科研 Skill

WRDS金融数据库访问

wrds-data-access

Accesses WRDS databases (Compustat fundamentals, CRSP returns, TAQ microstructure, linking tables) via the wrds Python package. Use when pulling data from a WRDS subscription account.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

WRDS Data Access for Academic Finance Research

One-line summary: Query Compustat, CRSP, and TAQ from WRDS using the Python API: financial fundamentals, stock returns, intraday microstructure, and CRSP-Compustat link table.


When to Use This Skill

  • When downloading Compustat balance sheet / income statement data
  • When fetching CRSP monthly or daily stock returns and market cap
  • When linking Compustat (gvkey) to CRSP (permno) via CCM link table
  • When extracting TAQ intraday trades and quotes for microstructure research
  • When building research-grade datasets (accounting ratios, anomalies)
  • When accessing I/B/E/S earnings forecasts or Execucomp compensation

Trigger keywords: WRDS, Compustat, CRSP, CCM, TAQ, gvkey, permno, cusip, annual fundamentals, stock returns, book-to-market, earnings surprise, I/B/E/S, Execucomp, institutional holdings


Background & Key Concepts

WRDS Architecture

WRDS (Wharton Research Data Services) exposes PostgreSQL databases via:

  1. Python package (wrds): Connects via SSH tunnel + PostgreSQL; requires institutional subscription
  2. SQL queries: Standard SQL on schemas like comp (Compustat), crsp, taq

Key Identifiers

IdentifierDatabaseDescription
gvkeyCompustatPermanent company ID
permnoCRSPPermanent security ID
cusipBothCommittee on Uniform Securities
tickerMultipleNot permanent — changes

CRSP-Compustat Merge (CCM)

The crsp.ccmxpf_lnkhist table maps gvkey → permno with date-valid links. Always filter by linktype IN ('LC','LU','LX','LD') and linkprim IN ('P','C').


Environment Setup

Install Dependencies

pip install wrds>=3.1 pandas>=2.0 numpy>=1.24 sqlalchemy>=2.0 matplotlib>=3.7

WRDS Credentials

# Credentials are stored securely in ~/.pgpass after first login
# Set environment variables (do NOT hardcode)
export WRDS_USERNAME="<your-wrds-username>"

Connect to WRDS

import os
import wrds

# Credentials via environment variable
wrds_username = os.getenv("WRDS_USERNAME", "")
if not wrds_username:
    raise ValueError("Set WRDS_USERNAME environment variable")

db = wrds.Connection(wrds_username=wrds_username)
print("Connected to WRDS")

# List available libraries
libs = db.list_libraries()
print("Available libraries (first 10):", libs[:10])

Core Workflow

Step 1: Compustat Annual Fundamentals

import os
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# ------------------------------------------------------------------ #
# Download Compustat annual fundamentals for S&P 500 firms
# Key variables for accounting research
# ------------------------------------------------------------------ #

try:
    import wrds
    db = wrds.Connection(wrds_username=os.getenv("WRDS_USERNAME", ""))

    # Query Compustat annual (comp.funda)
    query = """
        SELECT
            gvkey, datadate, fyear, conm, sich,
            at,         -- Total assets
            lt,         -- Total liabilities
            ceq,        -- Common equity (book value)
            ni,         -- Net income
            sale,       -- Net sales
            oibdp,      -- Operating income before depreciation
            dp,         -- Depreciation & amortization
            capx,       -- Capital expenditure
            che,        -- Cash and equivalents
            dltt,       -- Long-term debt
            dlc,        -- Short-term debt
            csho,       -- Common shares outstanding
            prcc_f,     -- Fiscal-year-end price
            sich        -- SIC code
        FROM comp.funda
        WHERE
            indfmt = 'INDL'       -- Industrial format
            AND datafmt = 'STD'    -- Standardized
            AND popsrc = 'D'       -- Domestic
            AND consol = 'C'       -- Consolidated
            AND fyear BETWEEN 2010 AND 2023
            AND at > 0
            AND sale > 0
        ORDER BY gvkey, datadate
        LIMIT 50000
    """
    comp = db.raw_sql(query, date_cols=['datadate'])
    print(f"Compustat: {comp.shape[0]:,} firm-years downloaded")

except Exception as e:
    print(f"WRDS connection failed ({e}), using synthetic data")
    # Synthetic Compustat-like data
    np.random.seed(42)
    n = 5000
    gvkeys = np.repeat(np.arange(1, 101), 50)[:n]
    fyears = np.tile(np.arange(2010, 2024), 500)[:n]

