AWARE-FX: An Auditable Knowledge-Guided AI System for Measuring Corporate Foreign-Exchange Hedging Disclosure

Battle Count: 62

The paper primarily addresses corporate disclosure measurement rather than trading signals. However, the strict FX hedging-disclosure score could serve as a screening variable for FX-risk-aware portfolio construction, fundamental factor models, or event-driven strategies monitoring corporate risk-management changes. The external construct validation (negative association with FX exposure) provides indirect relevance for risk-adjusted position sizing. Not directly applicable to short-term price prediction or algorithmic execution.

Multi-module pipeline requiring: (1) PDF parsing and page-level text extraction across 4.1M+ pages, (2) 80-rule professional lexicon with channel/signal/polarity metadata, (3) 15 negation and accounting-status rules, (4) four separately fine-tuned FinBERT classifiers with channel-specific epoch selection, (5) exact channel gates and top-three mean aggregation logic, (6) versioned evidence ledger with full provenance, (7) calibration and selective-prediction infrastructure, (8) multi-seed and temporal evaluation harness, (9) fixed-prompt LLM benchmarking pipeline. Requires GPU training infrastructure and careful version management.

Versioned lexicon (80 rules), fixed prompts, recorded model revisions, deterministic decoding for Qwen3-8B, archived run parameters (learning rate, batch size, seeds), incremental checkpoints, and hash-verified inputs. Code and derived tables promised upon publication. However, hosted Gemini comparator lacks stable API model ID and decoding parameters, reducing full reproducibility for that component.