Appendix T — Data Card — CRSP (Center for Research in Security Prices)

Provider & what it is. CRSP, maintained at the University of Chicago Booth School and distributed through WRDS, is the canonical record of U.S. stock prices, returns, shares outstanding, and — its single most valuable feature — delisting returns and a survivorship-bias-free universe of listed equities back to 1925. When a finance paper says “we use CRSP monthly returns,” this is what it means. The reveal-the-trick point: CRSP is not just “stock prices”; it is a carefully maintained, point-in-time-correct, survivorship-free panel — exactly the correctness free sources like Yahoo do not give you. Sam’s momentum backtests and every anomaly study in Week 5 rest on it.

Coverage

Span. Monthly stock file (crsp.msf) and daily stock file (crsp.dsf) back to 1925/1926; coverage deepens over time as exchanges were added (NYSE first, then AMEX, then Nasdaq from 1972). [CHECK] exact start dates per exchange.
Frequency. Monthly and daily. (Annual and weekly derived files exist; the two workhorses are msf and dsf.)
Universe. U.S. common stocks, ADRs, REITs, closed-end funds, and more, listed on NYSE/AMEX/Nasdaq/Arca. Crucially it includes securities that have since delisted, so a backtest that uses CRSP does not silently drop the losers — the survivorship-bias fix you study in Ch 7.4.

Key identifiers

PERMNO — a permanent security identifier. Unlike a ticker, it never changes when a firm renames or re-tickers, and it is never reused. This is the join key you carry everywhere.
PERMCO — a permanent company identifier (a company can have several PERMNOs, e.g., multiple share classes).
Tickers and CUSIPs are also present but are historical and can change; never join on ticker.
The link to Compustat’s GVKEY runs through the CRSP/Compustat Merged table crsp.ccmxpf_lnkhist (respect linktype, linkprim, and the link’s valid date range) — a Ch 7.4 crosswalk, not a string match.

Access path

WRDS is a remote PostgreSQL database; you send SQL to it with the wrds package and get back a DataFrame. Credentials come from .pgpass or the environment — never hard-coded (CONVENTIONS §5).

import os, wrds, pandas as pd, pathlib

db = wrds.Connection(wrds_username=os.environ["WRDS_USERNAME"])  # secret from env

# SELECT only the columns/rows you need; bound the dates. Never SELECT * on msf/dsf.
raw = pathlib.Path("data/raw/crsp_msf_2010_2020.parquet")        # git-ignored dir
if raw.exists():
    msf = pd.read_parquet(raw)
else:
    msf = db.raw_sql("""
        SELECT permno, date, ret, retx, prc, shrout, vol
        FROM   crsp.msf
        WHERE  date BETWEEN '2010-01-01' AND '2020-12-31'
    """, date_cols=["date"])
    # Delisting returns live in a SEPARATE table -- pull and merge them in:
    dl = db.raw_sql("""
        SELECT permno, dlstdt, dlret, dlstcd
        FROM   crsp.msedelist
    """, date_cols=["dlstdt"])
    msf.to_parquet(raw)        # cache the raw pull to GMU infrastructure only
db.close()

Explore the schema first with db.list_tables(library="crsp") and db.describe_table("crsp","msf").

License & the “stays on GMU infrastructure” note (CONVENTIONS §5)

CRSP is licensed, not public — GMU pays for the right to use it under terms that forbid redistributing the raw records. The connection is read-only. The hard consequence: you may query CRSP only on WRDS Cloud or GMU’s Hopper cluster, your derived extract is cached to a git-ignored data/raw/, and the licensed bytes are never committed, emailed off campus, posted to a public repo, or pasted into any commercial web service. What travels is the recipe — your query code and your pull log — not the data. Anyone with their own GMU WRDS seat can re-run your code; that is reproducibility done within the license.

Gotchas

RET vs. RETX. RET includes dividends (total return — what you almost always want); RETX excludes them. Using the wrong one silently biases every return-based result.
Delisting returns are in a separate table (crsp.msedelist). Ignore them and you get survivorship bias precisely when a firm fails — the worst time to be missing data. Merge dlret in.
Missing-return codes. RET can carry special non-numeric values (e.g., for halts/missing prices); coerce and inspect, do not blindly cast to float. [CHECK] the current code set.
Negative PRC means the price is a bid/ask midpoint, not a trade; take the absolute value before using it as a price.
Shares outstanding (shrout) are in thousands — market cap = |prc| * shrout * 1000.
Size of dsf. The daily file is tens of millions of rows; SELECT * will time out or exhaust memory. Bound by date and PERMNO.

“First 10 rows” schema sketch (illustrative — not real CRSP data)

permno (int)	date (date)	ret (float)	retx (float)	prc (float)	shrout (int, 000s)	vol (float)
10001	2010-01-29	0.0312	0.0298	24.13	41200	1.2e6
10001	2010-02-26	-0.0145	-0.0145	23.78	41200	9.8e5
10026	2010-01-29	0.0501	0.0501	112.40	88300	4.1e6
10026	2010-02-26	0.0088	0.0061	113.39	88300	3.7e6
14593	2010-01-29	0.0210	0.0210	-56.02	901000	2.2e7
…	…	…	…	…	…	…

(Values invented to show shape and types: note the negative prc in row 5 — a bid/ask midpoint — and ret≠retx when a dividend is paid.)

Which chapter/lab/capstone uses it

Week 2, Lab 2 — “Replicate a Textbook Fama–MacBeth on CRSP” (Path A pulls crsp.msf; Path B is the synthetic fallback). The lab pins the CRSP snapshot date.
Week 5 — nb5.1–nb5.3 (Fama–French 1992/1993 sorts, Jegadeesh–Titman momentum) build portfolios on CRSP/Compustat.
Week 7, Ch 7.2 — CRSP is the lead example of the WRDS access pattern and the GMU-infrastructure rule.
Sam’s anomaly/momentum thread; any capstone needing survivorship-free U.S. equity returns.