kurry/institutional-holdings-13f-quarterly

US Institutional 13F Holdings Record-level Embeddings Dataset (Merged Quarterly Snapshots 1980–2024)

This dataset provides a single merged Hugging Face DatasetDict combining all 179 quarterly snapshots of U.S. institutional 13F filings (1980 Q1 → 2024 Q3).

Every quarter’s folder has been concatenated into three splits:

• —holdings (113,114,724 rows): record-level positions with fields: • mgrno (string) – Institutional manager ID (SEC) • permco (string) – Permanent company identifier • fdate (timestamp[s]) – Quarter-end report date • shares (float64) – Shares held • price (float64) – Price per share on fdate • dollar_holding (float64) – Market value (shares × price)
• —asset_embeddings (919,599 rows): PERMCO-level 32-dimensional vectors • permco (string) – PERMCO identifier • embedding (float32[32]) – Asset embedding vector
• —investor_embeddings (445,036 rows): manager-level 32-dimensional vectors • mgrno (string) – Institutional manager ID • embedding (float32[32]) – Investor embedding vector

Quick Load Example:

from datasets import load_dataset

ds = load_dataset("kurry/institutional-holdings-13f-quarterly")
print({s: ds[s].num_rows for s in ds})  # {'holdings': 113114724, 'asset_embeddings': 919599, 'investor_embeddings': 445036}
print(ds["holdings"][0])