deeptrade-quant 0.0.2__py3-none-any.whl

deeptrade/strategies_builtin/volume_anomaly/volume_anomaly/pipeline.py

@@ -0,0 +1,327 @@
+"""LLM pipeline for the volume-anomaly走势分析阶段。
+
+Single-stage strategy: no R2 / final_ranking. Multi-batch is supported but the
+batch outputs are simply concatenated (no global re-rank), per the user's spec.
+
+Borrows the dual input/output token budget pattern from limit-up-board, but
+does NOT import from it (plugins are self-contained).
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Iterable
+from dataclasses import dataclass, field
+from typing import Any
+
+from pydantic import BaseModel
+
+from deeptrade.core.llm_client import (
+    LLMClient,
+    LLMTransportError,
+    LLMValidationError,
+)
+from deeptrade.plugins_api import StageProfile
+from deeptrade.plugins_api.events import EventLevel, EventType, StrategyEvent
+
+from .data import AnalyzeBundle
+from .profiles import STAGE_TREND_ANALYSIS, resolve_profile
+from .prompts import VA_TREND_SYSTEM, va_trend_user_prompt
+from .schemas import VATrendCandidate, VATrendResponse
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Token budgets / batching
+# ---------------------------------------------------------------------------
+
+
+# Per-candidate cost is higher than limit-up-board's R1 because we ship 60d
+# aggregates + 5d recent OHLCV + moneyflow summary per stock.
+DEFAULT_AVG_INPUT_TOKENS_PER_CANDIDATE = 1_500
+# v0.6.0 — bumped from 900 → 1_100 to absorb dimension_scores (6 ints) +
+# alpha_*_pct fields without crowding output budgets at 25-candidate batches.
+DEFAULT_AVG_OUTPUT_TOKENS_PER_CANDIDATE = 1_100
+DEFAULT_INPUT_BUDGET = 200_000  # matches continuation_prediction stage
+SAFETY_RATIO = 0.85
+
+
+@dataclass
+class BatchPlan:
+    batch_size: int
+    n_batches: int
+
+
+def plan_batches(
+    *,
+    n_candidates: int,
+    input_budget: int,
+    output_budget: int,
+    overhead_input_tokens: int = 5_000,
+    avg_in: int = DEFAULT_AVG_INPUT_TOKENS_PER_CANDIDATE,
+    avg_out: int = DEFAULT_AVG_OUTPUT_TOKENS_PER_CANDIDATE,
+) -> BatchPlan:
+    """Pick the largest batch_size satisfying BOTH input and output budgets."""
+    if n_candidates <= 0:
+        return BatchPlan(batch_size=0, n_batches=0)
+    in_room = max(0, input_budget - overhead_input_tokens)
+    by_in = max(1, in_room // max(avg_in, 1))
+    by_out = max(1, int(output_budget * SAFETY_RATIO) // max(avg_out, 1))
+    batch_size = max(1, min(by_in, by_out, n_candidates))
+    n_batches = (n_candidates + batch_size - 1) // batch_size
+    return BatchPlan(batch_size=batch_size, n_batches=n_batches)
+
+
+# ---------------------------------------------------------------------------
+# Set equality check (input candidate_id ⊆ ⊇ output)
+# ---------------------------------------------------------------------------
+
+
+def _ids(items: Any) -> set[str]:
+    return {(c["candidate_id"] if isinstance(c, dict) else c.candidate_id) for c in items}
+
+
+def _set_mismatch_repair_hint(expected: set[str], actual: set[str]) -> str:
+    missing = sorted(expected - actual)
+    extra = sorted(actual - expected)
+    parts = ["\n\n⚠ 上一次响应集合不一致，请严格按照原 candidate_id 列表重新输出。"]
+    if missing:
+        parts.append(f"missing (你必须包含): {missing}")
+    if extra:
+        parts.append(f"extra (你不能包含): {extra}")
+    parts.append("不可遗漏，不可新增，不可改名。")
+    return "\n".join(parts)
+
+
+class _SetMismatchError(Exception):
+    """LLM output's candidate_id set still differs after repair retry."""
