Research map / Label-efficient and robust learning / Recognition and long-tail
Linking Industry Sectors and Financial Statements: A Hybrid Approach for Company Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Classifies U.S. companies into 8 industry sectors from their quarterly financial statements (SEC EDGAR, FY2023), comparing LightGBM, LLM classification-head and generation-head models, and a proposed Text-Numeric transformer that fuses tag embeddings with log-scaled amounts.
Problem
Predicting a company's industry sector (SIC-based) from its financial statement records alone, to support semi-automatic accounting audit, where prior work relied on numerical features only and sometimes leaked company identity across splits.
Approach
Fuse sentence-encoder embeddings of account identifiers with signed log-scaled amounts, then process the sequence with a small transformer and a CLS-based classifier; also compare LLM classification-head and generation-head fine-tuning and generate explanations from the generative model.
Key results
- Text-Numeric transformer with FinLang embeddings reaches test MCC 0.71 and F1 0.76, the best of all variants.
- LightGBM with relative values reaches test MCC 0.69; LLM-clf and LLM-gen with LLaMA3.1/FinLLaMA3 reach test MCC about 0.64-0.66.
- Zero-shot LLaMA3.1 8B reaches only MCC 0.30 on test, and dropping amounts drops LLM-gen test MCC to 0.58.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use relative values instead of raw values (ablation) | LightGBM · Company classification, test set | MCC: 0.66 -> 0.69 | Helped |
| Use binary account presence instead of amounts (ablation) | LightGBM · Company classification, test set | MCC: 0.69 -> 0.56 | Hurt |
| Add account amounts to the relative-value input (ablation) | LLM-clf, LLaMA3.1 · Company classification, test set | MCC: 0.06 -> 0.64 | Helped |
| Use raw values instead of relative values (ablation) | LLM-clf, LLaMA3.1 · Company classification, test set | MCC: 0.64 -> 0.58 | Hurt |
| Add account amounts to the relative-value input (ablation) | LLM-gen, LLaMA3.1 · Company classification, test set | MCC: 0.58 -> 0.64 | Helped |
| Use raw values instead of relative values (ablation) | LLM-gen, LLaMA3.1 · Company classification, test set | MCC: 0.64 -> 0.65 | Helped |
| Omit task-specific fine-tuning (ablation) | LLM-gen, LLaMA3.1 · Company classification, test set | MCC: 0.30 | Hurt |
| Replace FinLang with BGE Base as the sentence encoder (ablation) | Text-Numeric Transformer · Company classification, test set | MCC: 0.68 -> 0.71 | Hurt |
| Add account amounts to the text embeddings (ablation) | Text-Numeric Transformer (FinLang) · Company classification, test set | MCC: 0.60 -> 0.71 | Helped |
| Use breadth-first instead of depth-first traversal to order accounts (ablation) | LLM-gen and LLM-clf · Company classification | MCC | No clear effect |
Surprising findings
- Zero-shot LLMs perform weakly and collapse onto two sectors (Manufacturing and Services).
- Ordering of tags (depth-first vs breadth-first) has minimal impact on LLM performance.
Limitations
- Class imbalance; oversampling minority classes reduced overall performance.
- LLM-gen explanations were only qualitatively evaluated on 10 samples, lacking a grounded reference.
- Larger LLMs (e.g., LLaMA 405B, Voyage-Finance embeddings) were not tested due to GPU limits.
Experimental setup
Datasets: SEC EDGAR quarterly financial statements (FY2023, US-GAAP tags, SIC labels)
Benchmarks: SEC EDGAR FY2023 company classification subset
Metrics: MCC weighted F1 MRR
Baselines: LightGBM LLM-clf LLM-gen Van Der Heijden (2022) random forest Vamvourellis et al. (2023) SBERT
Builds on: LightGBM BGE LLaMA 3.1 Finance-LLaMA3 Sentence-BERT
Closest papers by meaning
- The SEC Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data (NeurIPS 2026 · similarity 0.81)
- Explainable Earnings Call Representation Learning (Student Abstract) (AAAI 2024 · similarity 0.80)
- FinBen: A Holistic Financial Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.80)
- Breaking the Gradient Barrier: Unveiling Large Language Models for Strategic Classification (NeurIPS 2025 · similarity 0.78)
- Feature Responsiveness Scores: Model-Agnostic Explanations for Recourse (ICLR 2025 · similarity 0.78)
- Cost Efficient Fairness Audit Under Partial Feedback (NeurIPS 2026 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.