ETF Premium Rate Analysis Skill (v4)

Decision-tree driven. Anomaly-first. Built-in calculator.

Core Principle

Don't run every module every time. Follow the signal.

1. Grab data → 2. Check for anomalies → 3. Go deep only if needed

Data Strategy

Tier & Quality

Every output must declare its quality grade on the first line.

Browser Trigger Rules (automatic)

Use browser (mcp__playwright__browser_*) when:
  → Current premium > 8% AND last reliable data > 1 hour old
  → Single-day premium change > 3pp suspected (verify before alarming)
  → User requests --buy (need precise real-time price)

Browser strategy (tested in practice):
  ⚠️ Direct URL navigation to 雪球/集思录 often fails (404 or login walls).
  INSTEAD: use WebSearch FIRST to find a working page URL, THEN navigate.

  Approach that worked:
    1. WebSearch: "{code} 溢价率 IOPV 净值 site:xueqiu.com" → find the correct page URL
    2. Browser: navigate to the URL from search results
    3. Snapshot → extract data → close browser

  Approach that partially worked:
    1. Browser: navigate to https://www.jisilu.cn/data/etf/ → see column headers
    2. But detailed data rows are JS-dynamic, snapshot may not capture them
    3. Fallback: WebSearch with snippet extraction is more reliable for numbers

  Bottom line: Browser is a BEST-EFFORT enhancement, not a dependency.
  WebSearch (Tier 1) is the only guaranteed data source.

WebSearch Query Patterns

Single ETF:  "{code} {name} 最新净值 市价 溢价率 {YYYY}年"
History:     "{code} 溢价率 历史 走势 {YYYY}"
Peers:       "{category} ETF {code1} {code2} {code3} 溢价率 对比"
Quota:       "{code} 暂停申购 QDII额度"
Catalyst:    "QDII额度 外管局 {YYYY}年" / "跨境ETF 溢价风险 停牌"

Data Freshness (same-session reuse)

Session Memory

If the same ETF or category was analyzed earlier in this session:
  → Reference prior findings: "30分钟前的 scan 显示..."
  → Highlight what CHANGED since last analysis
  → Don't re-derive what's already known

Decision Tree (Main Workflow)

┌─ Step 0: Market Status Check (always first) ──────────────────┐
│  WebSearch: "{code} 实时行情" or quick check                   │
│  → Determine: 盘中(live) / 盘后(use last close) / 休市        │
│  → Label all outputs accordingly                               │
│  → 盘中用 IOPV(参考净值) / 盘后用最新NAV(单位净值)             │
└───────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─ Step 1: Quick Data Grab (always, parallel) ──────────────────┐
│  WebSearch: price + NAV for target ETF                        │
│  WebSearch: price + NAV for 2-3 category peers                │
│  WebSearch: quota/creation status                             │
│  [+ browser search-then-navigate if --buy flag]               │
└───────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─ Step 2: Triage ──────────────────────────────────────────────┐
│                                                               │
│  Calculate: premium%, peer_median%, deviation_from_peer        │
│                                                               │
│  ┌─ premium < 2% AND |deviation| < 1pp?                       │
│  │  → "正常范围" — output TLDR, STOP (no deep needed)         │
│  │                                                            │
│  ├─ deviation from peer > 3pp?                                │
│  │  → 🔴 ANOMALY: investigate why THIS ETF differs            │
│  │                                                            │
│  ├─ premium > 8%?                                             │
│  │  → 🔴 ANOMALY: run passive-inflation check                 │
│  │                                                            │
│  ├─ single-day premium change > 3pp?                          │
│  │  → 🔴 ANOMALY: run catalyst scan                           │
│  │                                                            │
│  └─ none of the above?                                        │
│     → "溢价偏高但在正常范围" — quick output, STOP             │
│                                                               │
└───────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─ Step 3: Anomaly Deep-Dive (only if triggered) ───────────────┐
│                                                               │
│  If deviation from peer:    → Check creation status, fund      │
│                               size, liquidity differences      │
│                                                               │
│  If premium > 8%:           → Passive-inflation check          │
│                               Compare: NAV Δ% vs Price Δ%      │
│                               over last 5/10/20 days            │
│                                                               │
│  If rapid change:           → Catalyst scan (quota news,       │
│                               regulatory actions, index moves) │
│                                                               │
└───────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─ Step 4: Calculator (if --buy N flag) ────────────────────────┐
│  See Calculator Mode section below                            │
└───────────────────────────────────────────────────────────────┘

Detection Rules

Rule 1: Passive Premium Inflation (被动溢价膨胀)

The most dangerous and most commonly missed pattern.