+
+
+def _complete_with_set_check(
+    llm: LLMClient,
+    *,
+    system: str,
+    user: str,
+    schema: type[BaseModel],
+    profile: StageProfile,
+    expected_ids: set[str],
+    output_attr: str = "candidates",
+    repair_retries: int = 1,
+    envelope_defaults: dict[str, Any] | None = None,
+) -> tuple[Any, dict[str, Any]]:
+    """Call LLM once; on candidate_id set mismatch, append a repair hint and retry."""
+    current_user = user
+    last_actual: set[str] = set()
+    for attempt in range(repair_retries + 1):
+        raw, meta = llm.complete_json(
+            system=system,
+            user=current_user,
+            schema=schema,
+            profile=profile,
+            envelope_defaults=envelope_defaults,
+        )
+        obj = raw if isinstance(raw, schema) else schema.model_validate(raw)
+        items = getattr(obj, output_attr)
+        actual_ids = {item.candidate_id for item in items}
+        if expected_ids == actual_ids:
+            return obj, meta
+        last_actual = actual_ids
+        if attempt < repair_retries:
+            current_user = user + _set_mismatch_repair_hint(expected_ids, actual_ids)
+    raise _SetMismatchError(
+        f"set mismatch after {repair_retries + 1} attempts; "
+        f"missing={sorted(expected_ids - last_actual)}, "
+        f"extra={sorted(last_actual - expected_ids)}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Pipeline result container
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class AnalyzeResult:
+    """Outcome of the走势分析 LLM phase."""
+
+    success_batches: int = 0
+    failed_batches: int = 0
+    candidates_in: int = 0
+    candidates_out: int = 0
+    predictions: list[VATrendCandidate] = field(default_factory=list)
+    failed_batch_ids: list[str] = field(default_factory=list)
+    batch_size: int = 0
+    market_context_summaries: list[str] = field(default_factory=list)
+    risk_disclaimers: list[str] = field(default_factory=list)
+
+
+# ---------------------------------------------------------------------------
+# Run analyze stage
+# ---------------------------------------------------------------------------
+
+
+def run_analyze(
+    *,
+    llm: LLMClient,
+    bundle: AnalyzeBundle,
+    preset: str,
+    input_budget: int = DEFAULT_INPUT_BUDGET,
+) -> Iterable[tuple[StrategyEvent, AnalyzeResult | None]]:
+    """Run all analyze batches, yielding (event, terminal_result_or_None).
+
+    The final iteration emits a STEP_FINISHED event paired with the populated
+    AnalyzeResult so the caller can hand it to render.
+    """
+    profile = resolve_profile(preset, STAGE_TREND_ANALYSIS)
+    candidates = bundle.candidates
+    plan = plan_batches(
+        n_candidates=len(candidates),
+        input_budget=input_budget,
+        output_budget=profile.max_output_tokens,
+    )
+    yield (
+        StrategyEvent(
+            type=EventType.LIVE_STATUS,
+            message=(
+                f"[走势分析] 待处理 {len(candidates)} 只，分 {plan.n_batches} 批提交..."
+            ),
+        ),
+        None,
+    )
+    yield (
+        StrategyEvent(
+            type=EventType.STEP_STARTED,
+            message="走势分析（主升浪启动预测）",
+            payload={"n_candidates": len(candidates), "n_batches": plan.n_batches},
+        ),
+        None,
+    )
+
+    result = AnalyzeResult(candidates_in=len(candidates), batch_size=plan.batch_size)
+    if plan.n_batches == 0:
+        yield (
+            StrategyEvent(
+                type=EventType.STEP_FINISHED,
+                message="走势分析（主升浪启动预测）",
+                payload={"predictions": 0},
+            ),
+            result,
+        )
+        return
+
+    for i in range(plan.n_batches):
+        batch = candidates[i * plan.batch_size : (i + 1) * plan.batch_size]
+        yield (
+            StrategyEvent(
+                type=EventType.LIVE_STATUS,
+                message=(
+                    f"[走势分析] 已提交第 {i + 1}/{plan.n_batches} 批 "
+                    f"({len(batch)} 只)，等待 LLM 响应..."
