Improve DEB calibration and training metrics

2026-06-07 22:51:18 +08:00
parent 12935f502f
commit 41e819735d
9 changed files with 690 additions and 57 deletions
@@ -19,6 +19,43 @@ type MetricPayload = {
  brier_score?: number;
 };

+type DebWindowSummary = {
+  start_date?: string | null;
+  end_date?: string | null;
+  samples?: number;
+  hits?: number;
+  hit_rate?: number | null;
+  mae?: number | null;
+  bias?: number | null;
+  city_count?: number;
+};
+
+type DebHistoricalSummary = {
+  city_count?: number;
+  avg_hit_rate?: number | null;
+  weighted_hit_rate?: number | null;
+  avg_mae?: number | null;
+  avg_days_per_city?: number;
+  sample_days?: number;
+  hits?: number;
+};
+
+type DebVersionSummary = {
+  version?: string;
+  samples?: number;
+  mae?: number | null;
+  rmse?: number | null;
+  bias?: number | null;
+  bucket_hit_rate?: number | null;
+};
+
+type DebSummaryPayload = {
+  historical?: DebHistoricalSummary;
+  recent_7d?: DebWindowSummary;
+  recent_14d?: DebWindowSummary;
+  versions?: Record<string, DebVersionSummary>;
+};
+
 type TrainingCity = {
  city_id: string;
  name: string;
@@ -26,6 +63,11 @@ type TrainingCity = {
  mu?: MetricPayload;
 };

+type TrainingAccuracyPayload = {
+  accuracy: TrainingCity[];
+  deb_summary?: DebSummaryPayload;
+};
+
 const STAT_CARD_CLASSES: Record<string, string> = {
  blue: "bg-blue-50 border-blue-200",
  emerald: "bg-emerald-50 border-emerald-200",
@@ -48,44 +90,48 @@ function barColor(hr: number) {
 const TRAINING_CACHE_KEY = "polyweather_training_accuracy_v1";
 const TRAINING_CACHE_TTL_MS = 24 * 60 * 60 * 1000; // 24 hours

-function readTrainingCache(): TrainingCity[] | null {
+function readTrainingCache(): TrainingAccuracyPayload | null {
  try {
    const raw = localStorage.getItem(TRAINING_CACHE_KEY);
    if (!raw) return null;
    const cached = JSON.parse(raw);
-    if (cached.ts && Date.now() - cached.ts < TRAINING_CACHE_TTL_MS && Array.isArray(cached.data)) {
-      return cached.data;
+    if (cached.ts && Date.now() - cached.ts < TRAINING_CACHE_TTL_MS) {
+      if (Array.isArray(cached.data)) return { accuracy: cached.data };
+      if (cached.data && Array.isArray(cached.data.accuracy)) return cached.data;
    }
  } catch { /* ignore */ }
  return null;
 }

-function writeTrainingCache(data: TrainingCity[]) {
+function writeTrainingCache(data: TrainingAccuracyPayload) {
  try {
    localStorage.setItem(TRAINING_CACHE_KEY, JSON.stringify({ ts: Date.now(), data }));
  } catch { /* ignore */ }
 }

 export function TrainingDashboard({ isEn }: { isEn: boolean }) {
-  const [data, setData] = useState<TrainingCity[] | null>(() => readTrainingCache());
+  const [payload, setPayload] = useState<TrainingAccuracyPayload | null>(() => readTrainingCache());

  useEffect(() => {
    let cancelled = false;
    fetch("/api/ops/training/accuracy", { cache: "no-store", headers: { Accept: "application/json" } })
      .then(async (res) => {
        if (!res.ok) return null;
-        return res.json() as Promise<{ accuracy: TrainingCity[] }>;
+        return res.json() as Promise<TrainingAccuracyPayload>;
      })
-      .then((payload) => {
-        if (cancelled || !payload?.accuracy) return;
-        const filtered = payload.accuracy.filter((c) => (c.deb || c.mu) && ((c.deb?.total_days ?? 0) + (c.mu?.total_days ?? 0)) >= 5);
-        setData(filtered);
-        writeTrainingCache(filtered);
+      .then((nextPayload) => {
+        if (cancelled || !nextPayload?.accuracy) return;
+        const filtered = nextPayload.accuracy.filter((c) => (c.deb || c.mu) && ((c.deb?.total_days ?? 0) + (c.mu?.total_days ?? 0)) >= 5);
+        const next = { ...nextPayload, accuracy: filtered };
+        setPayload(next);
+        writeTrainingCache(next);
      })
      .catch(() => {});
    return () => { cancelled = true; };
  }, []);

