Preregistered primary prediction
Ridge · realized_volatility_1_20
Rank IC0.2395
Raw p-value 0.00067; supports exploratory volatility-ranking evidence.
Methods / Audit · v4
Methods and Audit
Every claim must be linked to a timestamped, testable artifact.
Filings, labels, splits, features, models, predictions, inference, portfolio diagnostics, and audit checks form one evidence chain.
50_company_public_fmp_alpha_2016_2025_v4
Leakage-safe research design
Only Past Information Enters the Forecast
Every text feature and model decision must be available before the prediction timestamp.
1Filing timestampSEC acceptance time defines information availability.
2Section parsingBusiness, Risk Factors, Legal Proceedings, and MD&A are extracted.
3Forward labelsVolatility and CAR targets begin only after filing availability.
4Rolling splitsTrain, validation, and test windows move through time with embargo purges.
5Train-window featuresTF-IDF/SVD vocabularies are fit inside training windows only.
6Validation-only tuningHyperparameters are selected using validation Rank IC only.
7Tie-aware test Rank ICTest ranking uses average ranks; constant predictions return zero Rank IC.
8Evidence hierarchyPrimary, robustness, and exploratory specifications remain separate.
9Audit boundaryCoverage, parser quality, data limitations, and claim strength are disclosed.
Key controls
Research Controls at a Glance
| Control | Implementation |
|---|---|
| Event-time alignment | Filing timestamps define information availability. |
| Rolling OOS design | Train / validation / test windows roll through time. |
| Leakage control | Embargo purge and split-leakage logs. |
| Model selection | Validation-only Rank IC. |
| TF-IDF control | Train-window-only vocabulary fitting. |
| Incremental text diagnostic | Industry-neutral Rank IC and feature ablation. |
| Statistical uncertainty | Newey-West and clustered bootstrap confidence intervals. |
| Data snooping | Specification registry and multiple-testing report. |
| Parser quality | Manual review appendix for short or malformed sections. |
Feature construction
Financial Text Representations
Loughran-McDonald dictionary tone and TF-IDF/SVD are built over full filing, Business, Risk Factors, Legal Proceedings, and MD&A scopes.
| Feature set | Meaning |
|---|---|
industry_only | Training-window industry-mean baseline. |
dictionary_only | Dictionary-tone text features. |
tfidf_svd_only | TF-IDF/SVD text representations. |
combined_text | Combined dictionary and text representation. |
industry_plus_text | Industry features plus text features. |
Industry-neutral diagnosticDoes text retain information after removing split-industry means?
Industry-neutral Rank IC is a descriptive diagnostic, not a causal decomposition.
Bootstrap inference
How Uncertainty Is Reported
Split bootstrap
Inconclusive for v4 because there are only four OOS split clusters.
Event-date bootstrap
Supports a positive raw primary Rank IC interval.
Ticker-cluster bootstrap
Supports a positive raw primary Rank IC interval.
Industry-neutral bootstrap
Positive point estimate, but not bootstrap-robust.
Parser quality review
Section Extraction Is Audited, Not Assumed
2,000
Parsed section records
144
Manual-review records
144
Excluded section-level records
494
Short but included records
Item 1A and Item 7 below 100 words are excluded from section-level features. Core sections from 100 to 499 words remain included but carry a warning.
Evidence boundary
Prediction and Trading Evidence Are Judged Separately
Ridge ·
Preregistered primary portfolio
Monthly sector-neutral equal-weight
Sharpe-0.8539
Raw p-value 0.1147; does not establish tradable alpha.
Formal Result Boundary
Formal empirical-finance claims are blocked by data-boundary issues, not pipeline failures: mixed FMP/Yahoo data, applied-grade market-cap estimates, fixed 50-company panel, parser-quality limitations, and a small number of missing diagnostic model-label pairs.
The project should be interpreted as an applied-grade, auditable financial NLP workflow for exploratory volatility ranking.