+                ),
+            ),
+            None,
+        )
+        yield (
+            StrategyEvent(
+                type=EventType.LLM_BATCH_STARTED,
+                message=f"analyze batch {i + 1}/{plan.n_batches}",
+                payload={"batch_no": i + 1, "size": len(batch)},
+            ),
+            None,
+        )
+
+        user = va_trend_user_prompt(
+            trade_date=bundle.trade_date,
+            next_trade_date=bundle.next_trade_date,
+            batch_no=i + 1,
+            batch_total=plan.n_batches,
+            candidates=batch,
+            market_summary=bundle.market_summary,
+            sector_strength_source=bundle.sector_strength_source,
+            sector_strength_data=bundle.sector_strength_data,
+            data_unavailable=bundle.data_unavailable,
+        )
+        expected_ids = _ids(batch)
+        try:
+            obj, meta = _complete_with_set_check(
+                llm,
+                system=VA_TREND_SYSTEM,
+                user=user,
+                schema=VATrendResponse,
+                profile=profile,
+                expected_ids=expected_ids,
+                envelope_defaults={
+                    "stage": STAGE_TREND_ANALYSIS,
+                    "trade_date": bundle.trade_date,
+                    "next_trade_date": bundle.next_trade_date,
+                    "batch_no": i + 1,
+                    "batch_total": plan.n_batches,
+                    "market_context_summary": "",
+                    "risk_disclaimer": "",
+                },
+            )
+        except (LLMValidationError, LLMTransportError, _SetMismatchError) as e:
+            result.failed_batches += 1
+            result.failed_batch_ids.append(f"analyze.batch.{i + 1}")
+            yield (
+                StrategyEvent(
+                    type=EventType.VALIDATION_FAILED,
+                    level=EventLevel.ERROR,
+                    message=f"analyze batch {i + 1} failed: {e}",
+                    payload={"batch_no": i + 1},
+                ),
+                None,
+            )
+            continue
+
+        result.success_batches += 1
+        result.candidates_out += len(obj.candidates)
+        result.predictions.extend(obj.candidates)
+        result.market_context_summaries.append(obj.market_context_summary)
+        result.risk_disclaimers.append(obj.risk_disclaimer)
+        yield (
+            StrategyEvent(
+                type=EventType.LLM_BATCH_FINISHED,
+                message=f"analyze batch {i + 1}/{plan.n_batches} ok",
+                payload={
+                    "batch_no": i + 1,
+                    "input_tokens": meta["input_tokens"],
+                    "output_tokens": meta["output_tokens"],
+                },
+            ),
+            None,
+        )
+        n_imminent = sum(1 for c in result.predictions if c.prediction == "imminent_launch")
+        yield (
+            StrategyEvent(
+                type=EventType.LIVE_STATUS,
+                message=(
+                    f"[走势分析] 第 {i + 1}/{plan.n_batches} 批响应已收到 "
+                    f"(累计 imminent_launch {n_imminent} 只)"
+                ),
+            ),
+            None,
+        )
+
+    yield (
+        StrategyEvent(
+            type=EventType.LIVE_STATUS,
+            message=(
+                f"[走势分析] 完成 — 共 {len(result.predictions)}/{result.candidates_in} 个预测"
+            ),
+        ),
+        None,
+    )
+    yield (
+        StrategyEvent(
+            type=EventType.STEP_FINISHED,
+            message="走势分析（主升浪启动预测）",
+            payload={
+                "success_batches": result.success_batches,
+                "failed_batches": result.failed_batches,
+                "predictions": len(result.predictions),
+            },
+        ),
+        result,
+    )

deeptrade/strategies_builtin/volume_anomaly/volume_anomaly/plugin.py

@@ -0,0 +1,22 @@
+"""VolumeAnomalyPlugin — Plugin Protocol entry for the成交量异动策略."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from . import cli as _cli
+
+if TYPE_CHECKING:  # pragma: no cover
+    from deeptrade.plugins_api.base import PluginContext
+
+
+class VolumeAnomalyPlugin:
+    """Framework entry class for the volume-anomaly plugin."""
+
+    metadata = None  # injected by framework after install
+
+    def validate_static(self, ctx: PluginContext) -> None:  # noqa: ARG002
+        from . import schemas  # noqa: F401, PLC0415
+
+    def dispatch(self, argv: list[str]) -> int:
+        return _cli.main(argv)

deeptrade/strategies_builtin/volume_anomaly/volume_anomaly/profiles.py

@@ -0,0 +1,49 @@
+"""LLM stage profile presets for volume-anomaly.