Fetch: NAV change% and Price change% over last 5, 10, and 20 trading days.

IF |NAV decline| > 3% over period
   AND |Price decline| < |NAV decline| (price falling slower)
   AND premium increased > 1pp over same period
THEN:
   → Label: "被动溢价膨胀"
   → Interpretation: 溢价不是被人买上去的，是被 NAV 下跌垫高的
   → Risk: 市价滞后于 NAV，一旦补跌 → 溢价压缩 + 继续亏钱 = 双杀
   → Action: 在输出中显眼位置警告，这是最危险的溢价形态

IF |NAV increase| > 3%
   AND |Price increase| > |NAV increase| (price rising faster)
   AND premium increased > 1pp
THEN:
   → Label: "主动溢价扩张"
   → Interpretation: 资金在主动追高，溢价有动能支撑
   → Risk: 情绪驱动的溢价在情绪反转时会快速崩塌

Rule 2: Peer Divergence

IF |premium(T) - peer_median| > 3pp:
   → Investigate: 申购状态差异? 流动性差异? 规模差异?
   → Output: "这只 ETF 比同类多/少了 Xpp，原因可能是..."

Rule 3: Regime Assessment (Simplified)

Gone is the 7-regime classification requiring 250d data. Instead:

Using whatever historical data is available (annual stats, recent trend):

Q1: Is premium currently near its recent (60d) range high or low?
    → "处于近期偏{高/低}位置"

Q2: Has premium been trending up, down, or sideways over the past month?
    → "近一个月{上行/下行/横盘}"

Q3: Is volatility elevated or subdued vs typical?
    → "波动率偏{高/低}"

Combine into a one-line assessment:
  "溢价 9.9%，处近期偏高水平，近一月横盘但月末加速上行，波动率正常"

This replaces the old 7-regime classification with a continuous, data-driven description that works with partial data.

Rule 4: Cost of Switching

Always compute when analyzing a single ETF with peers. One-line output only:

savings = (premium_T - min_peer_premium) / (1 + premium_T/100) × amount

→ "选 {min_code} ({min_prem}%) 可省 ¥{savings} ({savings_pct}%本金)"

No need to expand into full table unless user asks. Just the number.

Calculator Mode (--buy)

Trigger: /etf-premium <code> --buy <amount> or user mentions specific investment amount.

Output Tiers

Calculator output ADAPTS to what's triggered:

TIER 1 — Normal (no anomaly triggered):
  → Shares + actual cost + entry fee + convergence table + cost-of-switching
  → MAX 15 lines. No passive-inflation check, no catalyst scan.
  → End with one-liner summary.

TIER 2 — Anomaly (passive inflation OR peer divergence triggered):
  → Tier 1 content
  → + compact anomaly alert block (3-5 lines, not full deep-dive)
  → + "→ 建议 /etf-premium --deep {code} 了解催化剂详情"

TIER 3 — Full (user explicitly added --deep):
  → Tier 1 + Tier 2 + full catalyst scan + quality assessment

Calculation (for A-Share ETFs, see Appendix D for rules)

买入金额:       ¥{amount}
当前市价:       ¥{price}
参考净值:       ¥{nav}  (盘中=IOPV/盘后=最新NAV，见术语说明)
当前溢价率:     {prem}%

实得份额:       {shares} 份 = floor(amount / price / 100) × 100  (A股ETF100份/手)
实际花费:       ¥{actual_cost} = shares × price
剩余现金:       ¥{remainder} = amount - actual_cost  (不足1手，留在账户)

底层资产价值:   ¥{nav_value} = shares × nav
"进门费":       ¥{premium_cost} = actual_cost - nav_value
               → 多付 {premium_cost} 元，买到价值 {nav_value} 元的资产
               → 占比: {premium_cost / actual_cost * 100}% 的本金

Convergence Scenarios

假设底层资产不涨不跌，仅溢价率变化:
(注: 溢价率 = 市价/参考净值 - 1，参考净值在盘中为IOPV、盘后为最新NAV)

