Determine the target date. If the user does not provide one, use the latest completed A-share trading day. Check get_last_trade_date and get_trade_cal; if the target date is closed, return a short "今日休市" note instead of running the scan.
Collect data in this order:
- Trading calendar and stock universe: get_last_trade_date, get_trade_cal, get_trade_list.
- Daily K-line (120 days for pattern calculation window): get_stock_daily.
- Fund flow per stock: 同花顺 (Tonghuashun/10jqka) stock_fund_flow_individual as primary source. Falls back to AKShare stock_individual_fund_flow_rank (东方财富), then direct East Money push2 API on failure. The flow source is tracked via df.attrs["flow_source"] and persisted in cache metadata.
- Stock basic info (name, industry): get_stock_detail; market cap computed from get_share_float × close price. 排除 ST / *ST / PT / 退市整理期股票（通过 list_status 字段检测，2026年7月 ST 涨跌幅放宽至 ±10% 后波动加剧，形态信号可靠性下降，排除策略暂不调整）。
Run 8 pattern detectors against each stock's K-line data (formulas in references/pattern-formulas.md). Each detector returns a continuous float strength value (> 0 = triggered, ≤ 0 = not triggered). The strength magnitude is used for backtest IC calculation; scoring uses binary weights.
- MA Golden Cross: MA5 crosses above MA20
- MACD Golden Cross: DIF crosses above DEA
- Bullish Alignment: MA5 > MA10 > MA20 > MA60
- Volume Breakout: close at 20-day high + volume > 1.5x 5-day avg volume
- Bollinger Breakout: bandwidth expansion + price breaks upper band
- Hammer: lower shadow >= 2x body + in downtrend
- Morning Star: bear candle → small body → bull candle (3-day reversal)
- RSI Oversold Rebound: RSI(14) < 30 + today closes bullish
Filter by capital flow:
- Main capital net inflow rate > 5%
- Turnover > 50M CNY (exclude illiquid stocks)
- Super-large order net inflow > 0 Note: Tonghuashun does not provide order-size breakdown (super_large/large/medium/small). Super-large and large net inflows are proxied as main_net_inflow / 2 each.
Cross-validate and rank. A stock must pass at least one pattern AND the flow filter. Compute stock_score = neutralized_pattern_score + flow_score + quality_bonus, then take the Top N.
Generate per-stock LLM analysis using references/report-template.md for the analysis prompt. Each stock gets detailed Chinese analysis (technical breakdown + fund verification + sector context + risk notes).
Render the report with data provenance table showing the actual source for each data category. Save to output/YYYY-MM-DD/daily_screener_YYYYMMDD.md and output/YYYY-MM-DD/daily_screener_YYYYMMDD.json unless the user gives another path.
Run scripts/validate_screener.py <md-path> <json-path> after writing the output. Production runs use --strict mode. The validator checks 17 rules across 5 categories (结构完整性 / 数据溯源 / 评分可审计性 / 内容合规 / 数据质量). See references/validate-rules.md for the complete rule checklist and exit codes. Fix any ERROR-level violations before presenting the result.
Periodic backtest & weight calibration. After accumulating ≥60 trading days of fresh K-line data in cache, run scripts/analyze_weights.py --data cache/ --json to compute per-detector Spearman rank IC (1d/3d/5d/10d/20d forward returns). Save the resulting weights to config.json under detector_weights. The pipeline automatically loads calibrated weights via apply_weights_from_config() at startup. Re-run quarterly to keep weights current.

回测方法说明（防止前视偏差）： 所有前向收益从 T+1 开盘价 起算（而非 T 日收盘），排除当日跳空和不可交易时段。例如 5d forward return = (T+6 收盘 − T+1 开盘) / T+1 开盘。这确保了 IC 计算结果反映的是实际可实现的交易收益，而非包含 T 日收盘已锁定的假收益。

生存偏差提示： 回测使用的股票 universe 来自回测日期的 get_trade_list，这意味着已退市股票不会出现在历史扫描中。在长周期回测（>1年）中，生存偏差可能导致 IC 被系统性高估约 0.02–0.05。建议在对比新规前后表现时，使用固定 universe（回测起点当日全 A 股列表）来消除此偏差。

2026-07-06 新规后校准： 主板 ST 涨跌幅由 ±5%→±10%、盘后固定价格交易扩容两项变更后，建议在 2026年10月（积累约 60 个交易日后）运行一次专项回测，对比新规前后的 IC 和权重变化。使用 scripts/analyze_weights.py --since 2026-07-06 --compare-pre-regulation 输出前后对比报告。

Pandadata Reference

Read references/pandadata-map.md when planning calls, selecting fields, or deciding how to degrade if a data interface is unavailable. The map documents which Pandadata endpoint serves each data requirement.