+
+v0.7 — stage 调参档归插件维护。本插件单 stage（``trend_analysis``），名称
+自然回归语义对应位（之前因为框架硬编码 stage 表，借用了 limit-up-board 的
+``continuation_prediction`` 名）。
+
+Preset 语义沿用 v0.6 ``PROFILES_DEFAULT`` 中 R2 阶段对应档（balanced/quality
+开 thinking，fast 关）。
+"""
+
+from __future__ import annotations
+
+from deeptrade.plugins_api import StageProfile
+
+STAGE_TREND_ANALYSIS = "trend_analysis"
+
+
+PROFILES: dict[str, dict[str, StageProfile]] = {
+    "fast": {
+        STAGE_TREND_ANALYSIS: StageProfile(
+            thinking=False, reasoning_effort="medium", temperature=0.2, max_output_tokens=32768
+        ),
+    },
+    "balanced": {
+        STAGE_TREND_ANALYSIS: StageProfile(
+            thinking=True, reasoning_effort="high", temperature=0.2, max_output_tokens=32768
+        ),
+    },
+    "quality": {
+        STAGE_TREND_ANALYSIS: StageProfile(
+            thinking=True, reasoning_effort="high", temperature=0.2, max_output_tokens=32768
+        ),
+    },
+}
+
+
+def resolve_profile(preset: str, stage: str) -> StageProfile:
+    """Look up the StageProfile for ``preset × stage``."""
+    if preset not in PROFILES:
+        raise KeyError(
+            f"unknown profile preset {preset!r}; expected one of {sorted(PROFILES)}"
+        )
+    table = PROFILES[preset]
+    if stage not in table:
+        raise KeyError(
+            f"unknown stage {stage!r} for preset {preset!r}; "
+            f"expected one of {sorted(table)}"
+        )
+    return table[stage]

deeptrade/strategies_builtin/volume_anomaly/volume_anomaly/prompts.py

@@ -0,0 +1,187 @@
+"""Prompt templates for the volume-anomaly走势分析阶段。
+
+Single-stage strategy (no R1/R2/final_ranking). Per the user's spec:
+    * 删除「异动确认」维度（已在筛选环节硬性满足，无需 LLM 重复判断）
+    * 增加「是否经过充分洗盘」维度（洗盘越充分启动概率越大）
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from .prompts_examples import VA_TREND_FEWSHOT
+
+_VA_TREND_SYSTEM_BASE = """\
+你是 A 股「主升浪启动预测」研究助手。你只能基于本次消息中提供的结构化数据进行分析。
+
+【硬性纪律】
+1. 严禁使用外部搜索、新闻网站、公告网站、实时行情、社交媒体、机构观点或任何未提供的数据。
+2. 严禁编造新闻、公告、盘口、传闻、龙虎榜席位（除非数据中明确提供）、资金分歧、ETF 申赎流向。
+3. 如果某字段缺失（出现在 data_unavailable 或单只 missing_data 中），必须显式声明，禁止猜测。
+4. 输入清单中的每一只标的都必须出现在 candidates 数组，candidate_id 与输入完全一致，不可漏不可加。
+5. 仅输出 JSON，不要 Markdown 代码块包裹，不要解释性前后缀。
+
+【任务】
+对本批次"已通过本地异动筛选并入池追踪"的标的进行『主升浪启动预测』。判断该标的在 1-3 个交易日内启动主升浪的概率。
+
+【判断维度】
+A. **是否经过充分洗盘** —— 启动前的洗盘越充分，启动概率越大；具体观察：
+   - 异动前的整理周期长度（base_days）；越长越好
+   - 整理期间的最大回撤（base_max_drawdown_pct）；幅度越深、量越缩越好
+   - 整理期间的成交量是否持续萎缩（base_vol_shrink_ratio）
+   - 整理期间换手率是否充分（base_avg_turnover_rate）
+   - 异动前距上一次涨停的天数（days_since_last_limit_up，若有）
+   - 整理期间的波动率是否收敛（atr_10d_quantile_in_60d / bbw_compression_ratio）；
+     **越低越好**——三维齐降（价收敛 + 量缩 + 波动率收敛）是 VCP 的教科书形态
+   → 对应 `dimension_scores.washout`
+   washout_quality ∈ {sufficient, partial, insufficient, unclear}
+B. **形态结构**：与 ma5/ma10/ma20/ma60 的关系；是否平台/底部突破；是否上升趋势右侧
+   → 对应 `dimension_scores.pattern`
