def14a-parser

Extract CEO compensation and governance data from SEC DEF 14A proxy statements using a local LLM via Ollama — no API keys, no cloud costs, runs on a laptop.

{
  "fiscal_year": 2023,
  "ceo": {
    "name": "Jane Doe",
    "title": "President & Chief Executive Officer",
    "is_chairman": false,
    "career_summary": "Ms. Doe joined Acme Bank in 2015 as CFO before becoming CEO in 2019."
  },
  "governance": {
    "board_size": 9,
    "pct_independent": 0.78,
    "ceo_chairman_combined": false,
    "say_on_pay_pct": 0.94,
    "auditor": "KPMG LLP"
  },
  "compensation": [
    {"name": "Jane Doe",  "title": "President & CEO", "is_ceo": true,  "year": 2023, "salary": 850000, "total_comp": 3200000},
    {"name": "John Smith","title": "Chief Financial Officer", "is_ceo": false, "year": 2023, "salary": 520000, "total_comp": 1100000}
  ]
}

Why another DEF 14A parser?

Existing tools take one of three approaches:

Approach	Example	Limitation
Rule-based HTML parsing	edgartools, ceopay	Breaks on inconsistent formatting across thousands of filers
Vision/ML table classifier	Execcomp-AI	Requires GPU, heavy dependency stack
Commercial API	sec-api	Paid, not reproducible

This library uses a Q&A prompt approach: instead of demanding that the LLM output a rigid JSON schema (which instruction-tuned models resist), we ask literal factual questions and parse the answers ourselves. This is significantly more reliable across the wide variation in proxy statement formatting found in the wild — especially for smaller regional filers who don't follow large-company templates.

Two calls per filing:

1. Comp table call — "list each NEO as a pipe-delimited row: Name | Title | CEO | Year | Salary | Total"
2. Governance Q&A — seven numbered questions answered one per line (board size, % independent, auditor, say-on-pay %, CEO/chair combined, fiscal year, CEO bio)

Requirements

• Python 3.10+
• Ollama running locally (ollama serve)
• At least one model pulled (see Model recommendations)

Installation

pip install def14a-parser

Or from source:

git clone https://github.com/ajwalsh08/def14a-parser
cd def14a-parser
pip install -e .

Quick start

Get a filing

DEF 14A filings are free from SEC EDGAR. Download the primary HTML document from any filing index page.

# Find a company's most recent proxy
# e.g. https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0000037808&type=DEF+14A

# Download the primary document (respect SEC rate limits: max 10 req/s)
curl -A "Your Name your@email.com" \
  "https://www.sec.gov/Archives/edgar/data/37808/.../primary.htm" \
  -o primary.htm

Python API

import def14a

result = def14a.extract("primary.htm")
# or with a specific model:
result = def14a.extract("primary.htm", model="qwen2.5:7b")

print(result["ceo"]["name"])
print(result["governance"]["auditor"])
for row in result["compensation"]:
    print(f"{row['name']}: ${row['total_comp']:,}")

CLI

# Single file — result to stdout
def14a run primary.htm

# Single file — write to JSON
def14a run primary.htm --output result.json

# Batch — each file gets a .json sidecar
def14a run filings/*/primary.htm --model qwen2.5:7b

Output schema

Field	Type	Description
fiscal_year	int	Year the compensation covers (from comp table, more reliable than Q&A)
ceo.name	str	CEO full name
ceo.title	str	CEO title as stated in the filing
ceo.is_chairman	bool	True if CEO is also Board Chair
ceo.career_summary	str | null	1-2 sentence bio from the proxy statement
governance.board_size	int | null	Total directors on board
governance.pct_independent	float | null	Fraction of independent directors (0.0–1.0)
governance.ceo_chairman_combined	bool	CEO and Chair are the same person
governance.say_on_pay_pct	float | null	Most recent advisory vote approval (0.0–1.0)
governance.auditor	str | null	External auditor firm name
compensation	list[dict]	All named executive officers (NEOs)
compensation[].name	str	NEO full name
compensation[].title	str	NEO title
compensation[].is_ceo	bool	True for the principal executive officer
compensation[].year	int	Fiscal year for this row
compensation[].salary	int | null	Base salary ($)
compensation[].total_comp	int | null	Total compensation ($)

Model recommendations

Pull models with ollama pull <name> before use.

Model	Size	Speed	Accuracy	Notes
llama3.2:3b	2 GB	~27s/filing	Good	Fastest option
qwen2.5:7b	4.7 GB	~50s/filing	Best	Recommended for production runs
gemma3:4b	3.3 GB	~35s/filing	Good	Strong instruction following

For a corpus of 1,000+ filings: run llama3.2:3b first (fast), then qwen2.5:7b --force on filings that produce bad CEO names (typically < 3%).

Expected results

These numbers come from a production run against 1,347 DEF 14A filings from 236 US regional bank holding companies (SIC 6020/6022), fiscal years 2017–2026, using qwen2.5:7b.

