AWARE-FX is an auditable AI/NLP decision-support system that converts corporate annual-report text into traceable firm-year FX hedging-disclosure measures. It combines a professional-source lexicon, negation and accounting-status logic, channel-specific financial encoders, exact evidence gates, conservative aggregation, and an audit ledger. Across 24,909 Hong Kong firm-years (2008–2025) it retrieves and scores 543,527 snippets, with a rigorous evaluation including ablations, a stratified human audit, multi-seed encoder comparisons, and temporal testing. FinBERT generally achieves higher mean F1 than Contemporary encoders across seven of eight encoder-task combinations; temporal F1 ranges from 0.702 to 0.872. Abstaining on the least-confident 20% of temporal observations raises retained-sample F1 by 0.050–0.077. Deterministic Qwen3-8B shows strength on commodity and negation evidence but weaker performance on foreign-debt and accounting-context labels, highlighting limits of general-purpose LLMs for specialized financial tasks. The strict FX score is negatively associated with linked FX exposure measures, while the generic broad score is not, providing external construct validation rather than causal hedging effects. The study contributes a tested, auditable decision-support pipeline where retrieval, status logic, classification, uncertainty handling, aggregation, and external validation remain independently auditable
以下对 AWARE-FX:An Auditable Knowledge-Guided AI System for Measuring Corporate Foreign-Exchange Hedging Disclosure 进行极为详尽的解读与评估,覆盖元数据、结构性解读、图表与表格逐条解读、方法论与估值性证据、风险分析、批判性视角,以及结论性综合。为溯源性强的分析,涉及每段论断我均在句末给出原文出处页码标识([page::页码]),必要时跨页引用则给出多页标识。
Figure 1: AWARE-FX measurement and validation architecture
Figure 2: A. Forward temporal generalization 与 B. Paired temporal differences
Figure 3: Selective prediction on the temporal test
Figure 4: External construct-validation coefficients in the main sample
Table 1: Positioning against recent financial language-model research
Table 2: AWARE-FX system modules and audit artifacts
Table 3: Domain-lexicon extraction coverage
Table 4: Selected held-out FinBERT checkpoints after platform-epoch assessment
Table 5: Manual audit of domain-lexicon v2 snippets
Table 6: Baseline and ablation comparison on the 300-snippet human audit
Table 7: Flow from annual-report firm-years to exposure-validation records
Table 8: External construct validation using FX exposure outcomes
Table 9: Multi-seed encoder performance under grouped-random and temporal evaluation
Table 10: Paired encoder comparison on the fixed 2023–2025 temporal test set
Table 11: Temporal generalization and selective prediction for FinBERT
Table 12: Fixed-prompt LLM performance on the stratified 300-snippet audit set
Table 13: Observed evidence-classification failure modes and system responses
Figure 2、Figure 3、Figure 4 的图像对象如上所述,均在相应页码处提供了可视证据与趋势描述。
其他图表:Figure 1(页码 9)作为架构示意图,为理解论文的模块化设计提供直观参照;Table 1/Table 2/Table 3/Table 4/Table 5/Table 6/Table 7/Table 8/Table 9/Table 10/Table 11/Table 12/Table 13 依序在文本对应段落中给出,均有上述页码标注可追溯到原始段落。对应图片引用如上列出。若需将图像嵌入,按上述 markdown 图像路径嵌入即可。
总体而言,AWARE-FX 提供了一个系统性、可审计且具外部构造效度检验能力的框架,用于将非结构化年报文本转化为可追溯的 FX 对冲披露证据集合,并通过多层评估(包括时序鲁棒性、对照实验、人工审计、外部暴露验证)来验证其在信息筛选任务上的有效性与局限性。研究的贡献不仅在于一个具体模型的性能提升,更在于将规则、编码、聚合与审计逐步分离,形成一个可更新、可追溯、可审查的金融文本分析“工艺流程”,为未来在金融文本分析领域的设计科学研究提供了一个具有可操作性的模板与参照框架。
附:本文引用/溯源的关键点清单(要点与页码)
如需,我可以将上述分析整理成更紧凑的摘要版,或将每个图表/表格的要点提炼成可直接用于研究报告的“要点卡片”,便于在后续研究中快速引用。
2607.27611.pdf