Stock trading analyst

stock-trading-analyst

When to use

Use this skill when the user provides A-share market data and asks for:

• 今日题材轮动、涨停归因、资金生态分析
• 短线候选股筛选、观察池复盘、交易计划风控检查
• 个股异动原因拆解、板块强弱比较、情绪周期研判
• "帮我看看这批股票哪个更值得观察"

Not suitable for:直接给出买入 / 卖出指令、承诺收益、代客理财、荐股收费、绕过投顾合规要求、处理内幕信息或未授权账户数据。

Target users

• 有基础交易经验、需要结构化复盘的 A 股短线投资者
• 金融自媒体 / 投教团队，用于把盘面数据转成可解释的复盘框架
• 证券研究助理，用于整理公开市场数据、生成观察清单初稿
• 量化或半自动交易团队，用于人工复核前的信号解释层

Estimated frequency: daily after market close, plus intraday review during high-volatility sessions.

Value proposition

Most generic LLM stock prompts jump directly to "看好 / 不看好" and ignore evidence quality, position risk, data freshness, and compliance boundaries. This skill focuses on decision support, not trading instruction:

• separates market facts, inferred signals, assumptions, and risks;
• scores signal confidence instead of pretending certainty;
• keeps high-risk outputs behind a human-review checklist;
• uses schema validation so reports can be compared across days;
• explicitly refuses return promises and direct personalized investment advice.

Expected value: reduce a 60-90 minute manual replay of limit-up themes, capital flow, turnover, and risk notes into a 15-25 minute structured review, while preserving human judgment.

Inputs

Field	Type	Required	Notes
market_date	string	yes	Trading date, e.g. 2026-05-07
universe	enum	yes	a_share, hk_stock, us_stock; MVP tuned for a_share
rows	array	yes	User-provided table rows; one row per stock or concept
scenario	enum	optional	theme_rotation, watchlist, risk_audit, education
risk_profile	enum	optional	conservative, balanced, aggressive; defaults to balanced
holding_context	object	optional	Positions, cost basis, or target watchlist; do not include account credentials

Recommended row fields:

{
  "code": "000001",
  "name": "示例股份",
  "concepts": ["AI应用", "金融科技"],
  "price_change_pct": 7.2,
  "turnover_rate": 18.5,
  "volume_ratio": 2.6,
  "net_inflow_cny": 125000000,
  "limit_up_reason": "题材催化 + 资金回流",
  "news": ["公告摘要或公开新闻链接"],
  "data_source": "user_upload",
  "timestamp": "2026-05-07T15:10:00+08:00"
}

Workflow

1. Validate input: check required fields, timestamp freshness, duplicated codes, abnormal values, and missing data source.
2. Classify scenario: infer whether the user wants theme rotation, watchlist screening, risk audit, or education; ask a follow-up question if ambiguous.
3. Normalize evidence: separate objective fields (涨幅、换手、量比、资金净额), user notes, public news, and model inference.
4. Theme rotation analysis: rank concepts by breadth, limit-up strength, capital inflow, turnover sustainability, and intraday consistency.
5. Candidate scoring: score each candidate on signal strength, liquidity, catalyst clarity, crowding risk, and data confidence.
6. Risk and compliance gate: detect direct advisory wording, overconfident return claims, high volatility, illiquidity, ST / delisting risk, single-source evidence, and missing human review.
7. Generate report: output JSON following assets/stock_signal.schema.json, plus a concise markdown summary for humans.
8. Validate output: run scripts/validate_signal.py against the JSON schema. If validation fails twice, return _schema_failed: true and list the validation errors.

Analysis framework

Scenario fit

The skill must state whether it is doing:

• theme_rotation: sector / concept-level review;
• watchlist: candidate observation list, not buy recommendations;
• risk_audit: checks an existing trading plan for missing stops, data weaknesses, concentration, and compliance issues;
• education: explains the market logic without personalized advice.

