I test whether filing text ranks companies by future 20-day realized volatility out of sample.
Auditable financial NLP research · v4 evidence
Financial 10-K Text Agent
10-K text ranks future volatility — but does not prove tradable alpha.
I designed and built an end-to-end research pipeline to test whether SEC 10-K disclosures contain out-of-sample information about future 20-day realized volatility.
500SEC 10-K filings
0.2395Preregistered primary Rank IC
0Critical audit failures
2Disclosed scope warnings
8,133 OOS predictions. Best exploratory Rank IC: 0.3668, reported separately from the preregistered primary result.
For admissions reviewers
The Project in 60 Seconds
Five answers connect the finance question, my contribution, the evidence, and its limits.
Inspect my contribution ↓SEC ingestion, section review, forward labels, rolling splits, train-only text features, models, diagnostics, and audit reports.
The fixed 50-company panel shows positive OOS information for volatility ranking.
The preregistered portfolio Sharpe is -0.8539; portfolio evidence is diagnostic only.
The universe is a fixed active-company panel and market data use a mixed public-source stack. These limits are disclosed rather than hidden.
How to read the evidence
Primary Claims Come Before Model Exploration
The specification registry separates the two preregistered tests from robustness checks and exploratory comparisons.
Preregistered
1 prediction + 1 portfolio testPrimary prediction: Rank IC 0.2395. Primary portfolio: Sharpe -0.8539.
Diagnostics & exploration
594 specifications across 26 familiesAblations, baselines, neutralization, return targets, and model comparisons are reported separately.
Audit boundary
0 critical failures · 2 scope warningsThe warnings disclose public-data and universe limitations; they do not convert exploratory evidence into a formal claim.
Current release: 50_company_public_fmp_alpha_2016_2025_v4 · Read the audit boundary
Result snapshot
Volatility Ranking Evidence
The strongest observed result is exploratory; the preregistered Ridge result remains positive.
TF-IDF/SVD onlyExploratory best
0.3668Industry + textIncremental diagnostic
0.3296Industry meanEconomic baseline
0.2924Combined text RidgePreregistered primary
0.2395Raw p-value 0.00067. Positive exploratory volatility-ranking evidence.
Raw p-value 0.1147. Tradable alpha is not established.
My contribution
What I Designed and Built
I developed the workflow from SEC filing ingestion to audited out-of-sample evidence.
Rolling train/validation/test splits, forward labels, preregistered specifications, and embargo-based leakage controls.
Section parsing, Loughran-McDonald tone features, train-window-only TF-IDF/SVD, and model manifests.
Rank IC, feature ablation, industry-neutral diagnostics, clustered bootstrap, coverage checks, and automated reports.
Python · scikit-learn · XGBoost · SEC EDGAR · pytest · Ruff · GitHub Actions
Why it matters
Beyond Sentiment Scores and Document Search
Forward labels
Links filing text to future volatility and abnormal-return targets.
Rolling OOS design
Separates training, validation, and test windows through time.
Train-only features
Fits TF-IDF/SVD vocabularies only inside each training window.
Evidence hierarchy
Separates preregistered claims from exploratory model comparisons.
Research audit
Discloses parser issues, bootstrap uncertainty, coverage, and multiple testing.
Pipeline
From SEC Filing to Audited Evidence
SEC Filingsacceptance timestamps
Parser Reviewsection quality flags
Labelsfuture targets
Rolling Splitsembargo + OOS
Text Featurestrain-window only
Diagnosticsablation + bootstrap
Auditclaims + boundaries
Inspect the evidence
Follow the Public Audit Trail
Compact, license-safe artifacts connect every headline number to a reproducible evidence file.
Usage Boundary
This is an applied-grade exploratory run, not CRSP/WRDS-equivalent formal asset-pricing evidence, a survivorship-free replication, a production trading system, proven tradable alpha, or investment advice.