    comp = pd.DataFrame({
        'gvkey': gvkeys,
        'fyear': fyears,
        'at':    np.random.lognormal(7, 1.5, n),    # Total assets
        'ceq':   np.random.lognormal(6, 1.5, n),    # Book equity
        'ni':    np.random.normal(0.1, 0.3, n) * np.random.lognormal(6, 1.5, n)/10,
        'sale':  np.random.lognormal(6.5, 1.5, n),
        'csho':  np.random.lognormal(4, 1.5, n),
        'prcc_f': np.random.lognormal(3.5, 0.8, n),
        'dltt':  np.random.lognormal(5, 2, n),
        'capx':  np.abs(np.random.normal(0, 0.05, n)) * np.random.lognormal(6.5, 1.5, n),
    })

# ---- Compute key accounting ratios ----------------------------- #
comp = comp.copy()
comp['mve']  = comp['csho'] * comp['prcc_f']                    # Market cap
comp['bm']   = comp['ceq'] / comp['mve'].replace(0, np.nan)      # Book/Market
comp['roa']  = comp['ni'] / comp['at'].replace(0, np.nan)        # ROA
comp['roe']  = comp['ni'] / comp['ceq'].replace(0, np.nan)       # ROE
comp['lev']  = comp['dltt'] / comp['at'].replace(0, np.nan)      # Leverage
comp['asset_growth'] = comp.groupby('gvkey')['at'].pct_change()  # Asset growth

# Winsorize at 1st/99th percentile
def winsorize(series, low=0.01, high=0.99):
    lb = series.quantile(low)
    ub = series.quantile(high)
    return series.clip(lb, ub)

for col in ['bm', 'roa', 'roe', 'lev', 'asset_growth']:
    comp[col] = winsorize(comp[col])

print("\nAccounting ratios summary:")
print(comp[['bm', 'roa', 'roe', 'lev', 'asset_growth']].describe().round(4))

# ---- Portfolio B/M quintiles ----------------------------------- #
latest = comp[comp['fyear'] == comp['fyear'].max()].copy()
latest = latest[latest['mve'] > 0].dropna(subset=['bm'])
latest['bm_quintile'] = pd.qcut(latest['bm'], 5,
                                  labels=['Growth','Q2','Q3','Q4','Value'])

fig, axes = plt.subplots(1, 2, figsize=(12, 5))

# ROA by BM quintile
roa_by_bm = latest.groupby('bm_quintile')['roa'].mean()
axes[0].bar(roa_by_bm.index, roa_by_bm.values, color='steelblue', edgecolor='black', linewidth=0.7)
axes[0].set_xlabel("B/M Quintile"); axes[0].set_ylabel("Mean ROA")
axes[0].set_title("ROA by Book-to-Market Quintile")
axes[0].grid(axis='y', alpha=0.3)

# Leverage distribution
axes[1].hist(latest['lev'].dropna(), bins=40, color='coral', edgecolor='white', linewidth=0.5)
axes[1].axvline(latest['lev'].median(), color='navy', linewidth=2, linestyle='--',
                label=f"Median={latest['lev'].median():.2f}")
axes[1].set_xlabel("Leverage (LT Debt/Assets)"); axes[1].set_title("Leverage Distribution")
axes[1].legend(); axes[1].grid(alpha=0.3)

plt.tight_layout()
plt.savefig("compustat_analysis.png", dpi=150)
plt.show()

Step 2: CRSP Monthly Returns

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# ------------------------------------------------------------------ #
# Download CRSP monthly stock file and compute factor portfolio returns
# ------------------------------------------------------------------ #

try:
    import wrds
    import os
    db = wrds.Connection(wrds_username=os.getenv("WRDS_USERNAME", ""))

    crsp_query = """
        SELECT
            permno, date, ret, retx, shrout, prc, exchcd, shrcd
        FROM crsp.msf
        WHERE
            date BETWEEN '2010-01-01' AND '2023-12-31'
            AND shrcd IN (10, 11)       -- Common shares, US incorporated
            AND exchcd IN (1, 2, 3)     -- NYSE, AMEX, NASDAQ
            AND ret IS NOT NULL
        ORDER BY permno, date
        LIMIT 200000
    """
    crsp = db.raw_sql(crsp_query, date_cols=['date'])
    print(f"CRSP monthly: {crsp.shape[0]:,} obs downloaded")

