feat: add historical report analysis for same market

- Add ReportHistoryService to find and summarize historical reports
- Integrate historical context into LLM analysis prompts
- Display market description in LLM input
- Show referenced historical report count in generated reports

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

src/main.py

@@ -105,7 +105,11 @@ class WhaleWatcher:
         decision = await self.llm_analyzer.analyze_whale_trade(whale_trade)
 
         # Print the full report (includes analysis + decision summary)
-        full_report = self.llm_analyzer.format_full_report(whale_trade, decision)
+        full_report = self.llm_analyzer.format_full_report(
+            whale_trade,
+            decision,
+            historical_report_count=self.llm_analyzer.last_historical_report_count,
+        )
         print(full_report)
 
         # Save report to file

src/prompts/whale_analyzer.py

@@ -69,22 +69,32 @@ class WhaleAnalyzerPrompts:
 **重要原则**：
 - **务必使用 Google Search！** 不要仅依赖你的历史知识
 - **重视交易者历史记录！** 这是判断交易者专业性的关键依据
+- **如果有历史报告，务必结合历史报告进行综合分析！** 这能帮助你了解该市场的交易模式
 - 关注过去24-72小时的最新动态
 - 如果搜索不到支持信息，内幕交易可能性应该降低
 - 信心不足时建议观望（HOLD）"""
 
     @staticmethod
-    def analyze_whale_trade(trade_context: str) -> str:
+    def analyze_whale_trade(trade_context: str, historical_context: str = "") -> str:
         """
         Get the prompt for analyzing a whale trade.
 
         Args:
             trade_context: Formatted trade context from AnomalyDetector
+            historical_context: Formatted historical reports context (optional)
 
         Returns:
             Complete prompt for LLM
         """
+        history_section = ""
+        if historical_context:
+            history_section = f"""
+{historical_context}
+
+---
+"""
         return f"""{trade_context}
+{history_section}
 
 ---
 
@@ -142,6 +152,27 @@ class WhaleAnalyzerPrompts:
 
 ---
 
+## 第三点五步：历史报告综合分析（如有历史报告）
+
+如果上文提供了历史报告，请进行以下分析：
+
+### 3.5.1 交易方向对比
+- 历史报告中的交易方向（BUY/SELL）与当前交易是否一致？
+- 如果方向一致，这可能表明多个交易者对同一结果有信心
+- 如果方向相反，需要分析原因（时间变化、新信息、不同交易者的判断）
+
+### 3.5.2 历史内幕交易评估
+- 历史报告对内幕交易的判断如何？
+- 如果历史报告也认为是内幕交易，这增强了当前交易的可信度
+- 结合历史报告的证据和当前搜索结果进行综合判断
+
+### 3.5.3 趋势演变分析
+- 该市场的交易模式是否有变化？
+- 价格从历史报告到现在有何变动？
+- 鲸鱼交易的频率和规模是否在增加？
+
+---
+
 ## 第四步：内幕交易判定
 
 ### 4.1 内幕交易可能性评估

src/services/anomaly_detector.py

@@ -235,8 +235,9 @@ class AnomalyDetector:
 - **隐含概率**: 交易者认为结果发生的概率约为 {context['implied_probability']:.2%}
 - **信心程度**: {context['conviction_level']}
 {ranking_str}{history_str}
-### 市场状态
+### 市场信息
 - **市场问题**: {context['market_question']}
+- **市场描述**: {whale_trade.market_description or 'N/A'}
 - **市场状态**: {context['market_state']}
 - **当前赔率**:
 {prices_str}

src/services/llm_analyzer.py

@@ -13,6 +13,7 @@ from src.config import get_settings
 from src.models.trade import WhaleTrade
 from src.models.decision import LLMDecision, TradeRecommendation, TradeAction, TraderCredibility
 from src.services.anomaly_detector import AnomalyDetector
+from src.services.report_history import ReportHistoryService
 from src.prompts.whale_analyzer import WhaleAnalyzerPrompts
 
 logger = logging.getLogger(__name__)
@@ -35,6 +36,15 @@ class LLMAnalyzer:
 
         self.anomaly_detector = AnomalyDetector()
         self.prompts = WhaleAnalyzerPrompts()
+        self.report_history = ReportHistoryService()
+
+        # Track the number of historical reports used in the last analysis
+        self._last_historical_report_count = 0
+
+    @property
+    def last_historical_report_count(self) -> int:
+        """Get the number of historical reports used in the last analysis."""
+        return self._last_historical_report_count
 
     def _extract_json_from_response(self, response: str) -> Optional[dict]:
         """
@@ -132,9 +142,21 @@ class LLMAnalyzer:
         # Format trade context for LLM
         trade_context = self.anomaly_detector.format_for_llm(whale_trade)
 
