NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 26·5 min read
From Accuracy to Auditability: A Survey of Determinism in Financial AI Systems
TL;DR
Financial AI systems face challenges in reproducibility due to mechanical nondeterminism from deep learning architectures. This survey analyzes three dominant modalities: tabular models, graph networks, and LLM-based workflows, assessing their auditability.
✦ Why It Matters
Engineers can enhance the auditability of financial AI systems by applying the proposed evaluation framework and metrics.
Key Takeaways
- Implement the proposed evaluation framework for financial AI systems.
- Focus on improving reproducibility in ML models used in finance.
- Measure explanation variance and prediction stability in deployed models.
- Explore the impact of hardware on model determinism and auditability.
Full Summary
In regulated financial environments, deploying machine learning (ML) raises concerns about reproducibility, particularly in areas like credit risk and fraud detection. The survey examines three key modalities in financial AI: tabular models, which struggle with explanation variance; graph networks, affected by stochastic sampling; and large language model (LLM)-based workflows, which experience batch-dependent divergence.
The authors conducted experiments on public financial datasets, measuring explanation rank instability in credit scoring and prediction flip rates in fraud detection. They also identified output divergence in LLM entity extraction due to tensor-parallel processing.
A layered evaluation framework was proposed, linking specific metrics to audit readiness, and the complementarity of logit-level and semantic-level determinism measures was empirically validated. These findings highlight the need for improved auditability in financial AI systems.