+  const data = payload?.accuracy ?? null;
+  const debSummary = payload?.deb_summary;
  const debSorted = useMemo(() => (data || []).filter((c) => c.deb).sort((a, b) => (b.deb?.hit_rate ?? 0) - (a.deb?.hit_rate ?? 0)), [data]);
  const muSorted = useMemo(() => (data || []).filter((c) => c.mu).sort((a, b) => (b.mu?.hit_rate ?? 0) - (a.mu?.hit_rate ?? 0)), [data]);

@@ -94,8 +140,15 @@ export function TrainingDashboard({ isEn }: { isEn: boolean }) {
    const avgHit = debSorted.reduce((s, c) => s + (c.deb?.hit_rate ?? 0), 0) / debSorted.length;
    const avgMae = debSorted.reduce((s, c) => s + (c.deb?.mae ?? 0), 0) / debSorted.length;
    const avgDays = Math.round(debSorted.reduce((s, c) => s + (c.deb?.total_days ?? 0), 0) / Math.max(debSorted.length, 1));
-    return { avgHit, avgMae, avgDays, cities: debSorted.length };
-  }, [debSorted]);
+    return {
+      avgHit: debSummary?.historical?.avg_hit_rate ?? avgHit,
+      avgMae: debSummary?.historical?.avg_mae ?? avgMae,
+      avgDays: debSummary?.historical?.avg_days_per_city ?? avgDays,
+      cities: debSummary?.historical?.city_count ?? debSorted.length,
+      sampleDays: debSummary?.historical?.sample_days,
+      weightedHit: debSummary?.historical?.weighted_hit_rate,
+    };
+  }, [debSorted, debSummary]);

  const muStats = useMemo(() => {
    if (!muSorted.length) return null;
@@ -122,6 +175,17 @@ export function TrainingDashboard({ isEn }: { isEn: boolean }) {
    () => [...muSorted].sort((a, b) => (a.mu?.brier_score ?? 99) - (b.mu?.brier_score ?? 99)).slice(0, 18).map((c) => ({ name: c.name, value: Number((c.mu?.brier_score ?? 0).toFixed(3)) })),
    [muSorted],
  );
+  const debVersionRows = useMemo(() => {
+    const versions = debSummary?.versions || {};
+    return [
+      { key: "deb_v1_raw", label: isEn ? "Raw DEB" : "原始 DEB" },
+      { key: "deb_v1_recent_bias_corrected", label: isEn ? "Mean Bias" : "均值偏差" },
+      { key: "deb_v2_bucket_calibrated", label: isEn ? "Bucket v2" : "桶校准 v2" },
+    ].map(({ key, label }) => ({ key, label, value: versions[key] })).filter((row) => row.value);
+  }, [debSummary?.versions, isEn]);
+
+  const formatPct = (value: number | null | undefined) => value == null ? "--" : `${value.toFixed(1)}%`;
+  const formatMaybeDeg = (value: number | null | undefined, digits = 1) => value == null ? "--" : `${value.toFixed(digits)}°`;

