Wrds

Building a proxy-voting panel? Use the npx-ownership-panel skill, not this one. It owns risk.voteanalysis_npx (238M rows / 329 GB), the ISS->CRSP fund crosswalk, and the four-leg SGE pipeline that produces the analysis-ready panel. This skill covers WRDS access patterns generally.

Contents

• WRDS Login Node Enforcement
• Query Enforcement
• SAS ETL Enforcement
• Quick Reference: Table Names
• Connection
• Critical Filters
• Parameterized Queries
• Additional Resources

WRDS Login Node Enforcement

IRON LAW: NEVER RUN COMPUTE ON THE WRDS LOGIN NODE

ALWAYS write an SGE submission script and submit via qsub. No exceptions.

• ssh wrds 'cat files.tsv | ./parser > output.tsv' → WRONG. Use qsub.
• ssh wrds 'nohup ./process &' → WRONG. Still the login node. Use qsub.
• ssh wrds 'python3 bulk_process.py' → WRONG. Use qsub.
• qsub -t 1-20 submit.sh → CORRECT.

The login node is for: qsub, qstat, qdel, scp, ls, head, short psql queries.

Submission patterns and working array jobs: references/edgar.md (§ SGE index build), scripts/sec_index/submit_array.sh, scripts/parse_13f/sge/submit_shards.sh, and ../npx-ownership-panel/scripts/run_pipeline.sh.

Running compute on the login node is NOT HELPFUL — it gets the user's account flagged, the job killed, and the work lost. You run on the login node because qsub feels like overhead. The overhead is 5 minutes of script writing. The downside is account suspension and a rerun from scratch.

Login Node & Infrastructure Facts

• Tests go through the scheduler too: qsub -t 1-1 submit.sh. The login-node "quick test" is the run that flags the account — one file becomes 100K when the command changes, and 173K filings over NFS is not 30 seconds.
• The quorum parser does not run on the login node and never did — it runs via submit_quorum.sh. Citing it as login-node precedent is an unverified claim presented as fact.
• The wrds_clean_filings path convention is cik_int.zfill(10)[:6]/{cik_int}/{accession}.txt (see references/edgar.md). Hand-rolled path logic gets this wrong.
• scan_covers profiles handle header extraction, body parsing, and custom extractors (Custom field type) — "this parser is different enough to need its own binary" has not yet been true once.
• A pixi/conda env under /scratch is not durable. A grid job that ran fine in August 2026 came back rc=127 weeks later: the env its submit script hard-coded had been swept, and no interpreter on WRDS had polars any more. Either rebuild the env as a step of the job, or keep the heavy pull on the grid and do the dataframe work locally. Do not hard-code an env path and assume it survives.

Red Flags — STOP Immediately If You're About To:

• Write ssh wrds '... | ./binary > output' → STOP. That's login-node compute. Write a submit script.
• Write ssh wrds 'nohup ... &' → STOP. nohup doesn't change the node. Use qsub.
• Write ssh wrds 'python3 ...' for anything that reads >10 files → STOP. Use qsub.
• Skip reading references/edgar.md before building a new WRDS file parser → STOP. The path conventions, SGE patterns, and existing parsers are already documented. Read them first.
• Create a new standalone Go binary for EDGAR extraction → STOP. scripts/scan_covers/ is a generic profile-based framework. Add a profiles_*.go file, not a new binary. The framework handles SGE sharding, path construction, concurrency, and form-type filtering.
• Build a new Go/Python parser without checking scripts/scan_covers/ → STOP. This framework exists precisely so you don't reinvent extraction infrastructure. Every standalone parser is technical debt that should have been a profile.

IRON LAW: USE SCAN_COVERS, NOT STANDALONE BINARIES

1. Read scripts/scan_covers/ — generic profile-based Go framework with SGE, concurrency, path handling
2. Add a profiles_*.go file — not a standalone binary. The Profile struct supports pattern-based fields AND custom extractors (set FullBody: true for body-text searches like prospectus 485 filings — see profiles_proxy_advisors.go)
3. Read references/edgar.md — path conventions, existing profiles, SGE submission patterns

Building a standalone parser when scan_covers exists is NOT HELPFUL — it reinvents infrastructure that already handles SGE sharding, NFS concurrency, path construction, form-type filtering, and error handling. You built a 300-line standalone Go binary, ran it on the login node, got the path convention wrong, and spent 5 iterations fixing it. Adding a 60-line profile to scan_covers would have worked on the first try.

Every standalone EDGAR parser is technical debt. The scan_covers framework exists to eliminate this class of mistake.