Extraction rate

Outcome	Count	%
Successfully extracted	1,206	89.5%
Irreducible failures	141	10.5%

The 141 failures fall into three categories that no parser can recover:

• Soliciting materials (14a-12 filings) — brief pre-proxy documents with no comp table
• Image-only tables — scanned PDFs where the comp table is a JPEG; BeautifulSoup returns empty text
• Word-per-line HTML — unusual layouts where table cells are individual <p> tags; column alignment is unrecoverable without the visual structure

Company coverage

Of 236 companies with DEF 14A filings, 229 (97%) have at least one successful extraction. The 7 with zero extractions are all image-heavy filers across every year.

Data quality of successful extractions

Check	Result
ceo.name populated	100%
Salary data present	98%
is_ceo=True row confirmed in comp table	91%
CEO career summary populated	~95%

The 9% without a confirmed is_ceo=True row are structurally unusual filings — ESOP-contribution tables (where all participants are listed as "Employee"), mutual thrift conversions, and holding companies where the comp table covers subsidiary officers only. The ceo.name field is still populated for these (using the first-listed executive as a fallback), but treat it as lower-confidence.

Known quality caveats

• Career summary accuracy (~62%) — The governance Q&A asks "describe the CEO career" without anchoring to the CEO's name. In ~38% of filings the model pulls a bio paragraph for the board chair or another prominent director instead of the CEO. This does not affect compensation data. If career summary accuracy matters for your use case, re-run the governance call with a name-anchored prompt: "What is [CEO name]'s career background?"

• Top-earner ≠ CEO (~10%) — In holding companies with separately-compensated subsidiary CEOs, the highest-paid executive may not be the parent company CEO. This is structurally correct, not an extraction error.

• total_comp null when salary > total — Some filings report multi-year data where a prior-year total looks smaller than the current-year salary (e.g., a year with no equity awards). The parser sets total_comp = None in these cases rather than returning an implausible value.

How section extraction works

DEF 14A filings contain several traps for a naive parser. The heading "Summary Compensation Table" appears in three places before the real section:

1. Table of contents — the heading followed immediately by a bare page number ("42")
2. Pay-vs-Performance column label — "Summary Compensation Table Total" as a column header in the PvP reconciliation (SEC rule added 2023)
3. Narrative cross-references — "as shown in the Summary Compensation Table above" mid-sentence

The section extractor works through three tiers:

Tier 1 — Exact heading match (~85% of filings) Find all short (<40 char) standalone occurrences of "Summary Compensation Table". Exclude PvP column labels, inline directional references ("above"/"below"), and sentence-final occurrences (trailing period with nothing after, which BeautifulSoup produces when a period falls at an HTML tag boundary). Skip TOC entries (lines immediately preceded or followed by a bare page number or "Page N"). Take the first remaining candidate whose extracted text contains both a fiscal year and a dollar amount — the data check distinguishes the real section from any remaining footnote cross-references.

Tier 2 — Broader "Executive Compensation" heading (~10% of filings) Some filers omit a standalone "Summary Compensation Table" heading entirely, or it lives inside a merged table cell that BeautifulSoup cannot extract. The broader pattern matches "Executive Compensation" section headers, which almost always precede the comp table. The first occurrence is tried before the last — the first is typically the actual section header, the last often appears in a say-on-pay advisory-vote narrative at the document end.

Tier 3 — Column-header cluster scan (last resort) A handful of filers have no recognizable section header at all but follow the standard SEC column layout. Scanning for adjacent "Salary" and "Total" lines within 5 rows locates the table directly.

Parser-only mode (no LLM)

If you only need the raw section text:

import def14a

lines        = def14a.html_to_lines("primary.htm")
comp_section = def14a.extract_compensation_section(lines)
dir_section  = def14a.extract_director_section(lines)

Known limitations

• Wrapper/index files — EDGAR sometimes stores a brief index file as the primary document. These have no compensation content and return None. Download the actual exhibit document instead.
• Non-standard headings — a small minority of filers label their comp section differently (e.g. "Named Executive Officer Compensation"). Tier 2 and Tier 3 catch most of these.
• Multi-subsidiary holding companies — holding companies that file for multiple subsidiaries may return data for only the parent entity.
• Governance field coverage varies — board_size is populated in ~34% of filings (the Q&A answer is often narrative rather than a bare number). auditor is cross-checked with a full-document regex scan as a fallback and populates in ~80% of filings.

Related projects

• edgartools — general-purpose SEC EDGAR library; DEF 14A support is broad but compensation extraction is shallow
• ceopay — Random Forest table classifier + rule-based parsing; compensation only, no governance
• Execcomp-AI — VLM pipeline covering 2005–2022; requires GPU

Contributing

Bug reports and PRs welcome. The most valuable contributions:

• Additional test cases (especially tricky real-world filings)
• Auditor name normalization entries
• Prompt improvements for edge cases

License

MIT