except Exception as e:
    print(f"Using synthetic CRSP data ({e})")
    np.random.seed(42)
    n = 50000
    permnos = np.repeat(np.arange(1, 1001), 60)[:n]
    dates = pd.bdate_range("2010-01-31", periods=60, freq='BM')
    dates = np.tile(dates[:60], 1000)[:n]

    crsp = pd.DataFrame({
        'permno': permnos,
        'date':   dates,
        'ret':    np.random.normal(0.01, 0.08, n),
        'prc':    np.abs(np.random.lognormal(3.5, 0.8, n)),
        'shrout': np.random.lognormal(8, 1.5, n) * 1000,
        'exchcd': np.random.choice([1, 2, 3], n),
    })

# ---- Market equity and value-weighted returns ----------------- #
crsp['me'] = crsp['prc'].abs() * crsp['shrout'] / 1000  # in thousands

# Value-weighted market return
crsp = crsp.sort_values(['permno', 'date'])
crsp['me_lag'] = crsp.groupby('permno')['me'].shift(1)

def vw_return(group):
    w = group['me_lag'].clip(lower=0)
    if w.sum() == 0:
        return np.nan
    return (group['ret'] * w).sum() / w.sum()

mkt_ret = crsp.groupby('date').apply(vw_return).rename('vw_mkt_ret')
ew_ret  = crsp.groupby('date')['ret'].mean().rename('ew_mkt_ret')

comparison = pd.concat([mkt_ret, ew_ret], axis=1).dropna()
cum_vw = (1 + comparison['vw_mkt_ret']).cumprod()
cum_ew = (1 + comparison['ew_mkt_ret']).cumprod()

fig, axes = plt.subplots(1, 2, figsize=(13, 5))
axes[0].plot(cum_vw.index, cum_vw, label='Value-Weighted', linewidth=1.5, color='navy')
axes[0].plot(cum_ew.index, cum_ew, label='Equal-Weighted', linewidth=1.5, color='coral')
axes[0].set_title("Cumulative Market Return"); axes[0].set_ylabel("Cumulative return (×)")
axes[0].legend(); axes[0].grid(alpha=0.3)

# Cross-sectional return distribution
axes[1].hist(crsp['ret'].dropna(), bins=60, range=(-0.5, 0.5),
             color='steelblue', edgecolor='white', linewidth=0.3, alpha=0.8)
axes[1].axvline(crsp['ret'].mean(), color='red', linewidth=2, linestyle='--',
                label=f"Mean={crsp['ret'].mean()*100:.2f}%")
axes[1].set_xlabel("Monthly Return"); axes[1].set_title("Return Distribution (CRSP)")
axes[1].legend(); axes[1].grid(alpha=0.3)

plt.tight_layout()
plt.savefig("crsp_returns.png", dpi=150)
plt.show()

Step 3: CRSP-Compustat Merge

import pandas as pd
import numpy as np

# ------------------------------------------------------------------ #
# Merge CRSP and Compustat via CCM link table
# Follow standard academic convention (Fama-French style)
# ------------------------------------------------------------------ #

try:
    import wrds
    import os
    db = wrds.Connection(wrds_username=os.getenv("WRDS_USERNAME", ""))

    # Download CCM link table
    ccm_query = """
        SELECT
            gvkey, lpermno AS permno,
            linktype, linkprim, linkdt, linkenddt
        FROM crsp.ccmxpf_lnkhist
        WHERE
            linktype IN ('LC', 'LU', 'LX', 'LD')
            AND linkprim IN ('P', 'C')
    """
    ccm = db.raw_sql(ccm_query, date_cols=['linkdt', 'linkenddt'])
    print(f"CCM link table: {ccm.shape[0]:,} records")

    # Fill missing end dates
    ccm['linkenddt'] = ccm['linkenddt'].fillna(pd.Timestamp('2099-12-31'))

    # Merge: given a date, find valid gvkey-permno link
    def get_permno(gvkey, date, ccm_df):
        """Return permno for a gvkey on a given date."""
        valid = ccm_df[
            (ccm_df['gvkey'] == gvkey) &
            (ccm_df['linkdt'] <= date) &
            (ccm_df['linkenddt'] >= date)
        ]
        if len(valid) == 0:
            return np.nan
        return valid['permno'].iloc[0]

    print("CCM merge function ready")
    print("Usage: permno = get_permno(gvkey='001234', date=pd.Timestamp('2020-06-30'), ccm_df=ccm)")

except Exception as e:
    print(f"WRDS not available ({e})")
    print("Demonstrating CCM merge logic with synthetic data...")