Scoring Formula

pattern_score = sum(triggered_pattern_weights)   # weights from config.json detector_weights or DETECTOR_REGISTRY defaults
flow_score    = min(main_inflow_rate_decimal / 0.05, 3)  # 1 point per 5% inflow rate, capped at 3
quality_bonus:
  +1  if market_cap >= 50B CNY
  +1  if turnover >= 200M CNY

Stage 1 — Industry neutralization (z-score within each industry):
  z_ind = (raw_pattern_score − μ_industry) / σ_industry
  z_ind = max(0, z_ind)   # floor at 0

Stage 2 — Market-cap neutralization (z-score within each market-cap quintile) [v2.1]:
  z_mcap = (z_ind − μ_mcap_quintile) / σ_mcap_quintile
  z_mcap = max(0, z_mcap)   # floor at 0

Blend:
  neutralized = z_ind × 0.6 + z_mcap × 0.4

Final score = neutralized_pattern_score + flow_score + quality_bonus

市场中性化说明 (v2.1)： A 股小市值因子异常显著——小盘股在技术形态检测器中系统性触发率更高（波动大、均线交叉频繁）。单纯行业中性化无法消除这个偏差。v2.1 引入市值五分位 Z-score 作为第二层中性化，60/40 混合权重保留更多行业信号同时修正小盘股优势。

Portfolio Optimization [v2.1]

After scoring and ranking, the pipeline computes three portfolio allocation schemes for the Top-N stocks:

Scheme	Method	Best For
等权 (Equal Weight)	1/N allocation	Benchmark, simplest
最小方差 (Min Variance)	Minimize w'Σw via quadratic programming, long-only	Conservative, low-volatility
风险平价 (Risk Parity)	Equal risk contribution via iterative scaling	Balanced risk exposure

The covariance matrix is estimated from 60-day daily log-returns. Portfolio weights are included in the JSON output and rendered as a comparison table in the Markdown report.

Report Rules

Write in Chinese unless the user requests another language.
Use absolute dates such as 2026-06-30; avoid ambiguous "today" in the final report body.
Each selected stock must show: triggered patterns, capital flow data, and the LLM-generated analysis.
Use cautious language: "值得关注", "可跟踪", "信号偏多"; never say "推荐买入", "目标价", "必涨".
When a data call fails, keep the report useful by skipping that stock (not the whole scan) and adding a missing-data count under "数据说明".
Include scoring breakdown in the JSON output for auditability.
Always include a 数据来源 (data provenance) table showing the actual source and authenticity status for each data category.

Automation

When the user asks for automated daily screening, create an after-close task for trading days only, preferably after 15:45 Asia/Shanghai (15 min after 盘后固定价格交易结束). A scripts/daily_screener.bat entry point for Windows Task Scheduler is provided. Make the task idempotent: if output/YYYY-MM-DD/daily_screener_YYYYMMDD.md already exists, regenerate and overwrite it.

2026-07-06 新规提示： 盘后固定价格交易已扩容至全部 A 股（15:05-15:30），收盘价和数据在 15:30 后才真正锁定。Pandadata get_stock_daily 的 K 线数据是否包含盘后固定价格成交需要在首次运行时验证（检查 close 字段与行情软件收盘价是否一致）。如有差异，在报告「数据说明」节中标注。

