Data

Optimizers consume a wide returns panel — rows are periods, columns are assets — as a pandas DataFrame or a 2-D array. jaxfolio provides a reproducible synthetic generator plus loaders for CSV, Parquet, and Yahoo Finance, and utilities to convert prices to returns and split them for out-of-sample testing.

Synthetic data

generate_returns (and its price-level sibling generate_prices) produce correlated geometric-Brownian-motion panels with no network access — this is what makes the tests, examples, and every snippet in these docs reproducible and offline-friendly.

import jaxfolio as jf

returns = jf.generate_returns(
    n_assets=12,
    n_days=756,                       # ≈ 3 years of trading days
    seed=7,                           # reproducible
    correlation_strength=0.5,         # average cross-asset coupling in [0, 1)
    annual_vol_range=(0.15, 0.45),    # per-asset vols drawn from this range
    mean_annual_return=0.08,
)

The generator builds a valid PSD correlation matrix from a low-rank factor model, Cholesky-factorizes it, and simulates GBM log-returns. Pass tickers=[...] for explicit column names (otherwise ASSET_00…).

Loading real data

Requires the data extra

The Yahoo Finance and Parquet loaders need uv add "jaxfolio[data]".

Yahoo FinanceCSVParquet

prices = jf.load_yfinance(["AAPL", "MSFT", "NVDA", "JPM"], period="2y", interval="1d")
returns = jf.to_returns(prices)

Returns a wide date-by-ticker panel of adjusted close prices.

# Wide layout: one column per asset + a date column
prices = jf.load_csv("prices.csv", date_col="date")

# Long/tidy layout: pivot date/asset/price columns into a panel
prices = jf.load_csv("prices.csv", date_col="date",
                     asset_col="ticker", price_col="close")

prices = jf.load_parquet("prices.parquet", date_col="date")

Prices → returns

to_returns converts a price panel to simple or log returns:

returns = jf.to_returns(prices, kind="simple")   # p_t / p_{t-1} − 1
returns = jf.to_returns(prices, kind="log")       # log(p_t / p_{t-1})

Out-of-sample splitting

train_test_split splits chronologically (never shuffled — this is time-series data), holding out the most recent test_size fraction:

train, test = jf.train_test_split(returns, test_size=0.25)

result = jf.maximum_sharpe(train)     # fit in-sample
# ... then evaluate `result` weights on `test`

Option chains

load_option_chain fetches and normalizes an option chain from Yahoo Finance (also the data extra) into a tidy frame with columns [type, strike, expiry, last, bid, ask, volume, open_interest, implied_vol]:

chain = jf.load_option_chain("AAPL")          # nearest expiry
chain = jf.load_option_chain("AAPL", expiry="2026-01-16")

Covariance estimators

The default moment estimator is the sample covariance, but the toolkit exposes robust alternatives you can pass anywhere moments are computed:

Estimator	Use when
sample_covariance	plenty of data relative to assets (default)
ewma_covariance	recent observations should weigh more (halflife)
ledoit_wolf_covariance	\(N\) is large vs. \(T\); shrinks toward scaled identity

from jaxfolio import toolkit as tk

mat, _ = tk.as_matrix(returns)
cov_ewma = tk.ewma_covariance(mat, halflife=63.0)
cov_lw, intensity = tk.ledoit_wolf_covariance(mat)   # also returns the shrinkage λ