jck3114/ashare-book-benchmark-train

China A-share Book Benchmark: Train

Training data for a benchmark where participants submit batches of JAX factor candidates and maintain their own evolving book. The benchmark performs fitting and admission privately. This repository contains data, not evaluator source or private scoring rules.

Coverage

• Train only: 2012-01-04 through 2017-11-01, inclusive; 1,416 trading days.
• Minute grid: 240 observations per trading day, from 09:31 through 15:00, with the exchange lunch break. Times are local China exchange times.
• One Parquet shard per trading day in each of the two configurations below.
• No Validation 1, Validation 2, Holdout/Test data, participant submissions, account information, evaluator configurations or scoring implementations.

Tables

Configuration	Keys	Values
bars	date, minute, integer sid	float32 open, close, high, low, qty, amt
labels	signal date, integer sid	float32 weight, next_return_30m, next_return_1d, next_return_2d, next_return_5d

sid is a stock identifier. Treat it as a categorical key, not a continuous explanatory variable. date is a Parquet date32 value; minute is an HH:MM string. Rows with no finite feature in any of the six fields are omitted from bars. Restore missing cells as NaN when constructing dense arrays; do not automatically forward-fill across missing stocks or trading days. labels includes rows with positive finite weights, including observations whose minute bars may be missing. The final labels shard is intentionally empty.

Label alignment and boundary protection

A signal uses information available through 15:00 on trading day t. Its label anchor and weight refer to 10:00 on the next trading day. The target suffix gives the forward return horizon from that anchor. All alignment uses trading days.

Any label whose forward horizon reaches outside Train is null. The conservative tail masks are 1, 2, 3 and 6 trading days for the 30m, 1d, 2d and 5d targets, respectively. The last trading day's weights are withheld because they depend on the following day. See manifest.json for the exact export contract.

Boundary-purged public labels can differ from the private evaluator's original unpurged Train labels. Local calculations on this release should account for that difference. No private performance metrics are provided by this dataset.

Loading

Install the Hugging Face datasets package, then use streaming to avoid loading the entire dataset into RAM:

from datasets import load_dataset

repo_id = "jck3114/ashare-book-benchmark-train"  # replace with this repository's ID
bars = load_dataset(repo_id, "bars", split="train", streaming=True)
labels = load_dataset(repo_id, "labels", split="train", streaming=True)
print(next(iter(bars)))

Use the same sid ordering for every day when creating dense factor inputs. A missing stock on a particular day remains missing; it does not shift the stock axis. The benchmark accepts batched candidates and returns approve/deny feedback with each factor's available Validation correlation and autocorrelation. Internal factor-return, Sharpe and Holdout numerical results are not part of the submission API.

Integrity and provenance

This release was exported from the dataset owner's C03 A-share minute panel. The source stock-name mapping and upstream source files are not distributed here. manifest.json lists the approved field/target names, Train date range, boundary masks, byte size and SHA-256 checksum of every Parquet file. Uploads are restricted to this Train export; the original full panel is never uploaded.

No redistribution license has been specified by the publisher. This dataset card does not assign or infer a license for the underlying data.