WRDS Schema Pre-loader

You are starting a session that involves WRDS PostgreSQL queries. Before writing any queries, load the schema knowledge you need by dispatching specialist agents in parallel.

Examples

• /wrds-schema crsp -- load CRSP schema only
• /wrds-schema crsp optionm -- load CRSP + OptionMetrics schemas
• /wrds-schema all -- load all available schemas

Connection

Connection details are in ~/.pg_service.conf; password in ~/.pgpass.

psql service=wrds

For Python, use psycopg2.connect("service=wrds"). Do NOT use pd.read_sql (broken with psycopg2 on Python 3.13) — use cursors directly and cast Decimal to float.

Prerequisite: The specialist agents (crsp-wrds-expert, optionmetrics-wrds-expert, taq-wrds-expert) must have Bash in their tools: frontmatter in ~/.claude/agents/. Without it, they cannot run psql queries.

What to do

Based on $ARGUMENTS (or "all" if none given), launch the appropriate specialist agents in parallel using the Task tool. Each agent should query the WRDS database to retrieve current schema details and return a concise reference.

Agent dispatches

For each requested database, launch a Task with subagent_type set to the matching specialist:

Keyword	Agent	Task
crsp	crsp-wrds-expert	Retrieve schema for crsp.dsf, crsp.dsi, crsp.msf, crsp.msi, crsp.stocknames, delisting tables. Confirm column names (date not caldt), PERMNO lookup for SPY (84398), index columns in dsi (spindx, sprtrn). Note Decimal type from psycopg2.
optionm	optionmetrics-wrds-expert	Retrieve schema for optionm.opprcd{YYYY} (yearly partitioned), optionm.securd, optionm.zerocd. Confirm: strike_price is strike*1000, SPX SECID=108105, SPY SECID=109820, XSP SECID=189691. Column names for options tables. LEAPS expiry cycles (Jun/Dec for 2Y).
comp	crsp-wrds-expert	Retrieve Compustat comp.funda/comp.fundq key columns, standard filters (INDL/STD/D/C), CCM linking via crsp.ccmxpf_lnkhist (columns, linktype LC/LU, linkprim P/C, deduplication pattern).
taq	taq-wrds-expert	Retrieve TAQ table structure, trade filtering rules (TR_CORR, TR_SCOND), NBBO tables. Note: TAQ requires SAS on WRDS, not SQL.
jkp	crsp-wrds-expert	Load Global Factor Data schema: contrib.global_factor key columns (permno, gvkey, eom, excntry, me, ret_exc, be_me, etc.), standard filters (excntry='USA', obs_main=1, common=1, exch_main=1, primary_sec=1), column categories, performance rules (always filter excntry + date range). 443 cols, 30M+ rows.
bonds	bonds-wrds-expert	Load Dickerson corporate bond schema: monthly (contrib.dickerson_bonds_monthly, 140 cols, 2.7M rows) and daily (contrib.dickerson_bonds_daily, 43 cols, 30M rows). Rating encoding (1-22), return measures, spread/duration metrics, identifier linkage to CRSP/Compustat (84%). Column names differ between tables.

Agent prompt template

Each agent should receive a prompt like:

You are pre-loading schema knowledge for a WRDS session. Query the database to confirm current table structures and return a concise reference card. Do NOT write analysis code — just retrieve and summarize schema.

Run these queries via psql:

1. SELECT column_name, data_type FROM information_schema.columns WHERE table_schema='...' AND table_name='...' ORDER BY ordinal_position;
2. Any quick validation queries (e.g. SELECT MIN(date), MAX(date) FROM crsp.dsi LIMIT 1)

Return a compact reference with: table name, key columns (name + type), date ranges, important IDs, and known gotchas.

After agents return

Compile their results into a single Schema Reference block and present it to the user. Keep it compact — a lookup table, not prose. Example format:

=== CRSP ===
crsp.dsi: date(date), vwretd(numeric), sprtrn(numeric), spindx(numeric) | 1926-2024
crsp.dsf: permno(int), date(date), prc(numeric), ret(numeric), vol(numeric) | 1926-2024
  SPY PERMNO: 84398
  Gotcha: psycopg2 returns Decimal, cast with .apply(float)

=== OPTIONMETRICS ===
optionm.opprcd{YYYY}: secid, date, exdate, cp_flag, strike_price(/1000!), best_bid, best_offer, impl_volatility, delta, volume, open_interest
  SPX SECID: 108105 | SPY SECID: 109820 | XSP SECID: 189691
  2Y LEAPS: expire Jun/Dec only, DTE 600-830 range for quarterly initiation
optionm.zerocd: date, days, rate | zero-coupon rates by maturity
optionm.securd: secid, ticker, cusip, index_flag
  Gotcha: strike_price is strike * 1000 (divide by 1000 in queries)

=== BONDS (DICKERSON) ===
contrib.dickerson_bonds_monthly: cusip, date, issuer_cusip, permno(84%), gvkey(84%), ret_vw, cs, md_dur, ytm, spc_rat(1-22) | 2002-2025
  Rating: 1=AAA...10=BBB-(IG), 11=BB+...21=CCC-(HY), 22=Default
  Credit spread column is `cs`, NOT `cs_sprd` (which is Corwin-Schultz liquidity spread)
  Gotcha: multiple rows per issuer (one per bond tranche); tret NULL on recent dates; gvkey is double not varchar
contrib.dickerson_bonds_daily: cusip_id, trd_exctn_dt, pr, prfull, credit_spread, mod_dur, ytm, prc_bid, prc_ask, trade_count | 2002-2025
  30M rows, sparse (only trading days), no returns (compute from prices), no issuer_cusip
  Column names differ from monthly! (cusip_id/credit_spread/mod_dur/spc_rating vs cusip/cs/md_dur/spc_rat)

Known gotchas (always include these)

1. pd.read_sql broken on Python 3.13 — use cursor.execute() + cursor.fetchall() + pd.DataFrame(rows, columns=[...])
2. Decimal type — psycopg2 returns decimal.Decimal for numeric columns. Always .apply(float) or float() before math/numpy operations.
3. CRSP column names — crsp.dsi uses date (not caldt), spindx (not sprindx)
4. OptionMetrics strike — strike_price column stores strike * 1000. Divide by 1000.
5. OptionMetrics yearly tables — option prices are in optionm.opprcd{YYYY}, NOT a single table. Loop years 1996-2025.
6. SPX LEAPS cycle — 2-year options only expire in June and December. Monthly initiations produce duplicate expiry targets. Use quarterly initiation for 2Y.
7. Exec pricing — use (best_bid + 3*best_offer) / 4 for realistic execution (between mid and ask).