  return (
    <div className="h-full overflow-auto bg-[#f5f7fa]">
@@ -143,12 +207,14 @@ export function TrainingDashboard({ isEn }: { isEn: boolean }) {
              <Thermometer size={14} className="text-amber-600" />
              {isEn ? "DEB Temperature Forecast" : "DEB 气温预报"}
            </h2>
-            <div className="grid grid-cols-4 gap-2 mb-3">
+            <div className="grid grid-cols-2 gap-2 mb-3 md:grid-cols-3 xl:grid-cols-6">
              {[
                { icon: Hash, label: isEn ? "Cities" : "城市数", value: debStats.cities, tone: "blue" },
-                { icon: Target, label: isEn ? "Avg Hit" : "平均命中", value: `${debStats.avgHit.toFixed(1)}%`, tone: "emerald" },
+                { icon: Target, label: isEn ? "Historical Avg" : "历史平均", value: `${debStats.avgHit.toFixed(1)}%`, tone: "emerald" },
+                { icon: TrendingUp, label: isEn ? "Recent 7d" : "近7天", value: formatPct(debSummary?.recent_7d?.hit_rate), tone: "emerald" },
+                { icon: TrendingUp, label: isEn ? "Recent 14d" : "近14天", value: formatPct(debSummary?.recent_14d?.hit_rate), tone: "purple" },
                { icon: Thermometer, label: isEn ? "Avg Error" : "平均误差", value: `${debStats.avgMae.toFixed(1)}°`, tone: "amber" },
-                { icon: TrendingUp, label: isEn ? "Avg Days/City" : "每城平均天数", value: debStats.avgDays.toLocaleString(), tone: "purple" },
+                { icon: Hash, label: isEn ? "Samples" : "样本天数", value: (debStats.sampleDays ?? debStats.avgDays).toLocaleString(), tone: "blue" },
              ].map(({ icon: Icon, label, value, tone }) => (
                <div key={label} className={`flex items-center gap-3 rounded-lg border ${STAT_CARD_CLASSES[tone]} p-3`}>
                  <Icon size={20} className={STAT_ICON_CLASSES[tone]} />
@@ -159,6 +225,25 @@ export function TrainingDashboard({ isEn }: { isEn: boolean }) {
                </div>
              ))}
            </div>
+            {debVersionRows.length ? (
+              <div className="mb-3 grid grid-cols-1 gap-2 md:grid-cols-3">
+                {debVersionRows.map((row) => {
+                  const bucketRate = row.value?.bucket_hit_rate == null ? null : row.value.bucket_hit_rate * 100;
+                  return (
+                    <div key={row.key} className="rounded-lg border border-slate-200 bg-white px-3 py-2">
+                      <div className="flex items-center justify-between gap-3">
+                        <span className="text-[11px] font-black uppercase text-slate-500">{row.label}</span>
+                        <span className="font-mono text-sm font-black text-slate-900">{formatPct(bucketRate)}</span>
+                      </div>
+                      <div className="mt-1 flex items-center justify-between text-[11px] text-slate-500">
+                        <span>{isEn ? "MAE" : "误差"} {formatMaybeDeg(row.value?.mae, 2)}</span>
+                        <span>{isEn ? "Samples" : "样本"} {row.value?.samples ?? 0}</span>
+                      </div>
+                    </div>
+                  );
+                })}
+              </div>
+            ) : null}
            <div className="grid grid-cols-2 gap-3 mb-4">
              <ChartCard title={isEn ? "Forecast Hit Rate by City" : "预报命中率 by 城市"}>
                <ResponsiveContainer width="100%" height="100%">
@@ -65,20 +65,44 @@ interface CityAccuracy {
  } | null;
 }

+interface DebSummary {
+  historical?: {
+    avg_hit_rate?: number | null;
+    weighted_hit_rate?: number | null;
+    avg_mae?: number | null;
+    sample_days?: number;
+    city_count?: number;
+  };
+  recent_7d?: {
+    hit_rate?: number | null;
+    mae?: number | null;
+    samples?: number;
+    hits?: number;
+  };
+  recent_14d?: {
+    hit_rate?: number | null;
+    mae?: number | null;
+    samples?: number;
+    hits?: number;
+  };
+}
+
 export function TrainingPageClient() {
  const [loading, setLoading] = useState(true);
  const [status, setStatus] = useState<SystemStatusPayload | null>(null);
  const [accuracy, setAccuracy] = useState<CityAccuracy[] | null>(null);
+  const [debSummary, setDebSummary] = useState<DebSummary | null>(null);

  const load = async () => {
    setLoading(true);
    try {
      const [s, accData] = await Promise.all([
        opsApi.systemStatus() as Promise<SystemStatusPayload>,
-        opsApi.trainingAccuracy().catch(() => ({ accuracy: [] as CityAccuracy[] })),
+        opsApi.trainingAccuracy().catch(() => ({ accuracy: [] as CityAccuracy[], deb_summary: null })),
      ]);
      setStatus(s);
      setAccuracy((accData as { accuracy: CityAccuracy[] }).accuracy ?? []);
+      setDebSummary((accData as { deb_summary?: DebSummary | null }).deb_summary ?? null);
    } catch { /* */ }
    setLoading(false);
  };
@@ -92,8 +116,15 @@ export function TrainingPageClient() {
    const avgMae = debCities.reduce((s, c) => s + (c.deb?.mae ?? 0), 0) / debCities.length;
    const best = debCities.reduce((a, b) => ((a.deb?.hit_rate ?? 0) > (b.deb?.hit_rate ?? 0) ? a : b));
    const worst = debCities.reduce((a, b) => ((a.deb?.mae ?? 0) > (b.deb?.mae ?? 0) ? a : b));
-    return { avgHit, avgMae, best, worst };
-  }, [accuracy]);
+    return {
+      avgHit: debSummary?.historical?.avg_hit_rate ?? avgHit,
+      avgMae: debSummary?.historical?.avg_mae ?? avgMae,
+      recent7Hit: debSummary?.recent_7d?.hit_rate,
+      recent14Hit: debSummary?.recent_14d?.hit_rate,
+      best,
+      worst,
+    };
+  }, [accuracy, debSummary]);

  const debChartData = useMemo(() => {
    if (!accuracy?.length) return [];
@@ -172,10 +203,18 @@ export function TrainingPageClient() {

