Lib Tushare

tushare

The cheapest route to real point-in-time A-share fundamentals — and the owner of the most dangerous adjustment implementation in the Chinese ecosystem, which is worse than ordinary qfq.

Access is token + 积分 (points). There are no client-side permission checks at all.

The trap that costs you money

pro_bar(adj="qfq") is anchored to YOUR end_date, not to today. Verified in tushare/pro/data_pro.py:

if adj == 'hfq': data[col] = data[col] * data['adj_factor']
if adj == 'qfq': data[col] = data[col] * data['adj_factor'] / float(fcts['adj_factor'][0])

fcts arrives newest-first, so [0] is the factor on the last date of your query window. The same bar returns different prices depending on what end_date you asked for. Two overlapping queries disagree; concatenating them produces a series with a silent discontinuity at the seam. Ordinary qfq is merely non-reproducible over time — this is non-reproducible over query parameters.

In the same function, pre_close = close.shift(-1) is correct only because rows arrive descending. Sort the frame ascending before letting tushare compute and pre_close, change and pct_chg all break silently — plausible numbers, off by one row, wrong sign of correction. adj_factor is also bfill()-ed, so a missing factor inherits a later one.

Rule: adj=None (raw) or adj="hfq". Never adj="qfq". Leave the rows descending.

Your token travels in plaintext HTTP

Verified in the 1.4.29 source: __http_url = 'http://api.waditu.com/dataapi'. No HTTPS option ships. Token and returned data are both readable on any hop. Treat the token as compromised-by-default: never reuse it as a password elsewhere, and do not run it over untrusted networks.

The client also validates nothing. DataApi.__getattr__ returns partial(self.query, name), so any api_name is accepted client-side. A typo, a renamed endpoint and an endpoint you lack 积分 for are indistinguishable until the server answers. There is no capability map — wrap every call and inspect the server error.

stock_basic is survivorship-biased by default

stock_basic defaults to list_status='L' — live names only. Omit the argument and your universe is silently survivorship-biased. Worse, delisting in China went from rare to common after the 2020 退市新规, so the bias is time-varying, which is harder to reason about than a constant one. Issue a second call with list_status='D' and union the results.

tushare also gives you no high_limit/low_limit and no per-bar 停牌 flag. You cannot refuse a limit-up fill from tushare data alone — cross-reference akshare's limit pools or a licensed vendor.

What it is genuinely best at: point-in-time fundamentals, cheaply

Every filing carries 报告期 end_date (the period) and 公告日 ann_date / f_ann_date (when it was published). Annual reports land up to four months after period end, so joining on end_date leaks up to four months of future information — on its own enough to turn a worthless value factor into a spectacular backtest.

• f_ann_date is what makes point-in-time possible — the actual announcement date of that version of the filing. Filter on it.
• update_flag distinguishes an original report from a later restatement. Keep the version whose f_ann_date <= as_of, not the latest one.
• The 业绩预告 → 业绩快报 → 正式财报 → restatement chain means one quarter has several rows.

积分 thresholds are verified: 120 = 日线 · 1000 = 港股日线 · 2000 = most 财务/基金/期货/指数 · 3000 = 限售股解禁 · 5000+ = fund portfolios, option daily, higher frequency. The per-minute call-rate table and the RMB prices circulating online are secondhand — do not treat them as fact.

Minimal correct call

import pandas as pd, tushare as ts
pro = ts.pro_api("TOKEN")            # token is sent over PLAINTEXT HTTP

# universe: BOTH halves, or you are survivorship-biased
live = pro.stock_basic(list_status="L", fields="ts_code,name,list_date,delist_date")
dead = pro.stock_basic(list_status="D", fields="ts_code,name,list_date,delist_date")
universe = pd.concat([live, dead], ignore_index=True)

# prices: raw or hfq, never qfq; leave rows DESCENDING
bars = ts.pro_bar(ts_code="000001.SZ", adj="hfq",
                  start_date="20200101", end_date="20260901")

# point-in-time fundamentals: filter on the publication date, not 报告期
inc = pro.income(ts_code="000001.SZ",
                 fields="ts_code,ann_date,f_ann_date,end_date,update_flag,revenue")
pit = inc[inc["f_ann_date"] <= as_of]
pit = pit.sort_values("f_ann_date").groupby("end_date").tail(1)   # latest version KNOWN at as_of

See also

• ../../../fin-china/skills/china-ashare-data/SKILL.md — 复权, 涨跌停, T+1, 停牌, 公告日 vs 报告期
• ../../../fin-china/skills/china-ashare-data/references/tushare.md — the verified reference card
• ../../../fin-china/skills/china-ashare-data/references/_source-matrix.md — capability matrix
• ../lib-akshare/SKILL.md — widest free coverage, limit pools, but no PIT

Where this sits

This file is the deep dive on one library and assumes the choice is already made. For which library to pick, how it compares with the alternatives, and the traps that span several of them, the entry point is the domain skill china-ashare-data (../../../fin-china/skills/china-ashare-data/SKILL.md).