docs: sync with time-decay sigma, rawOb precision, MGM filter, peak-hour arch

README.md

@@ -19,7 +19,13 @@ The system automatically tracks the historical performance of weather models (EC
 
 Automatically computes the probability for each possible WU settlement integer using a Gaussian distribution fitted to the ensemble forecast:
 
-- **Method**: Derives standard deviation (σ) from the 51-member ensemble P10/P90, centers the distribution (μ) on a weighted average of DEB/multi-model median (70%) and ensemble median (30%).
+- **Distribution Center μ**: Weighted average of DEB/multi-model median (70%) and ensemble median (30%). Auto-corrects upward when actual METAR max exceeds μ and is still rising.
+- **Standard Deviation σ**: Derived from the 51-member ensemble P10/P90 (σ = (P90-P10) / 2.56).
+- **Time Decay**: σ dynamically narrows based on the current time relative to the predicted peak window:
+  - Before peak: σ × 1.0 (maximum uncertainty)
+  - During peak window: σ × 0.7 (settling)
+  - After peak: σ × 0.3 (outcome mostly determined)
+- **Observed Floor**: Temperatures below the current METAR max WU value are automatically excluded (can't go back down).
 - **Interval Integration**: Integrates over each WU rounding interval [N-0.5, N+0.5) to compute the probability of settling at integer N.
 - **Display**: `🎲 Settlement Probability (μ=3.7): 4°C [3.5~4.5) 68% | 3°C [2.5~3.5) 32%`
 
@@ -27,15 +33,18 @@ Automatically computes the probability for each possible WU settlement integer u
 
 Feeds wind speed, wind direction, cloud cover, solar radiation, and METAR trend data into LLaMA 70B:
 
-- **Logical Reasoning**: Uses 2-3 sentences to deeply analyze airport dynamics—whether conditions promote or inhibit warming, and whether the forecast can be reached.
-- **Market Call**: Explicitly states the expected peak time window and the specific temperature betting range. Calls "dead market" when cooling is confirmed.
+- **Logical Reasoning**: 2-3 sentences analyzing airport dynamics, explicitly referencing Open-Meteo forecast and DEB blended values as benchmarks.
+- **Time Awareness**: Analysis considers how much time remains until the predicted peak, judging remaining warming potential.
+- **Market Call**: Explicitly states the expected peak time window and specific temperature betting range. Calls "dead market" when cooling is confirmed.
 - **Confidence Score**: Quantitative 1-10 confidence rating.
 - **High Availability**: Built-in auto-retry + fallback model degradation (70B → 8B) to withstand Groq API 500/503 outages.
 
 ### 4. ⏱️ Real-time Airport Observations (Zero-Cache METAR)
 
+- **Precise Timing**: Extracts actual observation time from raw METAR text (`rawOb`), not the API's rounded `reportTime`. Accurate to the minute.
 - **Live Passthrough**: Bypasses CDN caching via dynamic headers to obtain first-hand METAR reports.
 - **Settlement Warning**: Automatically calculates the Wunderground settlement boundary (X.5 rounding line).
+- **Anomaly Filtering**: Automatically filters out -9999 sentinel values from sources like MGM to prevent garbage data in output.
 
 ### 5. 📈 Historical Data Collection
 
@@ -97,13 +106,15 @@ graph TD
         Bot --> Collector[WeatherDataCollector]
         Collector --> OM[Open-Meteo Forecast/Ensemble]
         Collector --> MM[Multi-Model ECMWF/GFS/ICON/GEM/JMA]
-        Collector --> METAR[Live Airport METAR]
+        Collector --> METAR["Live Airport METAR (rawOb precise time)"]
     end
 
     subgraph Algorithm Layer
+        Collector --> Peak[Peak Hour Prediction]
         Collector --> DEB[DEB Dynamic Weighting]
         DEB --> DB[(daily_records Database)]
-        Collector --> Prob[Gaussian Probability Engine]
+        Peak --> Prob[Gaussian Probability Engine]
+        Collector --> Prob
         Collector --> Logic[Settlement Boundary / Trend Analysis]
     end
 
@@ -122,11 +133,12 @@ graph TD
 
 ## 💡 Trading Tips
 
-1. **Watch Settlement Probability**: The probability engine is math-based and more objective than AI subjective judgment. When one temperature has > 65% probability, the direction is relatively clear.
-2. **Reference DEB Blended Value**: When models diverge, the DEB corrected value is usually more reliable than any single forecast.
-3. **Observe AI Confidence**: A score below 5 indicates high uncertainty—consider staying on the sidelines.
-4. **Watch Settlement Boundaries**: When the observed high is near X.5, be wary of rounding jumps during WU settlements.
-5. **Distribution Center μ**: The μ value shown in the probability display represents the algorithm's expected most likely actual high temperature—compare it directly with the Polymarket odds.
+1. **Watch Settlement Probability**: The probability engine is math-based and more objective than AI judgment. When one temperature has > 65% probability, the direction is relatively clear.
+2. **Observe Time Decay**: Probabilities auto-lock as time progresses. After peak hours, the engine narrows σ dramatically, concentrating results around the observed max.
+3. **Reference DEB Blended Value**: When models diverge, the DEB corrected value is usually more reliable than any single forecast.
+4. **Observe AI Confidence**: A score below 5 indicates high uncertainty—consider staying on the sidelines.
+5. **Watch Settlement Boundaries**: When the observed high is near X.5, be wary of rounding jumps during WU settlements.
+6. **Distribution Center μ**: The μ value shown in the probability display represents the algorithm's expected most likely actual high temperature—compare it directly with the Polymarket odds.
 