Exactly two sanctioned exceptions: scripts/parse_13f/ and scripts/parse_npx/. scan_covers FullBody reads the whole file into one buffer per worker and its Field model reduces a regex to one value per column, so a record table — the 13F infoTable, the N-PX proxyTable whose nested voteRecords run to tens of thousands in a single filing — is neither; those two stream the record table instead, emitting one row per record: parse_13f with a hand-rolled information-table scanner, parse_npx with xml.Decoder. Cover-page or header extraction is still a profile, with no exception. Their TSV-to-parquet output contract is scripts/edgar_parquet/ (see references/edgar.md).

WRDS Data Access

WRDS (Wharton Research Data Services) provides academic research data via PostgreSQL at wrds-pgdata.wharton.upenn.edu:9737.

Query Enforcement

IRON LAW: NO QUERY WITHOUT FILTER VALIDATION FIRST

Before executing ANY WRDS query, you MUST:

1. IDENTIFY what filters are required for this dataset
2. VALIDATE the query includes those filters
3. VERIFY parameterized queries (never string formatting)
4. EXECUTE the query
5. INSPECT a sample of results before claiming success

This is not negotiable. Skipping sample inspection is NOT HELPFUL — the user builds analysis on data with undetected quality problems.

Red Flags

• Running a query without checking the Critical Filters section → standard filters apply even when the user doesn't mention them, and even for test queries.
• Pulling everything to filter in pandas later → filter at the database level first.
• Guessing a table name from the request → check the Quick Reference section for exact names.
• Claiming success before sample inspection → inspect .head()/.sample() first; query success ≠ data quality.

Query Validation Checklist

Before EVERY query execution:

For Compustat queries (comp.funda, comp.fundq):

• Includes indfmt = 'INDL'
• Includes datafmt = 'STD'
• Includes popsrc = 'D'
• Includes consol = 'C'
• Uses parameterized queries for variables
• Date range is explicitly specified

For CRSP v2 queries (crsp.dsf_v2, crsp.msf_v2):

• Post-query filter: sharetype == 'NS'
• Post-query filter: securitytype == 'EQTY'
• Post-query filter: securitysubtype == 'COM'
• Post-query filter: usincflg == 'Y'
• Post-query filter: issuertype.isin(['ACOR', 'CORP'])
• Uses parameterized queries

For Form 4 queries (tr_insiders.table1):

• Transaction type filter specified (acqdisp)
• Transaction codes specified (trancode)
• Date range is explicitly specified
• Uses parameterized queries

For ALL queries:

• Sample inspection with .head() or .sample() BEFORE claiming success
• Row count verification (is result size reasonable?)
• NULL value check on critical columns
• Date range validation (does min/max match expectations?)

SAS ETL Enforcement

IRON LAW: NO SAS CODE WITHOUT PERFORMANCE VALIDATION FIRST

1. MERGE STRATEGY — Is hash or sort-merge appropriate? Justify the choice.
2. WHERE CLAUSES — Are all date/string filters index-friendly? No functions on indexed columns.
3. PARALLELISM — Can this job run as an SGE array? Year-by-year is always parallelizable.
4. SQL OPTIMIZATION — For PROC SQL: pass-through opportunity? Indexed join columns?

Writing SAS code that forces full table scans when indexes exist is NOT HELPFUL — the user's job runs 100x slower than necessary and may timeout.

SAS Code Validation Checklist

Before EVERY SAS program execution:

For probing inputs (do this FIRST — metadata only, seconds):

• PROC CONTENTS data=lib.x varnum on every input — variables, types, lengths, formats
• Index section of the CONTENTS listing read — does the WHERE column actually have an index?
• Key lengths compared across datasets to be merged (mismatched $6/$8 gvkey = silent zero matches)
• PROC SQL; select memname, nobs from dictionary.tables where libname='LIB'; — row counts before committing to the job
• PROC PRINT data=lib.x(obs=20); var ...; — values look like the docs claim (always obs=, always var)
• PROC DATASETS library=scratch; — inventory intermediates; delete there, not via a rewriting DATA step

For merges/joins:

• Small lookup + large fact table → hash object (not PROC SORT + DATA merge)
• Hash uses defineKey/defineData/defineDone pattern correctly
• h.output() uses double quotes for macro resolution (not single quotes)
• call missing() initializes hash data variables for non-matches
• Both tables >50M rows → sort-merge is justified (document why)

For WHERE clauses (CRITICAL):

• NO year(date), month(date), datepart(dt) wrapping indexed columns
• Date filters use BETWEEN "01jan&year."d AND "31dec&year."d range pattern
• String filters avoid upcase(), substr() on indexed columns
• Compound date filters collapsed to single range (not year() = X AND quarter() = Y)

