Financial data analysis

Financial Data Analysis

A practical skill for sourcing, processing, and analyzing financial data in academic research contexts. Covers data acquisition from public APIs, cleaning workflows, and standard analytical techniques used in empirical finance research.

Data Acquisition

Public Financial Data Sources

Source	Data Type	Access	Python Package
Yahoo Finance	Prices, fundamentals	Free	yfinance
FRED (St. Louis Fed)	Macroeconomic indicators	Free (API key)	fredapi
SEC EDGAR	Company filings (10-K, 10-Q)	Free	sec-edgar-downloader
WRDS (Wharton)	CRSP, Compustat, IBES	University subscription	wrds
Alpha Vantage	Real-time and historical prices	Free tier	alpha_vantage

Fetching Price Data

import yfinance as yf
import pandas as pd

def fetch_stock_data(tickers: list[str], start: str, end: str) -> pd.DataFrame:
    """
    Fetch adjusted close prices for a list of tickers.

    Args:
        tickers: List of ticker symbols (e.g., ['AAPL', 'MSFT'])
        start: Start date (YYYY-MM-DD)
        end: End date (YYYY-MM-DD)
    Returns:
        DataFrame with adjusted close prices
    """
    data = yf.download(tickers, start=start, end=end, auto_adjust=True)
    prices = data['Close'] if len(tickers) > 1 else data[['Close']]
    prices.columns = tickers if len(tickers) > 1 else tickers
    return prices

# Fetch 5 years of data
prices = fetch_stock_data(['AAPL', 'MSFT', 'GOOGL'], '2020-01-01', '2025-01-01')
print(prices.head())

Macroeconomic Data from FRED

from fredapi import Fred

fred = Fred(api_key=os.environ["FRED_API_KEY"])

# Common series for finance research
series_ids = {
    'GDP': 'GDP',
    'CPI': 'CPIAUCSL',
    'Fed_Funds_Rate': 'FEDFUNDS',
    'Unemployment': 'UNRATE',
    '10Y_Treasury': 'DGS10',
    'VIX': 'VIXCLS'
}

macro_data = pd.DataFrame()
for name, sid in series_ids.items():
    macro_data[name] = fred.get_series(sid, observation_start='2000-01-01')

Data Cleaning Pipeline

Financial data requires careful cleaning before analysis:

def clean_financial_data(df: pd.DataFrame) -> pd.DataFrame:
    """Standard cleaning pipeline for financial time series."""
    cleaned = df.copy()

    # 1. Handle missing values
    missing_pct = cleaned.isnull().sum() / len(cleaned) * 100
    print(f"Missing data:\n{missing_pct}")

    # 2. Forward-fill for market holidays (max 5 days)
    cleaned = cleaned.ffill(limit=5)

    # 3. Remove remaining NaN rows
    cleaned = cleaned.dropna()

    # 4. Detect and flag outliers (>5 sigma daily returns)
    returns = cleaned.pct_change()
    z_scores = (returns - returns.mean()) / returns.std()
    outliers = (z_scores.abs() > 5).any(axis=1)
    print(f"Outlier days flagged: {outliers.sum()}")

    # 5. Verify data integrity
    assert cleaned.index.is_monotonic_increasing, "Index must be sorted"
    assert not cleaned.duplicated().any(), "No duplicate rows allowed"

    return cleaned

Standard Financial Metrics

Return Calculations

def compute_returns(prices: pd.DataFrame) -> dict:
    """Compute standard return metrics."""
    simple_returns = prices.pct_change().dropna()
    log_returns = np.log(prices / prices.shift(1)).dropna()

    annualized_return = simple_returns.mean() * 252
    annualized_vol = simple_returns.std() * np.sqrt(252)
    sharpe_ratio = annualized_return / annualized_vol

    # Maximum drawdown
    cumulative = (1 + simple_returns).cumprod()
    rolling_max = cumulative.cummax()
    drawdown = (cumulative - rolling_max) / rolling_max
    max_drawdown = drawdown.min()

    return {
        'annualized_return': annualized_return,
        'annualized_volatility': annualized_vol,
        'sharpe_ratio': sharpe_ratio,
        'max_drawdown': max_drawdown
    }

Event Studies

A common methodology in empirical finance research:

1. Define the event window (e.g., [-5, +5] trading days around earnings announcement)
2. Estimate normal returns using the market model over the estimation window (e.g., [-250, -30])
3. Compute abnormal returns: AR = R_actual - R_expected
4. Aggregate cumulative abnormal returns (CAR) across firms
5. Test statistical significance using parametric (Patell test) and non-parametric (sign test) methods

Always report both raw and risk-adjusted results, and perform robustness checks with different estimation windows and benchmark models.

Reproducibility

Store all data processing steps in version-controlled scripts. Use pandas.DataFrame.to_parquet() for efficient storage of intermediate datasets, and document data provenance including download dates, API versions, and any filters applied.