README.md

---
license: cc-by-4.0
language:
- en
pretty_name: OpenFundex
tags:
- finance
- value-investing
- sec-filings
- fundamental-analysis
- tabular
task_categories:
- tabular-classification
- tabular-regression
size_categories:
- 100K<n<1M
configs:
  - config_name: default
    default: true
    data_files:
      - { split: train,      path: train_clean.parquet }
      - { split: validation, path: validation_clean.parquet }
      - { split: test,       path: test_clean.parquet }
  - config_name: full
    data_files:
      - { split: train,      path: train.parquet }
      - { split: validation, path: validation.parquet }
      - { split: test,       path: test.parquet }
  - config_name: live
    data_files:
      - { split: clean, path: recent_clean.parquet }
      - { split: full,  path: recent.parquet }

---

# OpenFundex Dataset

A structured dataset of SEC financial filings for deep value analysis and financial distress prediction.

## Dataset Description

- **License**: CC-BY-4.0
- **Language**: English

### Summary

OpenFundex contains financial statement data extracted from SEC EDGAR filings,
enriched with derived financial metrics and labeled with established quality scores
(Piotroski F-Score, Altman Z'-Score, Graham metrics). Designed for training ML
models to assess company financial health and identify deep value opportunities.

**Key design decision:** This dataset uses only fundamental data from SEC filings.
No market prices or equity trading data are included, eliminating survivorship bias.

## Supported Tasks

- **Tabular Classification**: Predict financial distress, bankruptcy, value creation, fundamental improvement
- **Tabular Regression**: Predict quality scores (f_score, z_prime_score, composite_quality_score), growth rates
- **Anomaly Detection**: Identify companies in financial distress or with QA anomalies

## Dataset Structure

### Splits

| Split | Records | Companies | Date Range |
|-------|---------|-----------|------------|
| train | 221,779 | 11,786 | 2008-12-31 to 2019-12-31 |
| validation | 45,109 | 7,207 | 2020-01-31 to 2021-12-31 |
| test | 47,334 | 7,208 | 2022-01-31 to 2023-12-31 |
| recent | 43,288 | 6,507 | 2024-01-31 to 2026-02-28 |

**Total records:** 357,510

### Feature Groups

| Group | Count |
|-------|-------|
| Identifiers | 5 |
| Context | 4 |
| Raw Features (SEC XBRL) | 33 |
| Derived Features | 8 |
| Engineered Features | 23 |
| QA Flags | 4 |
| Prediction Targets | 21 |
| Rank Targets | 14 |

## Scoring Models

- **Piotroski F-Score** (0-9): Nine binary signals measuring profitability, leverage, and operating efficiency. Null when no prior quarter available for delta signals.
- **Altman Z'-Score** (Float): Private-firm bankruptcy risk variant with zone classification (safe/grey/distress). Null for financial firms (SIC 6000-6999).
- **Beneish Coverage** (0-8): Count of computable M-Score components. Full M-Score is computed transiently during enrichment but not retained.
- **Graham Metrics**: Graham Number, NCAV/share, tangible book value/share, net working capital/share, defensive score (0-5).
- **Quality Signals**: Cash conversion ratio, accrual ratio, free cash flow margin.
- **Composite Quality Score**: Z-score normalized average of key quality signals within each quarter cross-section.

## Target Columns

21 forward-looking prediction targets using same-quarter year-over-year comparisons:

### 1-Year Targets
- **Growth rates** (6): BVPS, equity, earnings, revenue, OCF, FCF growth
- **Level/delta** (2): Forward ROE, margin expansion
- **Binary** (5): ROA improved, fundamentals improved (≥3 of 5 metrics), value created (equity grew AND ROE>0), survived (filed Q+4 AND no bankruptcy in window; NULL when CIK is in dropout-cause cohort with cause ∈ {form_25_delisting, unexplained_dark} and the PIT gate is met — see Considerations), filed for bankruptcy (petition within 365 days)

### 2-Year Targets
- **Growth rates** (6): Same metrics as 1-year, over 2-year horizon
- **Binary** (2): Survived (filed Q+8 AND no bankruptcy in window; NULL when CIK is in dropout-cause cohort with cause ∈ {form_25_delisting, unexplained_dark} and the PIT gate is met — see Considerations), filed for bankruptcy (petition within 730 days)

Survived and bankruptcy targets are mutually exclusive by construction. All targets are null when forward quarter data is unavailable.
A third NULL state (CP-1 / v0.7.0) excludes Form-25 (M&A proxy) and unexplained-dark dropouts from the negative class to avoid mislabeling acquired companies as failed survivors.

### Rank-Transformed Targets (14 columns)

Cross-sectional percentile ranks (0-1] for all Float64 targets, computed per quarter
using `rank("average") / count()`. Raw growth targets are extremely skewed
(mean ~3.1, median ~0.03) and produce negative information coefficients for
regression models. Rank-transforming yields IC ~0.37.

## Dataset Creation

### Source Data

All data sourced from SEC EDGAR: Financial Statement Data Sets (FSDS) for fundamentals
and Full-Text Search API (EFTS) for 8-K Item 1.03 bankruptcy filings.
No market data providers. No third-party data. No equity pricing data.