For batch processing:

• Multi-year jobs use SGE array (#$ -t start-end) not sequential loop
• Year passed via -sysparm (not -set or %sysget)
• Per-year log files (not single shared log)
• Memory allocation appropriate for workload (#$ -l m_mem_free=4G minimum)
• Single-year benchmark run completed before full array submission

For PROC SQL:

• Join columns are not wrapped in functions
• calculated keyword used for computed column references in HAVING
• Pass-through SQL considered for direct WRDS PostgreSQL queries
• No redundant subqueries that could be hash lookups

For macros:

• Macro variables terminated with period (&year. not &year)
• Double quotes used where macro resolution is needed
• options mprint mlogic symbolgen used during development

SAS Performance Facts

• Hash lookup joins are ~10x faster than PROC SORT + MERGE and need no sorting; PROC SQL still sorts for joins. The hash is 5 extra lines — choosing sort-merge for a lookup join makes the user's job slower for your convenience.
• year(date) (or any function) on an indexed column forces a full table scan over millions of rows; BETWEEN with date literals uses the index.
• Sequential multi-year jobs run ~18x slower than the SGE array (18 years × 3 minutes = 54 minutes sequential vs 3 minutes parallel) — "I'll parallelize later" is anti-efficient on its own terms.
• Single quotes in h.output(dataset: '...') block macro resolution — the output dataset name comes out wrong. Always double quotes.
• %sysget is unreliable under SGE — it may return blank silently. Pass the year via -sysparm + &sysparm..

SAS Red Flags - STOP Immediately If You're About To:

• Write where year(date) = anything → STOP. Use BETWEEN with date literals.
• Write proc sort; data; merge for a lookup join → STOP. Use hash object.
• Write a %do year = start %to end loop → STOP. Use SGE array job.
• Use single quotes in h.output(dataset: '...') → STOP. Use double quotes.
• Submit a full array job without testing one year first → STOP. Benchmark first.
• Use -set or %sysget for SGE task parameters → STOP. Use -sysparm.

SAS Reference

See references/sas-etl.md for complete patterns:

• Probing data and metadata (PROC CONTENTS, PROC DATASETS, PROC PRINT, dictionary.tables)
• Hash object merge (basic, multidata, accumulator)
• Index-friendly WHERE clause quick reference table
• SGE array job templates with memory and logging
• PROC SQL pass-through and optimization
• Macro quoting and debugging

Quick Reference: Table Names

Connection

Initialize PostgreSQL connection to WRDS:

import psycopg2

conn = psycopg2.connect(
    host='wrds-pgdata.wharton.upenn.edu',
    port=9737,
    database='wrds',
    sslmode='require'
    # Credentials from ~/.pgpass
)

Configure authentication via ~/.pgpass with chmod 600:

wrds-pgdata.wharton.upenn.edu:9737:wrds:USERNAME:PASSWORD

Connect via SSH tunnel:

ssh wrds

This uses ~/.ssh/wrds_rsa for authentication.

Critical Filters

Compustat Standard Filters

Always include for clean fundamental data:

WHERE indfmt = 'INDL'
  AND datafmt = 'STD'
  AND popsrc = 'D'
  AND consol = 'C'

CRSP v2 Common Stock Filter

Equivalent to legacy shrcd IN (10, 11):

df = df.loc[
    (df.sharetype == 'NS') &
    (df.securitytype == 'EQTY') &
    (df.securitysubtype == 'COM') &
    (df.usincflg == 'Y') &
    (df.issuertype.isin(['ACOR', 'CORP']))
]

Form 4 Transaction Types

WHERE acqdisp = 'D'  -- Dispositions
  AND trancode IN ('S', 'D', 'G', 'F')  -- Sales, Dispositions, Gifts, Tax

Parameterized Queries

Always use parameterized queries (never string formatting):

Use scalar parameter binding for single values:

cursor.execute("""
    SELECT gvkey, conm FROM comp.company WHERE gvkey = %s
""", (gvkey,))

Use ANY() for list parameters:

cursor.execute("""
    SELECT * FROM comp.funda WHERE gvkey = ANY(%s)
""", (gvkey_list,))

Additional Resources

Reference Files

Detailed query patterns and table documentation:

• references/compustat.md - Compustat tables, ExecuComp, financial variables
• references/crsp.md - CRSP legacy (SIZ) stock data and CCM linking
• ${CLAUDE_SKILL_DIR}/../../skills/crsp-v2/SKILL.md - CRSP CIZ / v2 format (required for any data after 2024-12-31)
• references/insider-form4.md - Thomson Reuters Form 4, rolecodes, insider types
• references/iss-compensation.md - ISS Incentive Lab, peer companies, compensation
• references/formd.md - Form D / Reg D (canonical): two sources (WRDS wrds_vc_formd + SEC EDGAR TSV/XML), grain & keys, denormalization gotcha, exemption + industry codes, post-2020 gap, validated benchmarks
• references/boardex.md - BoardEx: 1999 coverage start and ~20k-company universe, wrds_org_composition is directors PLUS senior managers, sentinel dates, feed 4.2 succession ids, the reused-ticker linking trap, measured recall vs proxy statements
• references/people-linking.md - WRDS People Link: the pairwise person-id tables, which are readable per account, and the Execucomp chain when BoardEx↔CIQ is denied
• references/edgar.md - SEC EDGAR filings, URL construction, DCN vs accession numbers
• references/connection.md - Connection pooling, caching, error handling
• references/taq.md - TAQ: master files, IID, raw tick processing (NBBO, VWAP, closing auctions), CRSP–TAQ merge, era transition (legacy vs millisecond)
• references/sas-etl.md - SAS metadata probing (PROC CONTENTS/DATASETS/PRINT), hash objects, index-friendly WHERE, SGE array jobs, PROC SQL optimization
• references/postgres-vs-sas.md - Decision guide: when to use PostgreSQL vs SAS for WRDS ETL (benchmarks, constraints, hybrid pattern)
• references/fjc.md - FJC Integrated Database: civil/criminal case data, NOS codes, securities litigation queries, firm linking
• references/sdc-issuances.md - SDC New Issues: IPOs, SEOs, 144A equity, debt offerings — schema discovery, cleaning filters, CRSP/Compustat linking
• references/fisd-bonds.md - FISD/Mergent: corporate bond issuances, IG vs HY, 144A vs registered, rating classification, TRACE linking
• references/sdc-ma.md - SDC M&A: deal counts, PE/LBO vs strategic buyer, deal status codes, public vs private target
• references/fund-formation.md - Fund formation: Form D (pooled investment funds), EDGAR N-2 (closed-end fund IPOs), Form ADV (RIA registrations)
• references/capiq.md - Capital IQ: the eddyhu/edwin_hu account split (ciq_pplintel vs boardex_na, both plink tables denied), wrds_professional board panel + the boardflag/sponsorflag traps, ciqcompanyrel relationship types, the keyed sponsor→portco→director→sponsor-employment join
• references/pitchbook.md - PitchBook: schema architecture, dealsize/fundsize in USD millions, dealdate outliers, CIK crosswalk, fund performance (wrds_fund_returns), PE/VC/fund formation patterns
• references/proxy-advisors.md - Proxy-advisor customer identification: 485BPOS/485APOS body scan for ISS/Glass Lewis/Egan-Jones name variants; CRSP MFDB lift to mgmt_cd × year; validates against chongshu published CSV
• references/linkage.md - Cross-dataset linkage map: which identifiers are spines, the load-bearing link tables (CCM, wciklink, dswslink, MFDB), a "how do I join X to Y" table, and which vendor ids never cross
• references/blockholders.md - 13D/13G blockholder panel: Volkova replication, position %, the four mutually-exclusive holder flags
• references/execucomp.md - ExecuComp: CEO anncomp, legacy codirfin vs current directorcomp, firm-year aggregation
• references/iss-directors.md - ISS Directors: risk.directors + risk.rmdirectors, type harmonization, 1996 gender backfill, S&P 1500 filter
• references/iss-voting.md - ISS Voting Analytics: vavoteresults, voteanalysis_npx, base-conditional turnout/forpct, agenda codes
• references/tfn-ownership.md - Thomson 13-F (S34) institutional ownership and S12 mutual-fund holdings via MFLINKS, passive/index classification, and Known Data Defects (D1-D9: split mis-adjustment, post-2013 coverage collapse, 2017Q4 S12 feed change, 13F value unit break, and two that are yours not the vendor's — D8 silent Int8 date overflow, D9 ownership above 100%). Read the defects section before trusting any split-era or post-2013 quarter.
  • Detectors: scripts/ownership_dq.py (14 detectors, S12 and S34) — run these against any holdings panel before analysis. Tests: tests/ownership_dq_test.py (79 assertions, stdlib only).
  • Run detect_calendar_bucket_gap on every reference/dimension table at build time, not just on the output panel. It is the one detector that catches a root cause rather than a symptom: a reference table missing a whole calendar bucket makes every downstream join fall back to a default, silently, and the result looks like a vendor defect (see D8).
• references/lpc-dealscan.md - LPC DealScan: legacy vs 2021+ flat schema, borrower ids, the gvkey link and its grain caveats
• references/muni-bonds.md - Municipal bonds: MSRB RTRS trades, SDC municipals
• references/wrds-forms-tables.md - wrdssec_all.wrds_forms and friends: filing metadata tables and their columns

