| --- |
| license: other |
| pretty_name: OpenPM-Bench |
| viewer: false |
| language: |
| - en |
| tags: |
| - finance |
| - portfolio-management |
| - llm-agents |
| - point-in-time |
| - trading |
| - benchmark |
| size_categories: |
| - 1M<n<10M |
| --- |
| |
| # OpenPM-Bench |
|
|
| **Auditable, point-in-time evaluation data for LLM portfolio-management agents.** |
|
|
| OpenPM-Bench is the dataset behind **OpenPM**, a benchmark for the question *"can an LLM |
| manage a stock portfolio?"* A trading agent sees only point-in-time, leakage-controlled |
| information for the S&P 500 — the names that were *actually* in the index on each decision |
| date — and proposes target weights, which are backtested bar-by-bar on a 5-minute price panel. |
|
|
| - **Code:** <https://github.com/aslcai/OpenPM-Bench> |
| - **Paper:** *OpenPM: Auditable Point-in-Time Evaluation for LLM Portfolio-Management Agents* |
|
|
| ## What's in it |
|
|
| A single content-hashed freeze (`FREEZE_MANIFEST.json`, `dataset_sha256` `89e154ae…`, |
| **1,376 data files, ≈47 GB**) of **508 point-in-time S&P 500 members** active on at least one |
| date between **2026-02-19 and 2026-05-01**. The evaluation window is **2026-03-02 → 2026-05-01** |
| (44 trading days, 3,432 five-minute bars per name). The data is 2026-dated, after the knowledge |
| cutoff of most evaluated model backbones. |
|
|
| Every record carries an availability timestamp (`ts_available`); a field is visible at decision |
| time *T* only if its stamp ≤ *T*. Future-looking labels are held out in a separate, agent-invisible |
| layer. |
|
|
| ### Layers |
|
|
| | Path | Contents | |
| |---|---| |
| | `bars/` | 5-minute IEX-derived bar/quote feature parquets | |
| | `quotes/` | spread / liquidity features | |
| | `panel/` | `price_panel_5m.parquet` — the bid/ask/mid panel the backtester marks against | |
| | `feature_output/` | `feature_output.ndjson` — the point-in-time, label-free agent input stream | |
| | `labels/` | held-out evaluation labels (step / earnings / forward / executable returns) | |
| | `events/` | SEC event state (incl. PEAD), GDELT news state, earnings | |
| | `context/` | rates / FX / commodities macro context | |
| | `instruments/` | S&P 500 membership, sector history, corporate actions | |
| | `raw/` | raw provenance archives (IEX TOPS, SEC filings, FRED, GDELT, Wikipedia revisions) | |
|
|
| ## How much do you need? |
|
|
| The full release is **≈47 GB, but ~39 GB of that is `raw/`** — provenance archives needed **only to |
| rebuild the dataset from source.** **To run the full benchmark, download everything except `raw/` |
| (≈8 GB):** |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| |
| # Run the benchmark (~8 GB) — all derived layers, skip the raw provenance archives: |
| path = snapshot_download("aslcai/OpenPM-Bench", repo_type="dataset", ignore_patterns=["raw/*"]) |
| |
| # Full release (~47 GB, incl. raw/ to rebuild from source): |
| # path = snapshot_download("aslcai/OpenPM-Bench", repo_type="dataset") |
| ``` |
|
|
| This is a multi-layer parquet/NDJSON dataset, not a single table — read the layers directly with |
| pandas/pyarrow, or use the OpenPM code (see the repo) to run an agent. Download all derived layers, |
| not a subset: the flagship agent's `filings_meta` / `filings_body` analysts read `events/`, and a |
| missing layer silently degrades those channels rather than erroring. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{openpm2026, |
| title = {OpenPM: Auditable Point-in-Time Evaluation for LLM Portfolio-Management Agents}, |
| author = {Cai, Xinying and Guo, Minghao and Li, Raymond and others}, |
| year = {2026}, |
| url = {https://github.com/aslcai/OpenPM-Bench} |
| } |
| ``` |
|
|
