feat: 预测性成本智能升级 — 历史回归弹性校准 + 预测偏差告警

升级1: 宏观敏感性弹性历史校准
- 内置宏观历史数据(oil/usd/cpi 2026-01~07月度)
- 变化率弹性: 同period匹配KPI历史vs因素历史算弹性
- 合理性校验: |弹性|超出[0.01,0.5]视为噪声回退规则(诚实标注)

升级2: 预测偏差告警闭环
- 新表 kpi_forecast_log(预测历史)+模型KpiForecastLog
- 预测时落库(同KPI同预测期覆盖)
- alert_rules 支持 rule_type=forecast_deviation(threshold_pct)
- POST /alert-rules/run-forecast-deviation: 预测vs实际偏差>阈值生成预警(去重, 超2倍阈值红色)
- 端到端验证: 模拟实际500vs预测399.55→偏差20.1%>5%→红色预警生成

回归: pytest 40 passed(predict+alerts)

backend/app/api/alert_rules.py

@@ -107,7 +107,7 @@ def create_alert_rule(data: dict, db: Session = Depends(get_db), user=Depends(re
     kpi = db.query(KPIDefinition).filter(KPIDefinition.id == kpi_id).first()
     if not kpi:
         raise HTTPException(404, "KPI不存在")
-    if rule_type not in ("static", "dynamic", "trend_up", "trend_down"):
+    if rule_type not in ("static", "dynamic", "trend_up", "trend_down", "forecast_deviation"):  # 升级2b: 预测偏差
         raise HTTPException(400, f"不支持的规则类型: {rule_type}")
 
     rule = AlertRule(
@@ -619,3 +619,78 @@ def generate_alert_suggestions(db: Session = Depends(get_db)):
 
     db.commit()
     return {"message": f"已为{updated}条预警生成情景建议", "updated": updated}
+
+
+@router.post("/run-forecast-deviation")
+def run_forecast_deviation_check(
+    db: Session = Depends(get_db),
+    entity_id: int = Depends(get_entity_id),
+):
+    """预测偏差检查（升级2b, 2026-08-25）— alert_rules type=forecast_deviation
+    对每条偏差规则: 取最新预测log(kpi_forecast_log) vs 该期实际值(kpi_values)，
+    偏差 > threshold_pct → 生成/更新 pending 预警（去重）"""
+    from app.models import KpiForecastLog
+    rules = db.query(AlertRule).filter(
+        AlertRule.entity_id == entity_id,
+        AlertRule.rule_type == "forecast_deviation",
+        AlertRule.enabled == 1,
+    ).all()
+    if not rules:
+        return {"message": "无预测偏差规则，可先创建 rule_type=forecast_deviation 规则", "generated": 0}
+
+    generated = 0
+    for rule in rules:
+        try:
+            params = rule.params or {}
+            threshold = float(params.get("threshold_pct", 15))
+            # 最新预测
+            log = db.query(KpiForecastLog).filter(
+                KpiForecastLog.entity_id == entity_id,
+                KpiForecastLog.kpi_id == rule.kpi_id,
+            ).order_by(KpiForecastLog.created_at.desc()).first()
+            if not log or log.forecast_value is None:
+                continue
+            # 该预测期的实际值（同period匹配；兼容 2026-H1 等半年度）
+            actual = db.query(KPIValue).filter(
+                KPIValue.kpi_id == rule.kpi_id,
+                KPIValue.period == log.period,
+            ).order_by(KPIValue.id.desc()).first()
+            if not actual or not actual.actual_value:
+                continue
+            base = abs(actual.actual_value)
+            if base < 1e-9:
+                continue
+            deviation_pct = abs(log.forecast_value - actual.actual_value) / base * 100
+            if deviation_pct <= threshold:
+                continue
+            kpi = db.query(KPIDefinition).filter(KPIDefinition.id == rule.kpi_id).first()
+            kpi_label = f"{kpi.kpi_name}({kpi.kpi_code})" if kpi else f"KPI#{rule.kpi_id}"
+            alert_level = "red" if deviation_pct > threshold * 2 else "yellow"
+            alert_message = (
+                f"预测偏差 {deviation_pct:.1f}% > 阈值{threshold}%："
+                f"{kpi_label} 预测{log.period}={log.forecast_value}，实际={actual.actual_value}"
+            )
+            # 去重: 同KPI+period 已有 pending 偏差预警
+            existing = db.query(KPIAlert).filter(
+                KPIAlert.kpi_id == rule.kpi_id,
+                KPIAlert.alert_message.like(f"%预测偏差%{log.period}%"),
+                KPIAlert.status == "pending",
+            ).first()
+            if existing:
+                existing.alert_message = alert_message
+                existing.alert_level = alert_level
+            else:
+                db.add(KPIAlert(
+                    kpi_id=rule.kpi_id,
+                    kpi_value_id=actual.id,
+                    alert_level=alert_level,
+                    alert_message=alert_message,
+                    alert_type="forecast",
+                    status="pending",
+                ))
+            generated += 1
+        except Exception as e:
+            logger.error(f"预测偏差检查失败 rule_id={rule.id}: {e}")
+            continue
+    db.commit()
+    return {"message": f"预测偏差检查完成: {generated}条", "generated": generated}

backend/app/api/predict.py

@@ -719,7 +719,8 @@ def api_growth_quality(request: Request, data: dict):
 # ── KPI趋势预测（预测性成本智能 MVP） ────────────────────────────
 from app.utils.kpi_forecast_engine import (  # noqa: E402
     MODELS, forecast_kpi, forecast_finance_kpis,
-    MACRO_FACTORS, factor_sensitivity_for_kpi, adjusted_next_with_factor,
+    MACRO_FACTORS, factor_sensitivity_for_kpi, factor_sensitivity_with_history,
+    adjusted_next_with_factor, save_forecast_logs,
 )
 
 
@@ -758,6 +759,10 @@ def api_kpi_forecast_finance(
     if model not in MODELS:
         raise HTTPException(400, f"不支持的模型: {model}，可选: {'/'.join(MODELS)}")
     results = forecast_finance_kpis(entity_id, db, periods=periods, model=model)
+    try:
+        save_forecast_logs(entity_id, results, db, model=model)  # 升级2a: 预测落库(供偏差告警)
+    except Exception as e:
+        logger.warning(f"预测落库失败(不影响返回): {e}")
     return {
         "entity_id": entity_id,
         "model": model,
@@ -786,7 +791,9 @@ def api_kpi_forecast_sensitivity(
     matrix = []
     for r in results:
         kpi_info = r.get("kpi", {})
-        sens = factor_sensitivity_for_kpi(kpi_info.get("name", ""), kpi_info.get("code", ""))
+        # v2: 有历史数据用变化率弹性校准，无数据回退规则推断
+        sens = factor_sensitivity_with_history(
+            kpi_info.get("name", ""), kpi_info.get("code", ""), r.get("history", []))
         next_val = r.get("next_target")
         factor_effects = []
         for s in sens:
@@ -798,6 +805,9 @@ def api_kpi_forecast_sensitivity(
                 "factor_unit": s["factor_unit"],
                 "direction": s["direction"],
                 "elasticity": s["elasticity"],
+                "elasticity_source": s.get("elasticity_source", "rule"),
+                "matched_periods": s.get("matched_periods"),
+                "rule_direction": s.get("rule_direction"),
                 "adj_up": up_val,
                 "adj_down": down_val,
             })