Example Files

Working code from real projects:

• examples/form4_disposals.py - Insider trading analysis (from SVB project)
• examples/wrds_connector.py - Connection pooling pattern
• examples/formd_regd.ipynb - Form D / Reg D: dedup validation, SEC TSV download, exemption trend charts
• examples/sdc_issuances_eda.ipynb - SDC New Issues: annual IPO/SEO/debt counts, 144A share, IG vs HY breakdown
• examples/sdc_ma_eda.ipynb - SDC M&A: annual deal counts, PE/LBO vs strategic, public vs private target trends
• examples/fund_formation_eda.ipynb - Fund formation: Form D 3C.1/3C.7 counts, EDGAR N-2 closed-end fund IPOs, Form ADV RIA registrations
• examples/pitchbook_eda.ipynb - PitchBook: PE deal activity, VC rounds by stage, fund formation by vintage, IRR/TVPI by strategy
• npx-ownership-panel SKILL (promoted out of this skill's examples) - the full meeting-level proxy-voting x ownership panel: ISS N-PX fund votes reduced to (item x block) cells on the grid, joined to 13-F institutional and MF holdings. One bash command, verified end to end on 2026-07-25. Also carries the ISS->CRSP fund crosswalk. Use it for any N-PX or fund-level voting work.
• examples/blockholders_pipeline/ - 13D/13G → Volkova blockholder panel, end-to-end Python. redo_bridge.py is the reference implementation of TR personid → SEC rptOwnerCik name bridging (97.4% hit rate).
• examples/form4_pipeline/ - Two parallel Form 3/4/5 pipelines: the annualized SAS ownership panel and the XML owner bridge built from the raw filings.
• examples/proxy_advisors_pipeline/ - 485BPOS/485APOS scan for ISS / Glass Lewis / Egan-Jones customer relationships via the scan_covers Go framework + SGE.
• examples/fjc_eda.ipynb - FJC Integrated Database: securities cases (nos = 850), filing trends, court distribution
• examples/lpc_dealscan_eda.ipynb (paired script: examples/lpc_dealscan_eda.py) - LPC DealScan: ~171K US facilities 1990-2020 (the normalized facility table; queries are capped at 2020-12-31), volume by year, loan type and purpose mix
• examples/voting_ownership_eda.py - Standalone Python/PostgreSQL EDA of the same ISS-votes + ownership merge. For production work use the npx-ownership-panel skill, which is the SGE-ready, verified-end-to-end version of this analysis.

Scripts

• scripts/test_connection.py - Validate WRDS connectivity
• scripts/inventory_schemas.py - Inventory every accessible WRDS PostgreSQL schema, its tables, and row counts — run this before guessing at a table name
• scripts/scan_covers/ - Generic profile-based Go framework for EDGAR extraction (SGE sharding, NFS concurrency, path construction, form-type filtering). Add a profiles_*.go, never a new standalone binary — see the Iron Law above.
• scripts/parse_13f/, scripts/scan_headers/, scripts/sec_index/ - Companion EDGAR tooling: 13F table parsing, SEC header scanning, index building

Local Sample Notebooks

WRDS-provided samples at ~/resources/wrds-code-samples/:

• ResearchApps/CCM2025.ipynb - Modern CRSP-Compustat merge
• ResearchApps/ff3_crspCIZ.ipynb - Fama-French factor construction
• comp/sas/execcomp_ceo_screen.sas - ExecuComp patterns

Date Awareness

When querying historical data, leverage current date context for dynamic range calculations.

Current date is automatically available via datetime.now(). Apply this to:

• Data range validation (e.g., "get data for last 5 years")
• Fiscal year calculations
• Event study windows

Implement dynamic date ranges in queries:

from datetime import datetime, timedelta

# Query last 5 years of data
end_date = datetime.now()
start_date = end_date - timedelta(days=5*365)

query = """
SELECT * FROM comp.funda
WHERE datadate BETWEEN %s AND %s
"""
df = pd.read_sql(query, conn, params=(start_date, end_date))

Always incorporate current date awareness in date-dependent queries to ensure results remain fresh across time.