# ---- Standard merge protocol (Fama-French style) --------------- #
# 1. Use Compustat fiscal year ending in year t
# 2. Compute accounting variables in June of year t+1 (6-month lag)
# 3. Match to CRSP returns from July t+1 to June t+2

print("""
Standard CCM Merge Protocol:
1. Download comp.funda (fyear=t)
2. Set 'jdate' = fyear-end date + 6 months (June t+1)
3. Download crsp.msf for July t+1 onward
4. Merge via CCM: crsp.permno matched to comp.gvkey using jdate
5. Lag accounting variables by 6 months for look-ahead bias prevention

Example SQL (full merge):
    SELECT a.*, b.permno
    FROM comp.funda a
    LEFT JOIN crsp.ccmxpf_lnkhist b
        ON a.gvkey = b.gvkey
        AND b.linktype IN ('LC','LU')
        AND b.linkprim IN ('P','C')
        AND b.linkdt <= a.datadate
        AND (b.linkenddt >= a.datadate OR b.linkenddt IS NULL)
""")

Advanced Usage

TAQ Intraday Microstructure

import pandas as pd
import numpy as np

# ------------------------------------------------------------------ #
# Effective spread and price impact from TAQ trades and quotes
# ------------------------------------------------------------------ #

try:
    import wrds, os
    db = wrds.Connection(wrds_username=os.getenv("WRDS_USERNAME", ""))

    # TAQ daily trades for one stock (example: AAPL on 2023-01-03)
    taq_query = """
        SELECT time, price, size, tr_corr, tr_scond
        FROM taq.ct_20230103
        WHERE sym_root = 'AAPL'
            AND tr_corr = '00'   -- Regular trades only
            AND size > 0
        ORDER BY time
        LIMIT 10000
    """
    trades = db.raw_sql(taq_query)
    print(f"TAQ trades: {len(trades):,} records")
except Exception as e:
    print(f"TAQ not available ({e}) — using synthetic data")
    # Simulate intraday trades
    np.random.seed(42)
    n = 5000
    times = pd.date_range("2023-01-03 09:30", periods=n, freq='5s')
    mid_price = 130 + np.cumsum(np.random.randn(n) * 0.01)
    spread = 0.02  # $0.02 spread
    sign = np.random.choice([-1, 1], n)  # Buy/sell indicator
    trades = pd.DataFrame({
        'time':  times,
        'price': mid_price + sign * spread/2 + np.random.randn(n) * 0.001,
        'size':  np.random.randint(100, 2000, n),
        'mid':   mid_price,
        'sign':  sign,
    })

    # Effective spread = 2 * |price - mid| / mid
    trades['eff_spread'] = 2 * np.abs(trades['price'] - trades['mid']) / trades['mid']
    print(f"Mean effective spread: {trades['eff_spread'].mean()*10000:.2f} bps")

    # Amihud illiquidity: |return| / dollar volume
    trades['dollar_vol'] = trades['price'] * trades['size']
    trades['return_5min'] = trades['price'].pct_change(12)  # ~1min return
    trades['amihud'] = trades['return_5min'].abs() / (trades['dollar_vol'] / 1e6)
    print(f"Amihud illiquidity (per $M): {trades['amihud'].mean():.4f}")

Troubleshooting

Error: ConnectionError — WRDS authentication

Fix:

# Re-authenticate
python -c "import wrds; db = wrds.Connection()"
# Enter username and password when prompted — stored in ~/.pgpass

Error: psycopg2.OperationalError: SSL SYSCALL error

Cause: VPN/firewall blocking WRDS connection (port 5432).

Fix: Use WRDS web query interface or connect via campus network/VPN.

Memory error: large SQL query

Fix: Add row limits or filter by date range:

# Download in annual chunks
for year in range(2010, 2024):
    chunk = db.raw_sql(f"""
        SELECT * FROM comp.funda WHERE fyear = {year} AND at > 0
    """)
    chunk.to_parquet(f"compustat_{year}.parquet")

Version Compatibility

PackageTested versionsNotes
wrds3.1.xRequires institutional WRDS subscription
sqlalchemy2.0wrds 3.x requires SQLAlchemy 2.0+
pandas2.0, 2.1date_cols parsing in raw_sql

External Resources

Official Documentation

Key Papers

  • Fama, E.F. & French, K.R. (1993). Common risk factors in returns on stocks and bonds. Journal of Financial Economics.
  • Davis, J.L. et al. (2000). Characteristics, covariances, and average returns. Journal of Finance.