+C. **资金验证**：近 5 日 moneyflow 大单/特大单净流入趋势；net_mf 是否 5 日累计为正
+   → 对应 `dimension_scores.capital`
+D. **板块与市场相对强度**：
+   - 板块层：参考输入【板块强度摘要】sector_strength_source（可信度 limit_cpt_list > 行业聚合）
+   - 市场层：用 `alpha_5d_pct` / `alpha_20d_pct` / `alpha_60d_pct`（个股相对沪深 300）+
+     `rel_strength_label ∈ {leading, in_line, lagging}`（基于 alpha_20d_pct 分档）判断
+     抗大盘强度。**baseline 下跌时个股仍上涨 → 抗跌强势（强信号）**；
+     baseline 上涨而个股跟随 → 弱跟随。alpha 字段可能因数据降级为 None，
+     必须在 missing_data 中显式声明。
+   → 对应 `dimension_scores.sector`
+E. **历史强度**：近 60 日是否已有过涨停 / 是否已经处于二浪 / 距首次异动天数（tracked_days）
+   → 对应 `dimension_scores.historical`
+F. **风险**：是否高位放量出货 / 流通盘过大 / 题材孤立 / 超买连阳 / 缺数据
+   → 对应 `dimension_scores.risk`（**反向打分：分越高代表风险越大**）
+
+【evidence 要求】
+每只 1-5 条 key_evidence；每条必须引用真实出现在输入中的字段名 (`field`)，并填上对应数值 (`value`)、单位 (`unit`) 和你的解读 (`interpretation`)。
+任何无法用输入字段佐证的 rationale 都视为幻觉。
+rationale 不超过 200 字。
+
+【dimension_scores 评分尺度】
+对每个维度（washout / pattern / capital / sector / historical / risk）输出一个 0–100 的整数评分：
+- 0–30：明显不利 / 不充分
+- 30–60：中性 / 部分满足
+- 60–80：明显有利 / 较充分
+- 80–100：教科书级 / 极充分（保留给罕见的极端正例 / 极端风险）
+
+**风险维度方向相反**——分越高代表风险越大；其余维度都是正向评分。
+launch_score 应大致反映各维度综合，但不强制公式约束（保留模型自洽空间）。
+
+【输出语义】
+- launch_score (0-100): 主升浪启动概率分（模型内部排序分）
+- prediction:
+    * imminent_launch — 1-3 个交易日内启动概率高（强信号 + 充分洗盘 + 板块支撑）
+    * watching        — 形态正在构筑，需要再观察
+    * not_yet         — 时机未到 / 已在末段 / 数据矛盾
+- pattern:
+    * breakout              — 突破整理平台
+    * consolidation_break   — 缩量整理后温和放量启动
+    * first_wave            — 一浪初动（早期）
+    * second_leg            — 二浪起涨
+    * unclear               — 形态不清晰
+- washout_quality:
+    * sufficient   — 整理周期长、回撤充分、缩量明显、换手到位
+    * partial      — 仅满足部分维度
+    * insufficient — 几乎没有洗盘（追高风险）
+    * unclear      — 数据不足
+
+【输出格式】（严格按照此 JSON Schema 输出；不要省略任何字段，不要新增字段）
+{
+  "stage": "continuation_prediction",
+  "trade_date": "<原样回传输入中的 trade_date>",
+  "next_trade_date": "<原样回传输入中的 next_trade_date>",
+  "batch_no": <原样回传输入中的 batch_no>,
+  "batch_total": <原样回传输入中的 batch_total>,
+  "market_context_summary": "<整体市场背景 ≤ 100 字>",
+  "risk_disclaimer": "<风险提示 ≤ 80 字>",
+  "candidates": [
+    {
+      "candidate_id": "<原样回传输入中的 candidate_id>",
+      "ts_code": "<原样回传，含 .SH/.SZ 后缀>",
+      "name": "<原样回传输入中的股票名称>",
+      "rank": 1,
+      "launch_score": 0,
+      "confidence": "high",
+      "prediction": "imminent_launch",
+      "pattern": "breakout",
+      "washout_quality": "sufficient",
+      "rationale": "<≤ 200 字的核心判断>",
+      "dimension_scores": {
+        "washout": 60,
+        "pattern": 60,
+        "capital": 60,
+        "sector": 60,
+        "historical": 60,
+        "risk": 30
+      },
+      "key_evidence": [
+        {
+          "field": "<必须是输入字段名，如 base_days / vol_ratio_5d / ma60>",
+          "value": 0,
+          "unit": "<亿/万/%/日/次/无>",
+          "interpretation": "<对该数值的简短解读>"
+        }
