HaiwenWang/multimodal_stock_data

Multimodal Stock Data

中文: README.zh-CN.md

This is a cleaned raw-data entry point for internal U.S. multimodal stock forecasting work. It provides curated raw data files with semantic paths for downstream multimodal stock forecasting experiments.

Layout

multimodal_stock_data_v1/
  data/
    market/crsp/                       CRSP prices, adjusted returns, security history, market/index data
    fundamentals/                      Compustat annual and quarterly fundamentals
    analyst_estimates/                 IBES estimates, actuals, recommendations
    events_filings/                    Key Developments, SEC filing metadata/text, Insiders by WRDS
    structured_news/                   RavenPack Dow Jones structured news/events/sentiment
    factors/                           Fama-French factors
    link_tables/                       CRSP-Compustat link
    news/benzinga/                     Benzinga monthly gzip JSONL news
    news/fnspid_clean/                 cleaned FNSPID full-article news
    universe/                          survivorship-free stock universe and ticker history
  docs/schema_catalog.zh-CN.md          Chinese parquet schema catalog
  docs/schema_catalog.md                English parquet schema catalog

Why Symlinks

The view avoids duplicating tens of GB while presenting a clean, final data layout. To materialize the view for transfer, use:

rsync -aL multimodal_stock_data_v1/ target_dir/

Modules

Directory	Source	Format	Content	Key time/join fields
data/market/crsp	WRDS CRSP	Parquet	daily OHLCV, adjusted returns, security history, delisting, distributions, market index, S&P 500 membership	permno, dlycaldt, ticker, namedt/nameendt
data/fundamentals	Compustat	Parquet	annual and quarterly fundamentals	gvkey, datadate, rdq
data/link_tables	CRSP/CCM	Parquet	historical CRSP permno to Compustat gvkey links	lpermno, gvkey, linkdt, linkenddt
data/factors	Fama-French	Parquet	daily 3-factor/5-factor and risk-free returns	date
data/analyst_estimates	IBES	Parquet	EPS estimates, revisions, actuals, recommendations	ticker, cusip, statpers, actdats, anndats, anntims
data/events_filings/capital_iq_key_developments	Capital IQ Key Developments	Parquet	structured corporate events	gvkey, announceddateutc, eventtype
data/events_filings/sec_filings	WRDS SEC	Parquet	SEC filing metadata and 8-K items	cik, fdate, form
data/events_filings/sec_filing_text	external SharePoint SEC Markdown	Markdown inside Zip	SEC filing body text, complementing WRDS SEC metadata tables	filename ticker, filing_date, form_type; should be mapped to cik/accession
data/events_filings/insider_filings	Insiders by WRDS	Parquet	Form 3/4/5 documents, transactions, reporting owners, footnotes, signatures	accession, issuercik, issuertradingsymbol, fdate, transactiondate
data/structured_news/ravenpack_dow_jones	RavenPack Dow Jones	Parquet	structured news, events, sentiment, entity mappings	rp_story_id, rp_entity_id, timestamp_utc
data/news/benzinga	Benzinga API	gzip JSONL	Benzinga article text/title/ticker/channel/author	benzinga_id, published, tickers
data/news/fnspid_clean	cleaned FNSPID subset	Parquet	full-article news and article-ticker bridge	article_id, day, ticker
data/universe	local CRSP build	Parquet/CSV	security universe, historical memberships, ticker history, delisting status	permno, ticker, from_date/thru_date, namedt/nameendt

Scale

Module	Scale
CRSP market/index	54,006,627 parquet rows
WRDS fundamentals/analyst/events/filings/structured news	445,127,011 parquet rows
Benzinga news	2,828,464 articles in 207 gzip JSONL files
FNSPID clean	1,180,425 articles + 2,488,568 article-ticker pairs
Universe	10,187 permnos, 155,890 membership rows, 40,125 ticker-history rows
SEC filing text	286,290 Markdown filings, about 6.1 GB compressed, 2020-04-24 to 2026-06-04

Schema

See:

docs/schema_catalog.md

It lists every logical parquet table, row count, columns, and common field meanings.

Usage Notes

1. Build a point-in-time feature store before creating model windows.
2. Use permno + trading date as the main training key.
3. Join news tickers, IBES tickers, SEC CIKs, and Compustat GVKEYs through historical mappings rather than static ticker joins.
4. Use announcement/publication/filing timestamps for all event, text, analyst, and fundamental data to avoid look-ahead bias.
5. Use data/universe for historical membership control and survivorship-bias mitigation.

Universe Versus Source Tables

data/universe is not a regular feature table. It defines the modeling sample space and identifier history: which securities should be eligible on which dates. WRDS, Benzinga, FNSPID, RavenPack, and filing/text directories are feature and event sources: what information was available for a security at a point in time.

Use them together during feature-store construction, but keep them conceptually separate. A practical workflow is to build a permno + trading_date panel from universe, then point-in-time join market, fundamentals, analyst, news, event, and filing data onto that panel.

Source Traceability

Each major curated path keeps the source name in either the directory path or the module name: crsp, compustat, ibes, capital_iq_key_developments, ravenpack_dow_jones, benzinga, fnspid_clean, sec_filing_text, and related names. The source for each module is documented in the module table and schema catalog; directory names preserve source names such as crsp, benzinga, fnspid_clean, and ravenpack_dow_jones.