┌──────────────────────┬─────────────┬──────────────┬──────────────┐
│ 情景                 │ 溢价率      │ 你的份额市值   │ 盈亏         │
├──────────────────────┼─────────────┼──────────────┼──────────────┤
│ 当前                 │ {prem}%     │ ¥{actual_cost}│ —            │
│ 回归 20 日均线       │ {mean}%     │ ¥{v_mean}    │ {gain_loss}  │
│ 回归同类中位数       │ {peer_med}% │ ¥{v_peer}    │ {gain_loss}  │
│ 回归历史均值         │ {hist_avg}% │ ¥{v_hist}    │ {gain_loss}  │
│ 溢价归零             │ 0%          │ ¥{v_zero}    │ {gain_loss}  │
│ 触及近期高点上轨     │ {upper}%    │ ¥{v_upper}   │ {gain_loss}  │
└──────────────────────┴─────────────┴──────────────┴──────────────┘

盈亏平衡点:
  如果溢价回归 20 日均线，底层资产需涨 {breakeven}% 你才不亏。
  纳指一个月涨 {breakeven}% 的历史概率约为 {prob}%。

Compact Output (Tier 1 — normal)

## 买入计算: {name} ({code})

投入 ¥{amount} | 市价 ¥{price} | {nav_label} ¥{nav} | 溢价 {prem}% | {grade}

实得: {shares} 份 ({lots} 手) | 花费: ¥{actual_cost} | 剩余: ¥{remainder}
底层价值: ¥{nav_value} | 进门费: ¥{premium_cost} ({fee_pct}%本金)

回归情景 (底层不动):
  20日均线({mean}%): -¥{loss_mean} | 同类中位数({peer}%): -¥{loss_peer}
  年度均值({hist}%): -¥{loss_hist} | 归零(0%): -¥{loss_zero}

→ 选 {min_code} ({min}%) 可省 ¥{savings} | 需纳指涨 {be}% 回本 (概率~{prob}%)

Expanded Output (Tier 2 — anomaly triggered)

Append after compact output:

⚠️ {anomaly_label}: {one-line_explanation}
  {anomaly_detail}
→ 建议 /etf-premium --deep {code} 了解催化剂

Output Templates (Minimal)

TLDR Mode (--tldr)

{code} {name}: 溢价 {prem}% [{grade}] [{market}] | 偏离同类{+/-}pp | {anomaly_flag}
→ {actionable_one-liner}

例 (盘中):
513100 纳指ETF: 溢价 9.5% [B] [盘中] | +1.7pp | ⚠️被动膨胀 | 今日停牌
→ 停牌是监管升级信号，等冲击消化后再判断

例 (盘后):
159941 纳指ETF: 溢价 9.9% [B] [盘后] | +2.1pp | — | 基于7/30收盘
→ 同类最低 159632 (6.5%)，换一只每¥10,000省¥300

Quick Mode (default)

## {name} ({code})

{timestamp} | {market_status} | 数据质量: {grade} | 同类均值: {peer_med}%

### 核心数字
| 指标 | 数值 |
|------|------|
| 市价 / {nav_label} | ¥{price} / ¥{nav} |
| **溢价率** | **{prem}%** ({position}于近期区间) |
| 同类排名 | 第 {rank}/{n} | 
| 溢价质量 | {quality_label} |

### 检测结果
{only show triggered rules, skip clean ones}

### 盈亏快照
回归中枢: -¥{loss} | 触及上轨: +¥{gain}

→ 选 {min_code} ({min}%) 可省 ¥{savings}

{If anomalies: "→ 建议 /etf-premium --deep {code}"}

Deep Mode (--deep)

Only when anomalies are triggered or user explicitly requests:

## {name} ({code}) — 深度分析

{grade} | {timestamp} | {market_status}

### 1. 异常检测
{all triggered rules with detailed explanation}

### 2. 溢价全景
- 当前: {prem}% | 近期中枢: {mean}% | 近期波动带: [{low}%, {high}%]
- 走势: {trend_description}
- 历史参照: {annual_mean}% (年均), {annual_range} (年范围)
- 质量评估: {quality_detail}

### 3. 同类定位
{peer comparison table, sorted by premium}
结构性占比: {structural_pct}% → {interpretation}

### 4. 催化剂扫描
{only relevant catalysts, not all 7}
| 催化剂 | 状态 | 近期概率 | 影响 |
|--------|------|---------|------|
| {top 3-4 catalysts} |