+        # Find and format historical reports for the same market
+        historical_context = ""
+        historical_reports = self.report_history.find_historical_reports(
+            whale_trade.market_question,
+            similarity_threshold=0.5,
+            max_reports=5,
+        )
+        self._last_historical_report_count = len(historical_reports)
+        if historical_reports:
+            historical_context = self.report_history.format_historical_context(historical_reports)
+            logger.info(f"Found {len(historical_reports)} historical reports for market: {whale_trade.market_question}")
+
         # Build prompt (Gemini uses single prompt with system instruction)
         system_prompt = self.prompts.system_prompt()
-        user_prompt = self.prompts.analyze_whale_trade(trade_context)
+        user_prompt = self.prompts.analyze_whale_trade(trade_context, historical_context)
         full_prompt = f"{system_prompt}\n\n---\n\n{user_prompt}"
 
         try:
@@ -187,13 +209,19 @@ class LLMAnalyzer:
                 ),
             )
 
-    def format_full_report(self, whale_trade: WhaleTrade, decision: LLMDecision) -> str:
+    def format_full_report(
+        self,
+        whale_trade: WhaleTrade,
+        decision: LLMDecision,
+        historical_report_count: int = 0,
+    ) -> str:
         """
         Format a complete analysis report with trade info, analysis, and decision.
 
         Args:
             whale_trade: The whale trade
             decision: The LLM decision
+            historical_report_count: Number of historical reports used in analysis
 
         Returns:
             Formatted report string
@@ -242,12 +270,17 @@ class LLMAnalyzer:
             pnl_str = f"${tr.pnl:,.2f}" if tr.pnl else "N/A"
             trader_ranking_str = f"| **交易者排名** | {rank_str} (PnL: {pnl_str}) |"
 
+        # Historical reports info
+        historical_info = ""
+        if historical_report_count > 0:
+            historical_info = f"\n**参考历史报告**: {historical_report_count} 份 (已综合分析)"
+
         report = f"""
 {'='*70}
 # 🐋 鲸鱼交易分析报告
 {'='*70}
 
-**生成时间**: {datetime.utcnow().strftime('%Y-%m-%d %H:%M:%S')} UTC
+**生成时间**: {datetime.utcnow().strftime('%Y-%m-%d %H:%M:%S')} UTC{historical_info}
 