Examples

Example 1: Momentum Factor Construction (UMD)

import pandas as pd
import numpy as np

# ------------------------------------------------------------------ #
# Construct momentum factor: past 12-2 month return
# Following Jegadeesh & Titman (1993)
# ------------------------------------------------------------------ #
np.random.seed(42)
n_stocks = 500
n_months = 72  # 6 years
dates = pd.period_range("2018-01", periods=n_months, freq='M')

# Simulate monthly returns panel
ret_data = pd.DataFrame(
    np.random.normal(0.01, 0.08, (n_months, n_stocks)),
    index=dates,
    columns=[f"S{i:04d}" for i in range(n_stocks)]
)

def compute_momentum(ret_df, formation=12, skip=1):
    """Compute past (formation-skip) to (skip) month return."""
    cum_ret = (1 + ret_df).rolling(formation - skip).apply(np.prod, raw=True) - 1
    return cum_ret.shift(skip)  # Skip most recent month

mom = compute_momentum(ret_df, formation=12, skip=1)

# Sort into deciles each month, compute top-bottom spread
umd_monthly = []
for date in dates[12:]:
    m = mom.loc[date].dropna()
    if len(m) < 50:
        continue
    deciles = pd.qcut(m, 10, labels=False)
    winner = ret_df.loc[date][deciles == 9].mean()  # Top decile (Winners)
    loser  = ret_df.loc[date][deciles == 0].mean()  # Bottom decile (Losers)
    umd_monthly.append(winner - loser)

umd = pd.Series(umd_monthly)
print(f"UMD monthly mean: {umd.mean()*100:.3f}%  (annualized: {umd.mean()*1200:.2f}%)")
print(f"UMD monthly std:  {umd.std()*100:.3f}%")
print(f"UMD Sharpe ratio: {umd.mean()/umd.std()*np.sqrt(12):.3f}")

Example 2: Earnings Surprise (SUE)

import numpy as np
import pandas as pd

# Standardized Unexpected Earnings (SUE)
# SUE = (EPS_actual - EPS_expected) / std(forecast_error)

np.random.seed(0)
n = 2000
eps_actual   = np.random.normal(1.5, 0.8, n)
eps_expected = eps_actual + np.random.normal(0, 0.3, n)  # Analyst forecast
sue = (eps_actual - eps_expected) / (np.abs(eps_actual - eps_expected).rolling(8, min_periods=3).std()
       if hasattr(eps_actual, 'rolling') else np.std(np.abs(eps_actual - eps_expected)))

# Simple SUE
std_err = np.std(eps_actual - eps_expected)
sue_simple = (eps_actual - eps_expected) / (std_err + 1e-8)

# Portfolio returns by SUE quintile
next_ret = np.random.normal(0.01, 0.05, n) + sue_simple * 0.005  # Alpha correlated with SUE
quintiles = pd.qcut(sue_simple, 5, labels=['Q1 (neg)','Q2','Q3','Q4','Q5 (pos)'])
print("Mean next-month return by SUE quintile:")
for q in ['Q1 (neg)','Q2','Q3','Q4','Q5 (pos)']:
    mean_ret = next_ret[quintiles == q].mean() * 100
    print(f"  {q}: {mean_ret:.3f}%")

Last updated: 2026-03-17 | Maintainer: @xjtulyc Issues: GitHub Issues

示例提示词

我想用wrds包提取Compustat财务数据和CRSP股票收益率并做表连接。

所属插件

经济金融与管理研究面向经济、金融与管理学研究:DID/IV/RDD/合成控制、面板与时间序列计量、事件研究与因子模型、期权与组合、管理学论文写作、NSFC 申请与质性/混合方法。

安装该插件会一并获得同领域的 40 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

WRDS金融数据库访问能做什么?

Accesses WRDS databases (Compustat fundamentals, CRSP returns, TAQ microstructure, linking tables) via the wrds Python package. Use when pulling data from a WRDS subscription account。

如何安装WRDS金融数据库访问?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「WRDS金融数据库访问」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

WRDS金融数据库访问需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

WRDS金融数据库访问经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。