Explainable AI for Comprehensive Risk Assessment for Financial Reports: A Lightweight Hierarchical Transformer Network Approach

Better risk measurement beyond standard metrics: TinyXRA moves beyond the industry standard of using only standard deviation of returns to assess company risk. Instead, it incorporates skewness, kurtosis, and the Sortino ratio, providing a more nuanced understanding of financial risk that distinguishes between upside and downside volatility—a distinction that matters for investors.

Explainability makes AI decisions trustworthy: The model includes a "dynamic, attention-based word cloud mechanism" that visualizes which specific terms and passages in a company's 10-K report most influenced its risk assessment. This transparency allows users to understand why the system rated a company as risky, rather than treating it as a "black box."

Practical efficiency enables real-world deployment: By using TinyBERT as its foundation, TinyXRA is designed to be computationally lightweight, allowing it to process thousands of financial documents in real-time even on systems with limited computing resources. This matters for financial institutions that need scalable solutions for monitoring many companies simultaneously.

Improved classification methodology: The model uses "triplet loss" for categorizing companies into risk quartiles, which the authors note captures both the direction and magnitude of risk differences more effectively than previous approaches, leading to state-of-the-art predictive accuracy across seven years of test data (2013-2024).