 ## 交易摘要
 

src/services/report_history.py

@@ -0,0 +1,257 @@
+"""Report history service - finds and summarizes historical reports for the same market."""
+import os
+import re
+from pathlib import Path
+from typing import List, Optional
+from dataclasses import dataclass
+from datetime import datetime
+
+
+@dataclass
+class HistoricalReport:
+    """Represents a historical report for the same market."""
+    filepath: str
+    filename: str
+    timestamp: datetime
+    side: str
+    amount_usd: float
+    market_name: str
+    content: str
+
+    @property
+    def summary(self) -> str:
+        """Extract key decision info from the report."""
+        # Try to extract the JSON decision section
+        json_pattern = r'```json\s*([\s\S]*?)```'
+        matches = re.findall(json_pattern, self.content)
+
+        decision_info = ""
+        for match in matches:
+            try:
+                import json
+                data = json.loads(match.strip())
+                action = data.get('action', 'N/A')
+                confidence = data.get('confidence', 'N/A')
+                insider_likelihood = data.get('insider_trading_likelihood', 'N/A')
+                reasoning = data.get('reasoning', 'N/A')
+
+                if isinstance(confidence, (int, float)):
+                    confidence = f"{confidence:.0%}"
+                if isinstance(insider_likelihood, (int, float)):
+                    insider_likelihood = f"{insider_likelihood:.0%}"
+
+                decision_info = f"""
+- **操作建议**: {action}
+- **信心程度**: {confidence}
+- **内幕交易可能性**: {insider_likelihood}
+- **决策理由**: {reasoning}"""
+                break
+            except (json.JSONDecodeError, KeyError):
+                continue
+
+        return f"""**报告时间**: {self.timestamp.strftime('%Y-%m-%d %H:%M:%S')}
+**交易方向**: {self.side}
+**交易金额**: ${self.amount_usd:,.2f} USDC
+{decision_info}"""
+
+
+class ReportHistoryService:
+    """Service for finding and managing historical reports."""
+
+    def __init__(self, reports_dir: Optional[Path] = None):
+        """
+        Initialize the report history service.
+
+        Args:
+            reports_dir: Path to the reports directory. Defaults to project's reports dir.
+        """
+        if reports_dir is None:
+            self.reports_dir = Path(__file__).parent.parent.parent / "reports"
+        else:
+            self.reports_dir = reports_dir
+
+    def _parse_filename(self, filename: str) -> Optional[dict]:
+        """
+        Parse a report filename to extract metadata.
+
+        Filename format: {timestamp}_{side}_{amount}USD_{market_name}.md
+        Example: 20260105_150505_BUY_7520USD_Trump_out_as_President_before_2027.md
+
+        Args:
+            filename: The filename to parse
+
+        Returns:
+            Dictionary with parsed metadata or None if parsing fails
+        """
+        if not filename.endswith('.md'):
+            return None
+
+        # Pattern: timestamp_side_amountUSD_market_name.md
+        pattern = r'^(\d{8}_\d{6})_(BUY|SELL)_(\d+)USD_(.+)\.md$'
+        match = re.match(pattern, filename)
+
+        if not match:
+            return None
+
+        timestamp_str, side, amount_str, market_name = match.groups()
+
+        try:
+            timestamp = datetime.strptime(timestamp_str, '%Y%m%d_%H%M%S')
+            amount = float(amount_str)
+        except ValueError:
+            return None
+
+        return {
+            'timestamp': timestamp,
+            'side': side,
+            'amount_usd': amount,
+            'market_name': market_name,
+        }
+
+    def _sanitize_market_name(self, market_question: str, max_length: int = 50) -> str:
+        """
+        Sanitize market question for matching with filenames.
+
+        Args:
+            market_question: The market question to sanitize
+            max_length: Maximum length of the sanitized name
+
+        Returns:
+            Sanitized market name
+        """
+        # Remove special characters, keep alphanumeric and spaces
+        sanitized = re.sub(r'[^\w\s-]', '', market_question)
+        # Replace spaces with underscores
+        sanitized = re.sub(r'\s+', '_', sanitized)
+        # Truncate if too long
+        return sanitized[:max_length]
+
+    def _calculate_similarity(self, name1: str, name2: str) -> float:
+        """
+        Calculate similarity between two market names.
+
+        Uses a simple word overlap method for fuzzy matching.
+
+        Args:
+            name1: First market name (sanitized)
+            name2: Second market name (from filename)
+
+        Returns:
+            Similarity score between 0 and 1
+        """
+        # Convert to lowercase and split into words
+        words1 = set(name1.lower().replace('_', ' ').split())
+        words2 = set(name2.lower().replace('_', ' ').split())
+
+        # Remove common stop words
+        stop_words = {'the', 'a', 'an', 'is', 'are', 'will', 'by', 'to', 'of', 'in', 'on', 'for'}
+        words1 = words1 - stop_words
+        words2 = words2 - stop_words
+
+        if not words1 or not words2:
+            return 0.0
+
+        # Calculate Jaccard similarity
+        intersection = len(words1 & words2)
+        union = len(words1 | words2)
+
+        return intersection / union if union > 0 else 0.0
+
+    def find_historical_reports(
+        self,
+        market_question: str,
+        similarity_threshold: float = 0.5,
+        max_reports: int = 5,
+    ) -> List[HistoricalReport]:
+        """
+        Find historical reports for the same or similar market.
+
+        Args:
+            market_question: The market question to search for
+            similarity_threshold: Minimum similarity score to include a report
+            max_reports: Maximum number of reports to return
+
+        Returns:
+            List of HistoricalReport objects, sorted by timestamp (newest first)
+        """
+        if not self.reports_dir.exists():
+            return []
+
+        sanitized_question = self._sanitize_market_name(market_question)
+        matching_reports = []
+
+        for filename in os.listdir(self.reports_dir):
+            metadata = self._parse_filename(filename)
+            if metadata is None:
+                continue
+
+            # Calculate similarity between market names
+            similarity = self._calculate_similarity(
+                sanitized_question,
+                metadata['market_name']
+            )
+
+            if similarity >= similarity_threshold:
+                filepath = self.reports_dir / filename
+                try:
+                    with open(filepath, 'r', encoding='utf-8') as f:
+                        content = f.read()
+                except Exception:
+                    continue
+
+                report = HistoricalReport(
+                    filepath=str(filepath),
+                    filename=filename,
+                    timestamp=metadata['timestamp'],
+                    side=metadata['side'],
+                    amount_usd=metadata['amount_usd'],
+                    market_name=metadata['market_name'],
+                    content=content,
+                )
+                matching_reports.append((similarity, report))
+
+        # Sort by similarity (descending) then by timestamp (descending)
+        matching_reports.sort(key=lambda x: (x[0], x[1].timestamp), reverse=True)
+
+        # Return only the reports (without similarity scores)
+        return [report for _, report in matching_reports[:max_reports]]
+
+    def format_historical_context(
+        self,
+        reports: List[HistoricalReport],
+    ) -> str:
+        """
+        Format historical reports into a context string for LLM.
+
+        Args:
+            reports: List of historical reports
+
+        Returns:
+            Formatted string for LLM context
+        """
+        if not reports:
+            return ""
+
+        context = f"""
+### 历史报告分析 (共 {len(reports)} 份历史报告)
+
+**重要**: 该市场之前已经生成过分析报告，请结合历史报告进行综合分析。
+
+"""
+        for i, report in enumerate(reports, 1):
+            context += f"""
+---
+#### 历史报告 {i}
+{report.summary}
+---
+"""
+
+        context += """
+**综合分析要点**:
+1. 对比历史报告中的交易方向和当前交易方向，分析是否有趋势变化
+2. 对比历史的内幕交易可能性评估，判断该市场是否持续有异常交易
+3. 如果多份报告都指向同一方向，这可能加强信号的可信度
+4. 如果报告方向相反，需要分析原因并给出更审慎的判断
+5. 考虑时间因素：越近期的报告越有参考价值
+"""
+        return context