Narrowed scope — datasets that are genuinely cross-asset or not tied to a single asset class live here. SEC filings and 13F now live under data/equities/{fundamentals,positioning}/; CFTC Commitment of Traders under data/futures/positioning/; on-chain (DefiLlama, CoinGecko) under data/crypto/onchain/.

Directory

Subdir	Dataset	Download script	Loader(s)
`news/`	Bloomberg news archive + FNSPID financial headlines	`news/bloomberg_download.py`, `news/fnspid_download.py`	`load_bloomberg_news`, `load_fnspid`
`text/`	Reference / benchmark corpora (Financial Phrasebank)	— (HuggingFace cache)	`load_financial_phrasebank`

All loaders return Polars DataFrames. Each subdirectory has its own README with download instructions and storage footprint.

Quick Start

# FNSPID news (HuggingFace, ~1 GB)
uv run python data/alternative/news/fnspid_download.py

# Bloomberg news archive (HuggingFace mirror)
uv run python data/alternative/news/bloomberg_download.py

From Python

from data import load_bloomberg_news, load_financial_phrasebank, load_fnspid

news = load_fnspid(symbols=["AAPL", "MSFT"], start_date="2020-01-01")
phrasebook = load_financial_phrasebank(agreement="100")

When data is missing, loaders raise DataNotFoundError with a pointer to the relevant subdirectory README and the exact download command.