross seven covariance estimators and 10, 30 and 49 US industry portfolios over 432 months: joint reward calibration preserves each fitted mean-variance target, but does not improve realised investment performance. Together, these results provide a way to quantify whether a learned decision implements its stated financial objective, separately from how well that objective performs in a backtest.

q-fin.PM↗