AI
New Dataset Tackles Automated Extraction of Financial Metrics from Earnings Reports
Researchers release HiFi-KPI, a large-scale corpus designed to help machine-learning models automatically identify and classify key financial metrics buried in corporate filings.
Extracting actionable financial data from earnings reports at scale remains a significant challenge for investors and analysts. While public companies must submit filings in machine-readable format using inline eXtensible Business Reporting Language (iXBRL), the underlying taxonomies are complex and vary widely between firms, making it difficult for automated systems to generalize across different organizations.
The Dataset and Its Scale#
Researchers have introduced the Hierarchical Financial Key Performance Indicator (HiFi-KPI) dataset to help address this gap. The corpus comprises 1.65 million paragraphs drawn from earnings filings, annotated with 198,000 unique labels that follow a hierarchical structure. These labels are cross-referenced to iXBRL taxonomy standards, creating a bridge between human-interpretable categories and standardized financial reporting formats. The team also released a smaller, hand-curated version—HiFi-KPI-Lite—containing 8,000 paragraphs for faster evaluation and experimentation.
Model Performance and Capabilities#
Testing on HiFi-KPI-Lite revealed distinct strengths across different approaches. Encoder-based neural networks achieved macro-F1 scores exceeding 0.906 on classification tasks, indicating strong performance when asked simply to label which category a paragraph belongs to. Large Language Models (LLMs), meanwhile, reached F1 scores of 0.440 on more complex structured extraction tasks—where the goal is not just to identify metrics but to pull specific values and organize them into defined fields.
Known Limitations and Future Work#
Qualitative analysis of model errors highlighted a consistent weakness: accurately capturing dates and temporal references within the extracted data. This suggests that while current systems excel at recognizing financial metrics themselves, grounding those metrics to the correct time periods remains an open problem. The research team has open-sourced both the code and dataset to enable broader experimentation.