+      ],
+      "next_session_watch": ["<次日需要观察的 1-4 个关键点>"],
+      "invalidation_triggers": ["<会让预测失效的 1-4 个触发条件>"],
+      "risk_flags": [],
+      "missing_data": []
+    }
+  ]
+}
+
+【字段值约束】
+- rank:               本批内 1..N 连续唯一整数（不可重复、不可跳号）
+- launch_score:       0–100 浮点数
+- confidence:         "high" / "medium" / "low" 三选一
+- prediction:         "imminent_launch" / "watching" / "not_yet" 三选一
+- pattern:            "breakout" / "consolidation_break" / "first_wave" / "second_leg" / "unclear" 五选一
+- washout_quality:    "sufficient" / "partial" / "insufficient" / "unclear" 四选一
+- dimension_scores:   6 个维度（washout/pattern/capital/sector/historical/risk）每个 0–100 整数；不可省
+- key_evidence:       每只 1–5 条，每条 4 个字段不可省
+- next_session_watch / invalidation_triggers: 各 1–4 条字符串数组（不可为空）
+- risk_flags:         空数组或字符串数组，最多 5 条
+- missing_data:       数据缺失字段名数组（参见 data_unavailable）
+- 输入清单中每一只标的必须出现在 candidates 中，candidate_id 与输入完全一致。
+"""
+
+
+# v0.3.0 P0-5 — concatenate the few-shot examples block onto the system prompt
+# so all batches see the same anchoring scale across LLM providers / model
+# upgrades. The few-shot adds ~800–1000 tokens; well within the 200K budget.
+VA_TREND_SYSTEM = _VA_TREND_SYSTEM_BASE + VA_TREND_FEWSHOT
+
+
+def va_trend_user_prompt(
+    *,
+    trade_date: str,
+    next_trade_date: str,
+    batch_no: int,
+    batch_total: int,
+    candidates: list[dict[str, Any]],
+    market_summary: dict[str, Any],
+    sector_strength_source: str,
+    sector_strength_data: dict[str, Any],
+    data_unavailable: list[str],
+) -> str:
+    """Render the走势分析 user prompt for one batch."""
+    return (
+        f"trade_date     = {trade_date}\n"
+        f"next_trade_date= {next_trade_date}\n"
+        f"batch_no       = {batch_no}\n"
+        f"batch_total    = {batch_total}\n"
+        f"本批候选股 = {len(candidates)} 只\n"
+        f"全局 data_unavailable = {data_unavailable}\n\n"
+        "【市场摘要】\n"
+        + json.dumps(market_summary, ensure_ascii=False, indent=2)
+        + "\n\n【板块强度摘要】\n"
+        f"sector_strength_source = {sector_strength_source}\n"
+        "sector_strength_data = "
+        + json.dumps(sector_strength_data, ensure_ascii=False, indent=2)
+        + "\n\n【候选清单】（每只含异动入池后的追踪历史聚合 + 整理期/洗盘指标 + 资金摘要）\n"
+        + json.dumps(candidates, ensure_ascii=False, indent=2)
+        + "\n\n请对本批次每一只候选股输出 VATrendCandidate；"
+        "candidate_id 与输入一一对应；rank 在本批内唯一且 1..N 连续；"
+        "rationale ≤ 200 字。\n"
+    )

deeptrade/strategies_builtin/volume_anomaly/volume_anomaly/prompts_examples.py

@@ -0,0 +1,84 @@
+"""Few-shot anchoring examples for the volume-anomaly trend-prediction stage.
+
+Two synthetic candidates are wired into ``VA_TREND_SYSTEM`` (see prompts.py)
+to anchor LLMs onto a consistent ``launch_score`` scale and to demonstrate the
+expected ``key_evidence`` field-citation discipline.
