kuznets

Remote data access for economic and financial data — FRED, OECD, Eurostat, World Bank, Fama-French, Yahoo Finance, and more — returned as pandas, polars, pyarrow, or dask frames.

kuznets is a fork of pydata/pandas-datareader.

Installation

Install using pip

pip install kuznets

Usage

import kuznets as kz
kz.get_data_fred('GS10')

Every reader accepts an output_type argument to return polars, pyarrow, or dask frames instead of pandas, and multi-symbol daily reads fetch concurrently (tune with max_workers):

kz.DataReader(["AAPL", "MSFT"], "stooq", output_type="polars", max_workers=8)

Documentation

Stable documentation is available on github.io. A second copy of the stable documentation is hosted on read the docs for more details.

Development documentation is available for the latest changes in master.

Requirements

Using kuznets requires the following packages:

• pandas>=3.0
• narwhals>=2.0
• lxml
• requests

Non-pandas output backends are optional extras:

pip install kuznets[polars]    # or [pyarrow], [dask], [backends]

Building the documentation additionally requires:

• matplotlib
• ipython
• requests_cache
• sphinx
• pydata_sphinx_theme

Development and testing dependencies are defined in pyproject.toml: the dev extra for pip workflows, and the pixi test/lint features for the pixi workspace.

Install latest development version

python -m pip install git+https://github.com/jessegrabowski/kuznets.git

or

git clone https://github.com/jessegrabowski/kuznets.git
cd kuznets
python setup.py install

Development environment

The repo is a pixi workspace: pixi install creates the dev environment (editable install, test and lint tools), pixi run test runs the suite, and pixi run -e py314 test runs it on another Python. pip install -e ".[dev]" in a virtualenv works too.

Running the tests

The test suite runs fully offline by default: each reader is replayed against a recorded real response stored under tests/data/, so no third-party service is contacted.

pytest

Tests marked network are deselected by default. Each one runs a reader against the live service and asserts the shape of the result (columns, dtypes, non-empty) — an API-drift detector that skips gracefully when a service is unreachable:

pytest -m network

The offline fixtures are recordings, not hand-written mocks. To regenerate them from the live services, run the network tests in record mode:

RECORD=1 pytest -m network

This rewrites tests/data/ from the real responses (and a wrong parser fails here, at record time, rather than being hidden by a fixture built to match it). A weekly GitHub Actions workflow (refresh-fixtures) does this automatically and opens a PR when a fixture changes, failing if an upstream shape breaks. A few services can't be recorded and are documented in their tests: Stooq serves an anti-bot challenge to scripted clients, and the keyed readers (AlphaVantage, Quandl, Tiingo) need API keys.