### 5. 情景与计算
{scenarios with probabilities — only if useful for this specific case}

### 6. 综合判断
{one concise paragraph, separate data from opinion}

Scan Mode (--scan)

## {Category} ETF 扫描

{timestamp} | {grade} | {n} 只

| 代码 | 名称 | 溢价率 | vs同类 | 异常? | 申购 |
|------|------|--------|--------|-------|------|
| ... | ... | ...% | {+/-}pp | {flag} | {status} |

**类别中位数**: {median}% | **离散度**: {dispersion}pp
**最低成本**: {code} ({min}%) — 每 ¥10,000 比最高省 ¥{savings}
**异常标记**: {list of flagged ETFs with brief reason}
**今日事件**: {停牌/风险提示/其他值得注意的事件}

Data Sources (Quick Reference)

Primary (Tier 1 — always works)

WebSearch: targeted queries as defined in Data Strategy section
  → This is the ONLY guaranteed data source. Always start here.

Browser (Tier 2 — best-effort, NOT guaranteed)

⚠️ Direct URL navigation to financial sites often fails.
   Strategy: WebSearch first → find working page URL → browser navigate.

   Only attempt for: --buy flag (precision matters) or volatile markets.
   If browser fails after 2 attempts → fall back to WebSearch, downgrade grade to B.

Optional API (Tier 3 — may fail, never depend)

市价:  GET push2.eastmoney.com/api/qt/ulist.np?fields=f2,f3,f4,f12,f14&secids=1.{code}
净值:  GET fundgz.1234567.com.cn/js/{code}.js
历史:  GET api.fund.eastmoney.com/f10/lsjz?fundCode={code}&pageIndex=1&pageSize=250
K线:   GET push2his.eastmoney.com/api/qt/stock/kline/get?secid=1.{code}&klt=101&lmt=250

Appendix A: Category Reference Groups

纳指:       513100, 159941, 513300, 159632, 159659, 513110, 159501
纳指科技:   159509 (注意: 跟踪纳指科技指数，非纳指100)
标普500:    513500, 161125
日经:       513520, 159866
德国DAX:    513030
恒生:       159920, 513600, 510900
恒生科技:   513180, 159742, 513010
黄金:       518880, 159934, 159812
沪深300:    510300, 159919, 510310
中证500:    510500, 159922, 510510
科创50:     588000, 588080, 588380

Appendix B: ETF Personality Cards

Persistent traits that don't change day-to-day. Load these automatically when analyzing the relevant ETF.

纳指 ETF 族

关键日期

• 2024/11: 513100 暂停申购（至今未恢复，已超 8 个月）
• 2026/3/27: 外管局新批 53 亿美元 QDII 额度（2021 年来单次最大）
• 2026/5 初: 3 月批的额度耗尽，热门产品再次全线暂停
• 2026/6/2: 12 只纳指 ETF 首次同步停牌 1 小时
• 2026/6/17: 外汇局局长陆家嘴论坛表态"新增发放 QDII 额度"
• 2026/7/26: 15 只 QDII 同日发布溢价风险提示
• 2026/7/31: 513100 首次单独触发停牌（监管 escalation 信号）

Appendix D: ETF Trading Rules & Terminology

Trading Rules (A-Share ETFs)

最小交易单位: 100 份 = 1 手
份额计算:      floor(amount / price / 100) × 100
交收制度:      T+1 (今日买，明日到账)
交易时间:      9:30-11:30, 13:00-15:00 (交易日)
涨跌幅限制:    10% (大部分ETF)，20% (科创/创业板ETF)

NAV vs IOPV — When to use which

IOPV (实时参考净值):
  - 交易所每15秒计算并发布一次
  - 盘中交易时用作"实时净值"参考
  - 是估算值，不是精确值
  - 盘中使用: "参考净值 (IOPV)"

NAV (单位净值):
  - 基金公司 T+1 日公布
  - 基于收盘价精确计算
  - QDII ETF 的 NAV 通常滞后1-2个交易日(涉及境外市场结算)
  - 盘后使用: "单位净值 (NAV)"

溢价率计算:
  盘中: (市价 / IOPV - 1) × 100   ← 用IOPV
  盘后: (收盘价 / 最新NAV - 1) × 100  ← 用NAV

