Datasetpublic

TriadParty/jevm-news

jevm news screening dataset Chinese financial news flash, each item labelled by how the A-share market actually reacted to it. This is the news line training set of jevm: one row per news item, with every feature computed strictly as of the publish time τ. The labels come only from minute-bar price and turnover reactions and from objective propagation evidence. No labels come from model scores. 中文说明见下方 At a glance Split Period Rows Head-B labelled 30-min… See the full description on the dataset page: https://huggingface.co/datasets/TriadParty/jevm-news.

sourceHugging Faceotherupdated 2d agoView on Hugging Face

0likes46downloads

Dataset Card

jevm news screening dataset

Chinese financial news flash, each item labelled by how the A-share market actually reacted to it. This is the news line training set of jevm: one row per news item, with every feature computed strictly as of the publish time τ. The labels come only from minute-bar price and turnover reactions and from objective propagation evidence. No labels come from model scores.

中文说明见下方

At a glance

Split	Period	Rows	Head-B labelled	30-min "major" rate (labelled rows)
`train`	2025-02-01 → 2026-04-24	207,275	22,563	1.9%
`gap`	2026-04-25 → 04-30 (embargo, do not train)	2,935	420	2.1%
`calib`	2026-05-01 → 06-30	24,121	3,465	2.1%
`holdout`	2026-07-01 → 2026-09-23	48,443	7,143	2.1%
total		282,774	33,591

Sources: Jin10 (138.7k), CLS (60.0k), Wallstreetcn (57.5k), 36Kr (16.0k), Eastmoney (5.8k), Jiemian (3.9k) and a few others. The news line holds items that do not name a listed company in the title or lead. Items that do are in the separate company line, which is not included here. Community posts are excluded from this release.

The splits are chronological. The intended base model (Qwen3.8-27B) has a knowledge cutoff of 2026-05-01, so only calib and holdout measure out-of-knowledge performance.

Labels

Column	Meaning
`yB_q30`	Head B, primary target. Quantile q of the related-stock basket's 30-minute reaction after τ. Basket statistic = mean of the top-m \	SAR\	(m = min(3, basket size)); q = min(q vs 500 random same-minute, same-industry baskets, q vs the basket's own normal). NaN = no basket (missing label, not a negative).
`yB_sig30`	`yB_q30 ≥ 0.99` ("major")
`yB_q1`, `yB_sig1`	same, τ → T+1 close (auxiliary)
`yB_vol`	turnover burst: quantile of log(window turnover / 20-day same-slot mean) vs size- and industry-matched random baskets
`yB_src`	basket source: `newsbk` (light matcher on the raw text, ≤ 3 stocks) or `none`
`yA_src24`	Head A. Distinct sources that carried the same information within 24 h (cosine ≥ 0.90)
`yA_n24`, `yA_follow7`	reprint count within 24 h; same-topic follow-ups within 7 days (0.78 ≤ cos < 0.90)

SAR = (r_stock − median r_industry) / σ, where σ is the stock's own AR dispersion over the previous 20 days in the same slot. Grades: q ≥ 0.80 / 0.95 / 0.99 → normal / significant / major. The labels passed placebo tests (same basket, same clock time, τ shifted back 5/10 trading days). See docs/labels.md and docs/findings.md.

Features (all as of τ; daily statistics are taken from the previous trading day)

Column	Meaning
`text`	first 600 characters of the item
`f_session`, `f_hour`, `f_weekday`	pre-market / intraday / after-hours / non-trading day; hour; weekday
`f_source`, `f_len`, `f_platform_imp`	source; text length; platform importance flag (0 / 1 important / 2 highlighted)
`f_prior_n48`	same-information items in the 48 h before τ (novelty)
`f_mkt_*`	previous-day breadth: share above MA20, median change, advance ratio, 60-day highs − lows, limit-up/down counts, `rot10` sector-rotation index; `f_mkt_today` = market median return from the open to τ
`f_n_named`, `f_named_*`	named-company count; return today-to-τ, 5-day return and room to limit-up of the first named company
`f_rep_*`	rolling source reputation: major-reaction rate and mean reprint sources of the same source over [τ−90 d, τ−3 d]

Evaluation-only columns (do not use as features, since they do not exist at inference time): grp (P1/P2/P3/AMB: whether and how the existing LLM pipeline extracted the item), event1_id, invest_score (the existing LLM importance score, i.e. baseline B0). weight = 4 for down-sampled unextracted items and 1 otherwise.

Usage

python

from datasets import load_dataset
ds = load_dataset("TriadParty/jevm-news")
train = ds["train"].to_pandas()
lab = train[train.yB_q30.notna()]          # head B: train/evaluate only on labelled rows

The model-input renderer (prompting.render) and the trainers are in the GitHub repository.

License and terms

—The news text belongs to its original publishers. It is redistributed here in truncated form for non-commercial research only. Do not republish the text or use it commercially. Rights holders can request removal by opening a discussion on this repository.
—The labels and derived features are released for research use. They are derived from licensed market data, and the underlying price and turnover series are not included.
—Provided as is, without warranty. This dataset is not investment advice.

Citation

bibtex

@misc{jevm2026,
  title  = {jevm: market-labelled news screening for China A-shares},
  author = {zzlgreat},
  year   = {2026},
  url    = {https://github.com/zzlgreat/jevm}
}

中文说明

A 股财经快讯数据集。每条新闻的标签由市场的真实反应决定，所有特征都严格截止到源站发布时间 τ。这是 jevm 的新闻线训练集，只收未在标题 / 导语中点名上市公司的新闻；点名公司的公司线不在本次发布之内。本版本不含社群帖子类来源。

—头 B（主目标 yB_q30）：关联股票篮子在发布后 30 分钟的反应强度分位，对照为同一时刻、同行业构成的 500 个随机篮子和篮子自身常态，两个分位取较小者。NaN 表示没有篮子，是缺标签，不是负例。
—头 A（yA_src24）：24 小时内转载同一信息的来源数。
—切分按时间：train 到 2026-04-24，calib 为 2026-05 ~ 06，holdout 从 2026-07 起；gap 是隔离带，不要用于训练。
—grp、event1_id、invest_score 只用于评测，不能当特征（上线时这些信息还不存在）。
—新闻正文版权归原发布平台所有，仅供非商业研究使用，请勿转载或商用；权利方可在本仓库的 Discussions 中申请删除。