etfbench/etfbench_submission

Dataset bundle

This directory holds JSONL files (one JSON object per line) for multi-market, daily-frequency research: public information condensed into structured text summaries, plus a small standardized OHLCV sample for testing loaders, backtest scaffolding, or schema alignment.

For academic research and offline simulation only. Not financial advice.

All shards use UTF-8 encoding; calendar fields are YYYY-MM-DD strings.

us_news.jsonl and cn_news.jsonl

• —Role: Day-aligned “news window → text signal” records, intended to pair with daily returns, positions, or paper-trading logic.
• —Coverage: info_date runs from 2024-01-01 through 2025-12-31 (731 consecutive calendar days in each file, including leap year 2024).
• —Fields per line:
• —info_date: calendar date the summary refers to.
• —market_view: a single narrative paragraph (length varies with model and prompt).
• —key_points: array of strings with bullet-style takeaways.

Caveat: Text is produced by language models (or similar) summarizing public sources; omissions, lag, or factual errors are possible. Do not treat this as live quotes, execution data, or regulated research.

market_data_examples.jsonl

• —Role: Standalone, easy-to-parse daily bars so you can validate pipelines without wiring a full market database first. Independent of the news shards above.
• —Coverage: trade_date spans 2025-01-02–2025-01-08 on actual trading days only (a one-calendar-week window; weekends have no rows).
• —Fields per line:
• —pool_name: universe / bucket label (e.g. theme or market slice).
• —market: market tag (e.g. cn, us).
• —symbol: ticker / listing code (e.g. ETF code).
• —trade_date: session date.
• —open, high, low, close: floats.
• —volume: integer.

Caveat: Values come from offline, public-style market feeds used for method demos; coverage is not equivalent to a complete licensed history—use your own vendor data for production statistics.

Usage notes

1. 1.Read JSONL with one json.loads per line; stream for large files.
2. 2.When joining info_date to trade_date, define timezone and session rules yourself for cross-market work.
3. 3.For redistributions or derivatives, comply with Hugging Face and any upstream data terms; this README does not restate third-party licenses.

On the Hugging Face Hub, this repo declares three dataset configs (us_news, cn_news, market_data) so the viewer does not merge incompatible JSONL schemas. Load locally with:

from datasets import load_dataset
ds_us = load_dataset("<your_org>/<your_repo>", name="us_news")
ds_cn = load_dataset("<your_org>/<your_repo>", name="cn_news")
ds_px = load_dataset("<your_org>/<your_repo>", name="market_data")

Replace <your_org>/<your_repo> with your dataset id.