 ---
 

README_ZH.md

@@ -19,7 +19,13 @@ PolyWeather 是一款专为 **Polymarket** 等预测市场打造的天气分析
 
 基于集合预报的正态分布拟合，自动计算每个 WU 结算整数温度的概率：
 
-- **原理**：用 51 成员集合预报的 P10/P90 反推高斯分布标准差（σ），以 DEB/多模型中位数（70% 权重）和集合中位数（30%）的加权均值为分布中心（μ）。
+- **分布中心 μ**：以 DEB/多模型中位数（70% 权重）和集合中位数（30%）的加权均值为中心。当实测最高温已超过 μ 且仍在升温时，自动向上修正。
+- **标准差 σ**：由 51 成员集合预报的 P10/P90 反推（σ = (P90-P10) / 2.56）。
+- **时间衰减**：σ 会根据当前时间与预计最热时段的关系动态收窄：
+  - 峰值前：σ × 1.0（最大不确定性）
+  - 峰值窗口内：σ × 0.7（正在定型）
+  - 峰值后：σ × 0.3（大局已定）
+- **实测过滤**：已实测到的最高温对应 WU 值以下的候选温度自动排除（不可能回退）。
 - **区间积分**：对每个 WU 取整区间 [N-0.5, N+0.5) 做正态积分，得到结算落入该整数的概率。
 - **展示格式**：`🎲 结算概率 (μ=3.7)：4°C [3.5~4.5) 68% | 3°C [2.5~3.5) 32%`
 
@@ -27,15 +33,18 @@ PolyWeather 是一款专为 **Polymarket** 等预测市场打造的天气分析
 
 将风速、风向、云量、太阳辐射、METAR 趋势等全部投喂给 LLaMA 70B 大模型：
 
-- **逻辑推演**：用 2-3 句话深度分析机场动力因子对升温的促进/阻碍，判断最终能否达到预报值。
+- **逻辑推演**：用 2-3 句话深度分析机场动力因子对升温的促进/阻碍，明确引用 Open-Meteo 预报值和 DEB 融合值作为对标。
+- **时间感知**：分析中会考虑当前时间距预计最热时段的距离，判断剩余升温空间。
 - **盘口判定**：明确给出预计最热时段和具体博弈温度区间。降温确认后直接给出死盘结论。
 - **置信度评分**：1-10 分量化置信度参考。
 - **高可用保障**：内置自动重试 + 备用模型降级机制（70B → 8B），抵御 Groq API 503/500 故障。
 
 ### 4. ⏱️ 实时机场观测 (Zero-Cache METAR)
 
+- **精确时间**：从 METAR 原始报文 (`rawOb`) 中提取真实观测时间，而非 API 取整后的 `reportTime`，精确到分钟。
 - **实时穿透**：通过动态请求头绕过 CDN 缓存，获取机场第一手 METAR 报文。
 - **结算预警**：自动计算 Wunderground 结算边界（X.5 进位线），提醒潜在波动。
+- **异常过滤**：自动过滤 MGM 等数据源的 -9999 哨兵值，避免垃圾数据污染输出。
 
 ### 5. 📈 历史数据采集
 
@@ -97,13 +106,15 @@ graph TD
         Bot --> Collector[WeatherDataCollector]
         Collector --> OM[Open-Meteo 预报/集合]
         Collector --> MM[多模型 ECMWF/GFS/ICON/GEM/JMA]
-        Collector --> METAR[机场实时 METAR]
+        Collector --> METAR["机场实时 METAR (rawOb精确时间)"]
     end
 
     subgraph 算法层
+        Collector --> Peak[峰值时段预测]
         Collector --> DEB[DEB 动态权重融合]
         DEB --> DB[(daily_records 数据库)]
-        Collector --> Prob[正态分布概率引擎]
+        Peak --> Prob[正态分布概率引擎]
+        Collector --> Prob
         Collector --> Logic[结算边界 / 趋势分析]
     end
 
@@ -123,10 +134,11 @@ graph TD
 ## 💡 交易提示
 
 1. **紧盯结算概率**：概率引擎基于数学模型计算，比 AI 的主观判断更客观。当某个温度概率 > 65%，说明方向较为明确。
-2. **参考 DEB 融合值**：当多模型分歧较大时，DEB 的修正值通常比单一预报更具参考意义。
-3. **观察 AI 置信度**：置信度低于 5 分时，说明当前气象环境处于高度不确定状态，建议观望。
-4. **注意结算边界**：实测最高温接近 X.5 时，需警惕 Wunderground 结算时的进位跳动。
-5. **分布中心 μ**：概率展示中的 μ 值代表算法预期的最可能实际最高温，可与 Polymarket 盘口价格直接对比。
+2. **关注时间衰减**：概率会随时间推进自动锁定。峰值过后，概率引擎自动缩窄 σ，结果高度集中在实测最高值附近。
+3. **参考 DEB 融合值**：当多模型分歧较大时，DEB 的修正值通常比单一预报更具参考意义。
+4. **观察 AI 置信度**：置信度低于 5 分时，说明当前气象环境处于高度不确定状态，建议观望。
+5. **注意结算边界**：实测最高温接近 X.5 时，需警惕 Wunderground 结算时的进位跳动。
+6. **分布中心 μ**：概率展示中的 μ 值代表算法预期的最可能实际最高温，可与 Polymarket 盘口价格直接对比。
 
 ---