Directory Structure

skill-post-market-screener/
├── SKILL.md                         # Agent workflow entry point
├── README.md                        # Human documentation
├── OPERATION_MANUAL.md              # Detailed operation manual
├── LICENSE                          # GPLv3
├── config.json                      # Runtime config + backtest-calibrated weights
├── pyproject.toml                   # Python project metadata + linting config
├── .env.example                     # Environment variable template
├── requirements.txt                 # Python dependencies
├── run.py                           # CLI entry point (real data only, no --mock/--dry-run)
├── mcp_server.py                    # MCP server
├── core/
│   ├── data_fetcher.py              # Pandadata data acquisition (with retry + backoff)
│   ├── flow_fetcher.py              # Fund flow: 同花顺 → AKShare → East Money direct (3-path)
│   ├── pattern_detector.py          # 8 technical pattern detectors (float strength)
│   ├── flow_filter.py               # Capital flow filter + inflow rate normalization
│   ├── scorer.py                    # Scoring and ranking (z-score industry + mcap neutralization)
│   ├── reporter.py                  # Markdown + JSON report generator (data provenance + portfolio)
│   ├── pipeline.py                  # End-to-end pipeline orchestrator
│   ├── portfolio.py                 # Portfolio optimization (equal-weight, min-variance, risk-parity) [v2.1]
│   ├── cache.py                     # Date-keyed Parquet cache (with meta.json)
│   └── mock_data.py                 # Mock data generator (internal testing only)
├── llm/
│   └── analyst.py                   # LLM analysis (Claude/DeepSeek, concurrent API calls)
├── tests/
│   ├── test_pattern_detector.py     # 45 pattern detector tests
│   ├── test_integration.py          # 32 pipeline/integration tests
│   ├── test_analyst.py              # 24 LLM analyst tests
│   ├── test_api_integration.py      # 25 mocked API integration tests (3-path fallback)
│   └── test_reporter.py             # 28 report generation tests
├── references/
│   ├── pandadata-map.md             # Data routing reference
│   ├── pattern-formulas.md          # Detector formulas and weights
│   ├── report-template.md           # Report template + LLM prompt
│   └── validate-rules.md            # Output integrity validator rule checklist
├── scripts/
│   ├── validate_screener.py         # Output integrity validator
│   ├── daily_screener.bat           # Windows Task Scheduler entry point
│   ├── benchmark.py                 # Performance benchmark
│   └── analyze_weights.py           # IC analysis and weight optimization
├── agents/
│   ├── openai.yaml                  # OpenAI/Codex adapter
│   ├── cursor-rule.mdc              # Cursor IDE adapter
│   └── portable-loader.md           # Generic loader for any agent
└── output/                          # Reports grouped by date
    └── YYYY-MM-DD/
        ├── daily_screener_YYYYMMDD.md
        └── daily_screener_YYYYMMDD.json

Core Constraints

Constraint	Description
Dual-factor must hold	Stock passes only if pattern AND flow both trigger
ST stock exclusion	*ST / ST / PT / 退市整理期股票直接排除**。2026年7月6日起 ST 涨跌幅由 ±5%→±10%，波动空间翻倍，形态检测器假阳性率上升。待积累 ≥60 个交易日新规后数据后运行权重回测，再评估是否重新纳入
ST liquidity threshold	若未来重新纳入 ST 股，其 `min_turnover` 阈值需加倍至 100M CNY（涨跌幅放宽后流动性门槛相应提高）
Liquidity threshold	Turnover < 50M CNY → auto-exclude（非ST普通股）
No stock recommendation	Use "值得关注" / "可跟踪", forbid "买入" / "目标价"
Trading-day aware	Skip holidays, check trade calendar
Data fault tolerance	Skip individual failing stocks, do not halt the full scan
Audit trail	JSON output must include scoring breakdown per stock (raw + neutralized)
Data provenance	Report must label each data category's actual source and authenticity
Post-market data timing	2026-07-06起：盘后固定价格交易（15:05-15:30）扩容至全部 A 股，触发时间调整为 15:45。需验证 Pandadata K 线数据是否包含盘后成交
Survivorship bias	股票 universe 来自 `get_trade_list`（当日可交易），已退市股票不在历史回测中。长周期回测的 IC 可能被高估 0.02–0.05。使用固定起点 universe 可消除此偏差
Data freshness	数据拉取后校验目标日期数据时间戳：K 线的最大 `trade_date` ≠ 目标日期 → 数据尚未就绪，等待 60s 后重试（最多 3 次）。全部重试失败 → 报告标注「数据延迟」+ 实际最新数据日期
Runtime SLA	全市场扫描（~5000 只）：预期 8–15 分钟（4 线程 + Parquet 缓存命中）。>30 分钟 → WARNING 记录 API 延迟。>60 分钟 → 超时退出，输出已完成的部分结果

Disclaimer

This skill's output is for research reference only. It does not constitute any investment advice. Investors should make independent judgments and bear trading risks.

License

GPLv3

post-market-screener

Post-Market Screener

Workflow