评分改为±100分制；库存图y轴倒置；月差序列过滤远月前12月幽灵数据

spread_from_db.py

@@ -172,6 +172,24 @@ def _extract_year_prefix(contract: str) -> int | None:
         return None
 
 
+def _contract_year(contract: str) -> int | None:
+    """从合约代码推断完整年份。
+    上期所/大商所/中金所 4位数字: RB2610 → 2000+26=2026
+    郑商所 3位数字: MA609 → 2020+6=2026（当前十年2020s）
+    """
+    digits = "".join(c for c in contract if c.isdigit())
+    if len(digits) < 3:
+        return None
+    try:
+        prefix = int(digits[:-2])
+        if len(digits) == 3:  # 郑商所3位数字，前缀1位
+            return 2020 + prefix
+        else:  # 4位数字，前缀2位
+            return 2000 + prefix
+    except (ValueError, IndexError):
+        return None
+
+
 def _year_gap(near: str, far: str) -> int | None:
     """计算近月/远月合约的年份间隔，如 (SC2608, SC2609) → 0, (EG2609, EG2701) → 1。"""
     ny = _extract_year_prefix(near)
@@ -254,9 +272,23 @@ def build_historical_spreads(
                 spread_series = spread_series[spread_series.index >= f"{start_year}-01-01"]
                 if spread_series.dropna().empty:
                     continue
-                # 推断年份：取序列中间日期的年份
-                mid_idx = len(spread_series) // 2
-                year_val = int(spread_series.index[mid_idx].year)
+                # 只保留远月合约到期前约12个月开始的数据
+                # 远月合约 c2 交割年月 = (Y, far_month)，保留 date(Y-1, far_month, 1) 之后的数据
+                _far_year = _contract_year(c2)
+                if _far_year is not None:
+                    try:
+                        _far_start = pd.Timestamp(year=_far_year - 1, month=far_month, day=1)
+                        spread_series = spread_series[spread_series.index >= _far_start]
+                    except (ValueError, TypeError):
+                        pass
+                    if spread_series.dropna().empty:
+                        continue
+                # 推断年份：从近月合约代码提取完整年份
+                year_val = _contract_year(c1)
+                if year_val is None:
+                    # 降级方案：使用序列中间日期的年份
+                    mid_idx = len(spread_series) // 2
+                    year_val = int(spread_series.index[mid_idx].year)
                 hist_series_list.append({
                     "pair_label": f"{c1} - {c2}",
                     "series": spread_series,
@@ -336,16 +368,19 @@ def _builtin_band_score(value: float, history: pd.Series, day: int, inverse: boo
     low = mid - (mid - q_low) * width_scale
     high = mid + (q_high - mid) * width_scale
     if high == low:
-        score = 50.0
-    elif inverse:
-        score = 100.0 if value <= low else (0.0 if value >= high else 100.0 - (value - low) / (high - low) * 100)
+        raw_score = 100.0 if value >= high else -100.0
+    elif value >= high:
+        raw_score = 100.0
+    elif value <= low:
+        raw_score = -100.0
     else:
-        score = 100.0 if value >= high else (0.0 if value <= low else (value - low) / (high - low) * 100)
+        raw_score = 200.0 * (value - low) / (high - low) - 100.0
+    score = -raw_score if inverse else raw_score
     below = float((sample < value).sum())
     equal = float((sample == value).sum())
     raw_pct = 100.0 * (below + 0.5 * equal) / len(sample)
     return {
-        "score": round(max(0.0, min(100.0, score)), 1),
+        "score": round(max(-100.0, min(100.0, score)), 1),
         "raw_percentile": round(raw_pct, 1),
         "band_low": round(low, 4),
         "band_high": round(high, 4),