### Pipeline

1. **Ingest**: Download quarterly SEC FSDS ZIP files and bankruptcy events from EDGAR
2. **Parse**: Extract XBRL financial data, normalize 32 tags to standard fields
3. **Enrich**: Compute derived ratios (8), scoring models (5), bankruptcy flags, and QA flags
4. **Label**: Generate 21 forward-looking prediction targets (including bankruptcy)
5. **Split**: Temporal train/validation/test/recent splits with leakage validation
6. **Evaluate**: Quality checks, ML fitness, and publication readiness
7. **Publish**: Stage and upload to Hugging Face Hub

## Considerations

### Known Limitations

- XBRL coverage varies: some companies report fewer standardized tags
- F-Score delta components require prior quarter data (null for first appearance)
- Z'-Score was designed for manufacturing firms; interpretation varies by sector
- No market data: cannot compute price-based metrics (P/E, market cap, etc.)
- `bankruptcy_chapter` is populated for bankrupt CIKs (values like `"7"`, `"11"`,
  `"15"`, `"9"` extracted via 8-K Item 1.03 filing bodies); non-bankrupt rows
  are null. Edge cases not yet handled: chapter spelled out as a word
  ("chapter eleven") and USC-only citations with no "chapter" keyword both
  yield null.

### Bias Considerations

- **Survivorship bias** (Audit C1 / C1.1, 2026-05-05/06): the panel is built from
  SEC FSDS XBRL filings, so companies that stopped filing leave the panel
  silently. An audit of the 2009Q2–2026Q1 window classified 8,847 stopped-filing
  CIKs by cause: 54.4% Form-15 deregistration, 33.3% unexplained-dark,
  8.7% bankruptcy, 3.6% Form-25 delisting (M&A proxy). LoPucki BRD intersect
  analysis (Audit C1.1) shows 21.4% of in-window Chapter 7/11 bankruptcies
  (83/388) are entirely absent from the panel, concentrated 2009–2011 during
  the XBRL ramp-up — a structural lower bound on missed bankruptcies.
- **Survival target NULL semantics** (CP-1 / v0.7.0): `target_survived_1y` and
  `target_survived_2y` are set to NULL on rows where the CIK appears in the
  bundled dropout-cause cohort with cause ∈ {`form_25_delisting`,
  `unexplained_dark`} and the PIT gate is met (Form-25 `evidence_date` falls
  inside the label horizon, or `unexplained_dark` always-NULL). This avoids
  mislabeling acquired companies (Form-25 = M&A proxy) and unexplained
  dropouts as failed survivors. Bankruptcy and Form-15 CIKs remain in the
  negative class. The cohort artifact is shipped with the package at
  `openfundex/cohorts/_data/dropout_causes.parquet` and regenerable via
  `ofx cohorts refresh dropout-causes`.
- **Temporal integrity**: Strict time-based splits prevent data leakage
- **Sector bias**: Z'-Score thresholds may not be equally applicable across all sectors
- **Financial firms excluded from Z'-Score**: Financial companies (SIC 6000-6999) have null Z'-Score values

## License

This dataset is released under the [CC-BY-4.0 license](https://creativecommons.org/licenses/by/4.0/).

The underlying SEC data is in the public domain.

## Citation

```bibtex
@dataset{openfundex,
  title={OpenFundex: SEC Financial Filings for Deep Value Analysis},
  author={Danielson, Luke},
  year={2026},
  url={https://github.com/danielukea/openfundex},
  license={CC-BY-4.0}
}
```

## Changelog

### v0.7.0 — 2026-05-14
- **`target_survived_{1y,2y}` NULL semantics widened (CP-1):** rows where the CIK
  appears in the bundled dropout-cause cohort with cause ∈ {`form_25_delisting`,
  `unexplained_dark`} and the PIT gate is met are now NULL instead of False.
  This excludes M&A proxies and unknown-cause dropouts from the survival
  negative class to avoid mislabeling acquired companies. Bankruptcy and
  Form-15 CIKs remain negatives. Schema unchanged; this is a value-contract
  change. See Considerations § Survivorship bias.
- New bundled package data: `src/openfundex/cohorts/_data/dropout_causes.parquet`
  (8,847 CIKs classified by precedence). Regenerable via `ofx cohorts refresh
  dropout-causes`.

### v0.6.0 — 2026-04-17
- **`bankruptcy_chapter` populated:** Chapter number (`"7"`, `"11"`, `"15"`, `"9"`)
  is now extracted from the 8-K Item 1.03 filing body for each bankrupt CIK and
  stored in the `bankruptcy_chapter` column. Non-bankrupt rows remain null.
  Extraction uses a guardword-gated Item 1.03 window search; known limitations
  (word-spelled chapters, USC-only citations) are documented above.

### v0.5.0 — 2026-04-17
- **Loader contract:** dataset card now declares explicit `configs` and `data_files`. `load_dataset("ttchopper/openfundex")` returns the QA-filtered historical splits (`train`/`validation`/`test`); unfiltered historical data is available via `load_dataset(..., "full")`; live 2024+ inference data via `load_dataset(..., "live")`.
- Previously the card omitted `configs:`, causing the Hub to glob all parquets and misassemble splits (paired `_clean` files were concatenated or surfaced as extra splits, and `recent*` was not declared at all).
- Raw parquet access (`hf://datasets/ttchopper/openfundex/train.parquet`) is unaffected.