+
+CRITICAL CONSTRAINT: every ``"field": "<X>"`` referenced below MUST appear in
+either the ``hits`` rows produced by ``screen_anomalies`` or the candidate
+rows produced by ``_build_candidate_row``. The
+``test_prompt_consistency.py`` test enforces this — any field rename in
+``data.py`` will turn that test red, preventing silent prompt-schema drift.
+"""
+
+from __future__ import annotations
+
+VA_TREND_FEWSHOT = """\
+
+【参考示例】（仅展示判断尺度与字段引用规范，不是输入的一部分）
+
+示例 A — 教科书式 VCP 缩量充分洗盘后放量突破
+{
+  "candidate_id": "000XXX.SZ",
+  "ts_code": "000XXX.SZ",
+  "name": "示例 A",
+  "rank": 1,
+  "launch_score": 78,
+  "confidence": "high",
+  "prediction": "imminent_launch",
+  "pattern": "breakout",
+  "washout_quality": "sufficient",
+  "rationale": "整理 24 日，回撤 12%，波动率分位 0.08；T 日放量站上 MA20；moneyflow 5 日累计净流入；板块为当日主线。三维 VCP 齐降 + 主线共振 → 启动概率高。",
+  "dimension_scores": {
+    "washout": 80,
+    "pattern": 75,
+    "capital": 70,
+    "sector": 75,
+    "historical": 60,
+    "risk": 25
+  },
+  "key_evidence": [
+    {"field": "base_days",                  "value": 24,    "unit": "日",  "interpretation": "整理周期较长，洗盘相对充分"},
+    {"field": "atr_10d_quantile_in_60d",    "value": 0.08,  "unit": "无",  "interpretation": "波动率处于近 60 日 8% 分位，VCP 收敛"},
+    {"field": "anomaly_vol_ratio_5d",       "value": 2.4,   "unit": "倍",  "interpretation": "异动当日相对前 5 日放量 2.4 倍"},
+    {"field": "dist_to_250d_high_pct",      "value": -3.5,  "unit": "%",   "interpretation": "距 250 日新高仅 3.5%，临近年线突破口"},
+    {"field": "alpha_20d_pct",              "value": 8.2,   "unit": "%",   "interpretation": "近 20 日相对沪深 300 跑赢 8.2%，主线领涨"}
+  ],
+  "next_session_watch": ["次日开盘是否站稳 MA10", "板块强度是否延续主线地位"],
+  "invalidation_triggers": ["收盘跌破 MA10 且 moneyflow 转为净流出", "板块跌出主线前 5"],
+  "risk_flags": [],
+  "missing_data": []
+}
+
+示例 B — 高位长上影线 + 资金外流
+{
+  "candidate_id": "600YYY.SH",
+  "ts_code": "600YYY.SH",
+  "name": "示例 B",
+  "rank": 2,
+  "launch_score": 22,
+  "confidence": "medium",
+  "prediction": "not_yet",
+  "pattern": "unclear",
+  "washout_quality": "insufficient",
+  "rationale": "异动当天上影线占振幅 0.42，body_ratio 仅 0.55；过去 60 日已两次涨停且距 120 日新高 < 1%；moneyflow 5 日累计净流出。高位放量诱多概率高。",
+  "dimension_scores": {
+    "washout": 30,
+    "pattern": 35,
+    "capital": 25,
+    "sector": 50,
+    "historical": 30,
+    "risk": 75
+  },
+  "key_evidence": [
+    {"field": "upper_shadow_ratio",         "value": 0.42,  "unit": "无",  "interpretation": "上影线偏长，疑似冲高回落"},
+    {"field": "prior_limit_up_count_60d",   "value": 2,     "unit": "次",  "interpretation": "近 60 日已 2 次涨停，浪型偏后"},
+    {"field": "dist_to_120d_high_pct",      "value": -0.8,  "unit": "%",   "interpretation": "贴近 120 日新高，套牢盘压力大"},
+    {"field": "alpha_20d_pct",              "value": -3.0,  "unit": "%",   "interpretation": "近 20 日相对沪深 300 跑输，板块支撑弱"}
+  ],
+  "next_session_watch": ["放量回踩 MA10 是否守住", "moneyflow 是否转为净流入"],
+  "invalidation_triggers": ["再放量阴线击穿 MA20"],
+  "risk_flags": ["high_bull_trap_risk", "late_stage_pattern"],
+  "missing_data": []
+}
+"""