VERAFI: Verified Agentic Financial Intelligence through Neurosymbolic Policy Generation

Battle Count: 62

While VERAFI is primarily focused on financial question-answering and regulatory compliance rather than direct trading strategy development, it has indirect relevance to quantitative trading through: (1) accurate extraction and computation of financial metrics (ROA, debt-to-equity, free cash flow) that feed into fundamental analysis and factor models, (2) ensuring mathematical accuracy in financial calculations critical for backtesting and risk assessment, (3) regulatory compliance verification for trading strategies, and (4) reliable analysis of SEC filings for event-driven strategies. However, it does not address market prediction, signal generation, or execution optimization directly.

The system requires integration of multiple components: dense embedding models (Qwen3-Embedding-4B), cross-encoder reranking (Jina-reranker-v3), agentic framework (Strands with Claude Sonnet 4), computational tools (Python REPL, calculator, web search), vector database (Chroma), neurosymbolic autoformalization pipeline (Amazon Bedrock Automated Reasoning), and SMT-lib policy generation. The multi-stage pipeline with policy loading, retrieval, agentic reasoning, and answer extraction adds significant orchestration complexity. However, the use of existing frameworks and APIs reduces some implementation burden.

The paper provides detailed system architecture, algorithm pseudocode, evaluation dataset construction methodology, and prompt templates in the appendix. However, no code repository is mentioned. The system relies on proprietary models (Claude Sonnet 4, Amazon Bedrock) and specific commercial tools (Tavily, Strands framework), limiting full reproducibility. The evaluation dataset is constructed from publicly available FinanceBench and ConvFinQA data for four specific companies.