ZanderL1337/financial-data-pipeline

Financial Data Pipeline — Full Curated Snapshot

A comprehensive financial dataset covering 237 tables and 152,222,232 rows across macro, market, and alternative data sources.

Data Sources

Category	Tables	Key Sources
Market Prices	18	Tiingo, Schwab, Finnhub, CBOE
Macro & Economic	91	FRED, BLS, BEA, Treasury, EIA, EPA
Fundamentals	82	SEC EDGAR, Finnhub, SimFin, Alpha Vantage
Alternative Data	32	Congressional trades, insider transactions, patents, OpenFDA
Crypto & Forex	8	CoinGecko, Tiingo
Index & Holdings	4	Wikipedia, BlackRock, EdgarTools, OpenFIGI
Sentiment & News	2	Finnhub, Reddit, Google Trends, Fed speeches

Usage

from datasets import load_dataset

# Load entire dataset
ds = load_dataset("ZanderL1337/financial-data-pipeline", trust_remote_code=True)

# Load specific table
df = ds["agriculture"].to_pandas()

Or load individual parquet files directly:

import pandas as pd

df = pd.read_parquet("path/to/parquet/file.parquet")

Schema

Each table is stored as a separate parquet file. Key columns:

• date / snapshot_date: Temporal columns (where applicable)
• symbol / ticker: Security identifiers
• fetched_at: UTC timestamp when data was fetched
• year / month: Hive partition columns (where applicable)

Engineering & data quality

This snapshot is the output of a tested pipeline, not a one-off scrape:

• 761 automated tests, including guard tests that fail the suite if a pipeline is added but not wired into the query catalog, schema registry, and curated-table dedup step — so a table can't silently go missing from downstream queries.
• Schema/null-rate/range validation (validate.py) runs as an operational health check against every table after each pipeline run.
• Raw vs. curated separation: pipelines write Hive-partitioned raw Parquet, which can contain overlapping re-fetches (measured up to ~42% duplicate rows on some raw tables); a dedup step (curated.py) produces the deduplicated tables published here.
• Point-in-time discipline: downstream feature joins use filing/publication date with explicit lags rather than observation date, so a backtest can't accidentally see information before it was actually available. fetched_at records ingestion time separately from the business dates already present in the source data (e.g. dividend tables carry declaration_date/record_date/payment_date/ ex_dividend_date, which are four different points in time for the same event).
• Iceberg pilot: a subset of tables (prices, macro, fundamentals_annual, fundamentals_quarterly) are additionally mirrored into an Apache Iceberg warehouse with real snapshot-based reads and automated snapshot expiration (30-day retention), as a pilot toward full lakehouse-style versioning of the rest of the store.

Full source, tests, and architecture docs: https://github.com/Zanderl1987/financial-data-pipeline

Build Info

• Generated: 2026-10-09
• Pipeline: financial-data-pipeline (https://github.com/Zanderl1987/financial-data-pipeline)
• Tables: 237
• Total Rows: 152,222,232
• Total Size: 5426.9 MB

License

CC BY 4.0 — data sourced from public APIs and government databases.