Changahou/MATM

A-share daily quotes, screened panel for the MambaATM rework

Daily quotes for Chinese A-share stocks, retrieved through the public baostock interface and screened for price-series integrity. The panel was built for a study of whether a hybrid of causal attention and a selective state-space model can predict the daily direction of a stock, and it is published so that the study's negative result can be checked and reused.

This dataset replaces an earlier upload of the same name. The earlier version held four unnamed stocks with 2,435 rows each, of which 248 of 258 feature columns were constant, and its quoted return column was uncorrelated with its own price columns. It is withdrawn and must not be used. CORRECTION.md records what was wrong with it and how each defect was measured.

Contents

config	stocks	rows	what it is
adj	232	610,988	the screened panel the study uses
adj_unscreened	240	628,554	the same panel before screening, so the screen can be re-run

One file per stock, named by exchange code, for example 000001.SZ.csv. Both directories carry the same ten columns, in this order, and the files are UTF-8 with a byte-order mark.

column	meaning
日期	trading date
当日开盘	open
收盘价	close, forward adjusted for splits and distributions
最高价	high
最低价	low
成交量	volume, in shares
成交额	turnover value, in yuan
涨跌幅	quoted close-to-close return
换手率	turnover rate
昨收	previous close

Prices are forward adjusted, so a return computed from two consecutive rows is the return an investor would have received. The eight columns used as features are open, close, high, low, volume, turnover value, turnover rate and previous close. The label in the study is the sign of the next trading day's close-to-close return.

How the panel was screened

The adjusted series of some stocks contains daily moves that A-shares cannot produce, either where a stock was suspended and resumed into the delisting process so that the price limit does not span the suspension, or where the adjustment factor breaks mid-series. Without a corporate-action database the two cannot be told apart, and both leave a meaningless label, so the rule is the simple one in screen_panel.py: a stock whose adjusted series contains any daily move beyond 25 per cent is excluded in full.

Excluding the whole stock rather than the offending row is deliberate. Dropping an interior row leaves the next return spanning a two-day gap, which manufactures a large spurious return, and that is exactly the defect that made the earlier upload unusable. A whole-stock screen keeps every remaining series continuous. Eight stocks were dropped for this reason, which is why adj has 232 files and adj_unscreened has 240.

Screening also drops the first row of each file, which has no previous close, and recomputes the previous close and the quoted return from the adjusted close column so that every column agrees with the price column.

Caveats

Three things to know before reusing this data.

Suspension days are still present. A suspension day is a row with zero volume. There are 22,798 of them, 3.7 per cent of the rows, and they are removed inside the modelling pipeline rather than in the files, so that the files remain exactly what was retrieved and screened. Their quoted return is zero by construction, so leaving them in would both add untradable observations and push the label distribution toward "down".

The screen is survivorship-biased. Stocks whose adjusted series contain an impossible move are removed, and stocks that delisted during the sample are absent. That flatters every long-only strategy including buy and hold, and the study says so in its conclusion.

The rows do not all carry a usable label. Of the 610,988 rows, 232 are the last row of a stock and have no next-day return, 232 are the first row and have no quoted return, and 25 rows are both a first row and a suspension day. 23,237 distinct rows are therefore removed before modelling and 587,751 usable stock-days remain.

Usage

import pandas as pd

d = pd.read_csv("000001.SZ.csv", encoding="utf-8-sig")

from datasets import load_dataset

ds = load_dataset("Changahou/MATM", "adj")

Code and the study

The code that builds the panel, screens it, trains the models and produces every table and figure is at https://github.com/ChenJiahaoJNU/Mamba_ATM, together with the manuscript and a verification record in which every check names the command that produced it. screen_panel.py is in code/panel/. Running it on adj_unscreened reproduces adj.

Licence

MIT for the files and the code in this repository. The quotes themselves come from the public baostock interface and carry whatever terms that source imposes. Nothing here is investment advice.