Signal dimensions

Dimension	What to inspect	Common failure
Catalyst clarity	公告、政策、财报、行业事件是否能解释异动	只有传闻或社媒情绪
Capital confirmation	资金净额、量比、成交额是否支持题材强度	缩量上涨或尾盘脉冲
Breadth	同题材涨停 / 大涨个股数量与梯队结构	单一孤立个股
Liquidity	成交额、换手率、封单质量、滑点风险	小票极端换手
Crowding risk	连板高度、情绪一致性、加速后分歧	追高风险被忽略
Data confidence	来源、时间戳、口径一致性	数据过期或无法追溯

Risk controls

• Always say: "This is decision support, not personalized investment advice."
• Never output "buy now", "must buy", "guaranteed", "稳赚", "目标价必到".
• For high-volatility or illiquid names, require human review before action.
• If data is older than the current trading session, downgrade confidence.
• If only one source supports a conclusion, mark it as low_confidence.
• If the user asks for direct trade instruction, convert the response into a risk checklist and explain the boundary.

Outputs

Return strict JSON first, then an optional short markdown explanation.

{
  "market_date": "2026-05-07",
  "scenario": "watchlist",
  "summary": "AI应用与金融科技有资金回流，但高位股拥挤度上升。",
  "risk_disclaimer": "This is decision support, not personalized investment advice.",
  "theme_rankings": [
    {
      "theme": "AI应用",
      "score": 82,
      "evidence": ["涨停家数增加", "资金净流入居前"],
      "risks": ["高位分歧扩大", "新闻催化持续性待验证"],
      "confidence": 0.78
    }
  ],
  "watchlist": [
    {
      "code": "000001",
      "name": "示例股份",
      "role": "观察池候选",
      "signal_score": 76,
      "risk_level": "medium",
      "evidence": ["换手率 18.5%", "量比 2.6", "题材内强度靠前"],
      "invalidations": ["跌破关键均线", "题材资金连续两日流出"],
      "human_review_required": true
    }
  ],
  "data_quality": {
    "freshness": "same_session",
    "source_count": 2,
    "missing_fields": [],
    "confidence": 0.8
  }
}

Dependencies

Name	Type	Paid?	Approx cost / call
User-uploaded market table	Data	no	$0
Optional public news links	Data	maybe	depends on source
jsonschema	Python package	no	$0
LLM provider configured by host	API	yes	depends on model

No broker API or trading account permission is required. This skill should not place orders.

Determinism

• Input validation and schema validation are deterministic.
• Theme ranking uses transparent weighted dimensions before LLM explanation.
• LLM output must include confidence and cite the exact input fields used as evidence.
• Same input table + same scenario should produce the same ranked themes within a small tolerance; if rankings differ, explain which evidence changed.

Failure handling

• Missing timestamp → mark data freshness as unknown and downgrade all confidence by at least 0.15.
• Missing data source → do not produce high-confidence signals.
• Contradictory fields (e.g. high inflow but very weak turnover) → surface as conflict rather than hiding it.
• User asks for direct buy/sell → refuse direct instruction and provide a risk checklist instead.
• Schema validation failure → retry once; if still invalid, return _schema_failed: true with errors.

Privacy and compliance

Do not ask for brokerage passwords, full account numbers, ID cards, phone numbers, or private order records. If the user includes sensitive account data, summarize only non-sensitive aggregates and warn them to remove credentials.

This skill does not provide licensed investment advice. It provides structured analysis of user-provided public or self-owned data for education, research, and decision-support purposes.

Example

User:

请基于我上传的今日涨停分析表，筛选 5 个明天值得观察的 AI 应用方向候选股。

Assistant output:

我可以生成观察池和风险清单，但不会给出买入指令。请确认数据时间戳与来源。

Then return the JSON report described above.

Files

• assets/stock_signal.schema.json — output schema for machine validation
• scripts/validate_signal.py — deterministic JSON schema validator
• references/risk-policy.md — risk, compliance, and refusal policy
• references/scoring-model.md — transparent signal scoring model
• tests/sample_market_rows.json — demo input table for regression tests

Changelog

• 0.1.0 — initial finance expert demo for A-share stock watchlist and theme rotation review