输出时标注使用的是哪个:
  "参考净值(IOPV) ¥{x}" 或 "单位净值(NAV) ¥{x}"

Market Status Detection

Before any analysis, determine market status:
  → 盘中 (A股 9:30-15:00 交易日): 用 IOPV，标注 "盘中实时"
  → 盘后 (15:00后): 用收盘价 + 最新NAV，标注 "盘后，基于{date}收盘"
  → 周末/节假日: 用最近交易日数据，标注 "休市，最近数据{date}"
  → 美股时段 (21:30-04:00): 对 QDII ETF 特别标注 "美股{open/closed}"

检测方式: WebSearch "{code} 实时行情" → 看是否有今日盘中数据

Execution Rules

1. Check market status first — determine 盘中/盘后/休市 before any analysis
2. Use correct NAV term — 盘中=IOPV(参考净值), 盘后=NAV(单位净值), always label which
3. Grade your data — first line of every output: 数据质量: {A/B/C/D}
4. Tier output to anomalies — normal → compact; anomaly → expand; --deep → full
5. Highlight anomalies, skip normal — if nothing is wrong, say so briefly
6. Always show cost of switching (one line) — "→ 选 {code} ({prem}%) 可省 ¥{X}"
7. Passive inflation check on every premium > 8% — this pattern kills returns
8. Reference session memory — if data already fetched, use it and note deltas
9. Browser is best-effort — WebSearch first, browser only as enhancement, don't depend on it
10. Calculator when asked — user says a number → run --buy logic, compact format
11. Separate data from opinion — "数据显示 X，但这取决于 Y"
12. Acknowledge ETF floor lots — calculator always shows remainder cash

Appendix C: Quick Reference Card

ETF 溢价分析 v4.1

状态:  先检查: 盘中?/盘后?/休市? → 决定用IOPV还是NAV

数据:   [A]浏览器实时 / [B]WebSearch可靠 / [C]估算 / [D]参考
        ⚠️ 浏览器是best-effort增强，不依赖

检测:   被动膨胀? (NAV↓ > 价↓) | 同类偏离? (>3pp) | 溢价>8%? | 单日跳变? (>3pp)

计算:   --buy N → 份额(100份/手) / 进门费 / 回归情景 / 剩余现金
        Tier 1(正常)→紧凑 | Tier 2(异常)→追加警告 | Tier 3(--deep)→完整

换一只: → 选 {code} ({prem}%) 可省 ¥{X}  ← 永远是1行

人格:   513100=老大哥(关闭>8月) | 159941=开关不定 | 159632=省钱之选
       159509=科技集中(波动最大) | 159659=新秀偏低

关键日: 2024/11 513100关 | 2026/3/27 批53亿$ | 6/2 首次同步停牌
       6/17 局长表态 | 7/26 15只QDII同日预警 | 7/31 513100单独停牌

Changelog

• 2026-07-31 v4.1: Bugfix & polish — Fixed browser URLs (direct navigation unreliable; changed to search-then-navigate). Added calculator output tiering (Tier1=compact 15 lines, Tier2=+anomaly alert, Tier3=full deep). Added Appendix D (ETF trading rules, IOPV vs NAV terminology, market status detection). Added Step 0: market status check before every analysis. Unified NAV/IOPV terminology throughout all templates. Compressed cost-of-switching to single-line output. Calculator now shows remainder cash. Execution rules expanded with market check, tiering, terminology rules.
• 2026-07-31 v4: Major refactor — replaced 7-module framework with decision-tree workflow. Added passive-inflation detection rule, built-in purchase calculator (--buy), cost-of-switching comparison, data quality grading system (A/B/C/D), ETF personality appendix with persistent knowledge, session memory rules, and browser trigger conditions. Simplified regime assessment from 7-category classification to continuous 3-question description. Reduced output templates to minimal "anomaly-first" format.
• 2026-07-31 v3: Percentile-based regime detection, catalyst hunting, duration-aware probabilities, premium quality assessment, TLDR mode, data freshness tiers, skew-aware bands.
• 2026-07-31 v2: Dynamic regime detection, dynamic thresholds, premium decomposition, scenario simulation, quota tracking.
• 2026-07-31 v1: Initial skill with static thresholds.