waylonli/FINSABER-V2-Data

FINSABER V2 Data

FINSABER V2 Data is the parquet dataset used by the upgraded FINSABER-2 backtesting framework. It contains S&P 500 daily market data, financial news items, and SEC filing text organized by year for reproducible financial strategy research.

Dataset Structure

The dataset is partitioned by modality and year:

price_daily/year=<YYYY>/part-000.parquet
news_items/year=<YYYY>/part-000.parquet
filingk/year=<YYYY>/part-000.parquet
filingq/year=<YYYY>/part-000.parquet

Current local snapshot summary:

Split	Years	Files	Rows	Size
price_daily	2000-2025	26	4,901,012	0.13 GB
news_items	2009-2025	17	2,908,022	0.46 GB
filingk	2000-2025	26	17,707	2.96 GB
filingq	2000-2025	26	53,377	3.97 GB

Total size is about 7.52 GB across 95 parquet files.

Columns

price_daily:

• date, symbol, cik
• open, high, low, close, adjusted_close
• volume

news_items:

• date, symbol, cik
• item_index, news_text
• text_len, text_crc32, year

filingk and filingq:

• date, symbol, cik
• filing_idx, filing_text
• text_len, year, accession

Usage

Install the FINSABER-2 package:

pip install finsaber

Download this dataset with huggingface_hub:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="waylonli/FINSABER-V2-Data",
    repo_type="dataset",
)
print(local_dir)

Point the FINSABER-2 parquet loader at the downloaded snapshot directory. It should contain the price_daily, news_items, filingk, and filingq folders at its root.

Notes

• Use adjusted prices for split-adjusted price simulation.
• Use raw volume or dollar volume when applying liquidity constraints.
• News and filings should be handled with explicit timing assumptions to avoid look-ahead bias.
• Users are responsible for ensuring their use of source-derived market, news, and filing data complies with applicable licenses and data-provider terms.