A-Share Flow and Ownership Features

中文页面

This page defines the integration boundary for A-share money flow, institutional holdings, and shareholder-structure data. The platform retrieves, cleans, applies point-in-time (PIT) availability rules, and publishes assets. Research repositories consume the published feature Parquet files read-only.

Provider permissions

As of 2026-06-16, the TuShare points table states that accounts above 5,000 points receive 500 requests per minute for standard data with no standard-data total limit. Above 10,000 points, the standard-data rate is unchanged and access to specialty data is added at 300 requests per minute. Above 15,000 points, specialty data has no total limit. Minute, news, announcements, and HK/US market access are separate permissions and are not included automatically.

TUSHARE_TOKEN_2 is currently used as a 15,000-point account. If that token requires a proxy domain, configure TUSHARE_API_URL_2 or pass --api-url. This changes the TuShare SDK client's API URL; it does not configure the machine's HTTP proxy.

Candidate data families

Family	TuShare interfaces	Candidate use	Limitation
Equity money flow	moneyflow, moneyflow_dc, moneyflow_ths	Rolling standardized net inflow for large and extra-large orders	Vendor-defined measures, not regulator-verified institutional accounts
Cross-market flows	moneyflow_hsgt	Northbound/southbound state and market context	Better suited to market- or industry-level features
Public-fund holdings	fund_portfolio	Holdings, market value, and changes	Must be PIT by disclosure date; default availability is the next trading day
Dragon-Tiger institutional details	top_inst	Sparse event features	Covers only listed stocks, not all institutional trading
Shareholder structure	top10_holders, top10_floatholders	Concentration and institutional-holder measures	Slow updates; more suitable for low-frequency cross-sections
Shareholder transactions	stk_holdertrade	Major-holder increase/decrease events	Sparse events; missing values must not filter samples directly

Asset boundaries and available pipelines

Keep these features outside daily_clean. The repository uses separate raw and derived assets, including flow_ownership_features, fund_portfolio_features, holder_structure_features, top_inst_events, holdertrade_events, and hsgt_market_features. Exact asset paths and CLI invocations are maintained in the Chinese companion; machine-readable options and output keys are unchanged between locales.

The money-flow pipeline mirrors raw moneyflow, then builds and validates derived features. Current generated measures include mf_net_amount_*d_to_amount, mf_elg_net_amount_*d_to_amount, mf_lg_net_amount_*d_to_amount, mf_net_amount_20d_to_float_mv, mf_buy_sell_imbalance_20d, and 20-day cross-sectional ranks/z-scores. When --industry-dir is provided, it also builds mf_net_amount_20d_industry_zscore. daily.amount is in thousands of yuan; the builder divides it by 10 to match the ten-thousand-yuan unit used by TuShare money-flow amount fields.

The public-fund pipeline mirrors raw fund_portfolio by report period, then publishes PIT features using the first eligible trading day after disclosure. It includes holding-fund counts, holding market value and amount, ratios to total/float market value and share count, and related quarter-over-quarter changes. The builder maintains each fund's latest disclosed state and emits zero rows for stocks no longer held in the next disclosure, preventing downstream forward fills from retaining stale positions.

Separate raw assets also feed shareholder-structure, top-institution, and shareholder-transaction event builders. Use each builder's matching validate-a-share-* command and the thresholds recorded in the companion before treating an output as a candidate asset.

PIT and feature rules

• Same-day money-flow data may be used for a signal formed after the close and traded on the next session. Intraday-trading assumptions require at least a one-day lag.
• Fund holdings, shareholder structure, and holder transactions derive available_date from ann_date or disclosure_date. Do not align historical observations using report_period alone.
• Normalize money-flow amounts by turnover, float market value, or total market value before calculating rolling windows, cross-sectional ranks, or z-scores.
• Dragon-Tiger and shareholder-transaction data are sparse events. Preserve missingness instead of applying complete-case filtering.
• Before promotion to a production research candidate, evaluate coverage, missingness, PIT leakage, single-feature IC, correlation with existing technical features, benchmark ladder, CPCV, turnover, and cost drag.

Suggested research sequence

1. Build PIT features from fund_portfolio.
2. Add rolling standardized money-flow features from moneyflow or moneyflow_dc.
3. Add shareholder-structure features from top10_holders and top10_floatholders.
4. Add sparse event features from top_inst and stk_holdertrade.

For short-term incremental validation, a research repository can load external feature Parquet through fundamentals.source=file. Longer term, these data should use a separate panel_features or alternative_features configuration entry so money-flow and ownership features are not represented as fundamentals.