50-company public result package

Results and Artifacts

A compact public summary of the latest applied-grade SEC 10-K text factor run: `50_company_public_fmp_alpha_2016_2025_v1`.

Run summary

50-Company Applied Pilot

Run ID	`50_company_public_fmp_alpha_2016_2025_v1`
Universe	50 U.S. large-cap firms
Sample	FY2016-FY2025
SEC 10-K filings	500
Labels	1,500
OOS predictions	4,716
Feature records	520,465
Tested specifications	568

Preregistered primary result

Main Result

Ridge on `realized_volatility_1_20`, evaluated by ALL_SPLITS Rank IC.

Rank IC 0.2606

Raw p-value 0.00017

Exploratory comparison

Best Observed Prediction

XGBoost on `realized_volatility_1_20` reported as model-comparison evidence, not as the preregistered primary claim.

Rank IC	0.3133
Newey-West t-stat	6.8479
RMSE	0.00834

Model comparison

Rank IC by Model

A compact view of ALL_SPLITS test Rank IC on realized_volatility_1_20. Ridge is the preregistered primary model; XGBoost is exploratory.

XGBoost exploratory text model

0.3133

Industry mean benchmark

0.2952

Ridge preregistered primary

0.2606

Historical mean benchmark

-0.0206

Bars are scaled to the largest displayed Rank IC. Negative values indicate the score ranks future volatility in the opposite direction.

Audit

Coverage and Controls

100%

Eligible OOS coverage

Audit failures

Audit warnings

568

Tested specifications

The audit trail separates raw label coverage from eligible OOS prediction coverage, discloses multiple testing, and records the applied-grade data boundary.

Raw labels 48.7%

Eligible OOS 100%

Primary specs 2 / 2

Artifacts

Representative Evidence Files

Result PackageLatest public package Factor CardFast result summary Empirical ReportPaper-style narrative Audit ReportFormal gate and warnings Coverage WaterfallRaw vs eligible OOS coverage Multiple TestingBonferroni, Holm, BH-FDR

Data Boundary

SEC EDGAR provides official 10-K filings and filing timestamps. Market data uses a mixed FMP/Yahoo public-source stack, and market-cap-at-selection values are applied-grade estimates.

This is not a CRSP/WRDS-equivalent survivorship-free replication.

Interpretation Policy

Use this package as evidence of an auditable financial NLP research workflow and exploratory volatility-prediction evidence.

Portfolio outputs are diagnostic only; this package does not establish formal tradable alpha or provide investment advice.

Reproducibility

Run the Public Code Locally

The public repository can be cloned, installed, linted, and tested without private data. Full real-data runs require API keys and local private data directories.

git clone https://github.com/uiclxh/financial-10k-text-agent.git
cd financial-10k-text-agent
python -m pip install -e ".[dev]"
python -m ruff check .
python -m pytest

Open-source code is released under the MIT License.