      {/* Accuracy KPI row */}
      {kpis ? (
-        <div className="grid grid-cols-2 md:grid-cols-4 gap-4">
+        <div className="grid grid-cols-2 md:grid-cols-6 gap-4">
          <KpiCard
            icon={Target} color="bg-cyan-500/20 text-cyan-400"
-            label="DEB 平均命中率" value={`${kpis.avgHit.toFixed(1)}%`}
+            label="DEB 历史平均命中" value={`${kpis.avgHit.toFixed(1)}%`}
+          />
+          <KpiCard
+            icon={Target} color="bg-emerald-500/20 text-emerald-400"
+            label="近 7 天命中" value={kpis.recent7Hit == null ? "—" : `${kpis.recent7Hit.toFixed(1)}%`}
+          />
+          <KpiCard
+            icon={Activity} color="bg-violet-500/20 text-violet-400"
+            label="近 14 天命中" value={kpis.recent14Hit == null ? "—" : `${kpis.recent14Hit.toFixed(1)}%`}
          />
          <KpiCard
            icon={Activity} color="bg-blue-500/20 text-blue-400"
@@ -154,6 +154,7 @@ export const opsApi = {
          hit_rate: number;
          mae: number;
          total_days: number;
+          hits?: number;
          details_str: string;
        } | null;
        mu?: {
@@ -161,9 +162,49 @@ export const opsApi = {
          hit_rate: number;
          brier_score: number | null;
          total_days: number;
+          hits?: number;
          details_str: string;
        } | null;
      }>;
+      deb_summary?: {
+        historical?: {
+          city_count?: number;
+          avg_hit_rate?: number | null;
+          weighted_hit_rate?: number | null;
+          avg_mae?: number | null;
+          avg_days_per_city?: number;
+          sample_days?: number;
+          hits?: number;
+        };
+        recent_7d?: {
+          start_date?: string | null;
+          end_date?: string | null;
+          samples?: number;
+          hits?: number;
+          hit_rate?: number | null;
+          mae?: number | null;
+          bias?: number | null;
+          city_count?: number;
+        };
+        recent_14d?: {
+          start_date?: string | null;
+          end_date?: string | null;
+          samples?: number;
+          hits?: number;
+          hit_rate?: number | null;
+          mae?: number | null;
+          bias?: number | null;
+          city_count?: number;
+        };
+        versions?: Record<string, {
+          version?: string;
+          samples?: number;
+          mae?: number | null;
+          rmse?: number | null;
+          bias?: number | null;
+          bucket_hit_rate?: number | null;
+        }>;
+      };
    }>("/api/ops/training/accuracy");
  },
  telegramAudit() {
@@ -1267,8 +1267,10 @@ def calculate_deb_prediction(
    adjustment when enough settled samples exist.
    """
    from src.analysis.deb_evaluation import (
+        DEB_BUCKET_CALIBRATED_VERSION,
        DEB_RAW_VERSION,
        DEB_RECENT_BIAS_CORRECTED_VERSION,
+        build_bucket_calibrated_corrector,
        build_recent_bias_corrector,
        flatten_daily_records,
    )
@@ -1297,6 +1299,14 @@ def calculate_deb_prediction(
        lookback_days=bias_lookback_days,
        min_samples=bias_min_samples,
    ).apply(city_name, raw_prediction)
+    bucket_corrected = build_bucket_calibrated_corrector(
+        history_rows,
+        lookback_days=bias_lookback_days,
+        min_samples=max(5, int(bias_min_samples or 0)),
+    ).apply(city_name, raw_prediction)
+    if int(bucket_corrected.get("samples") or 0) > 0:
+        corrected = bucket_corrected
+
    bias_adjustment = float(corrected.get("bias_adjustment") or 0.0)
    bias_samples = int(corrected.get("samples") or 0)
    if bias_samples <= 0:
@@ -1311,14 +1321,19 @@ def calculate_deb_prediction(

    next_weights_info = weights_info
    if abs(bias_adjustment) >= 0.05:
+        correction_label = (
+            "bucket_calibration"
+            if corrected.get("version") == DEB_BUCKET_CALIBRATED_VERSION
+            else "recent_bias"
+        )
        next_weights_info = (
            f"{weights_info or 'DEB'} | "
-            f"recent_bias({bias_adjustment:+.1f},n={bias_samples})"
+            f"{correction_label}({bias_adjustment:+.1f},n={bias_samples})"
        )
    return {
        "prediction": corrected["corrected_prediction"],
        "raw_prediction": corrected["raw_prediction"],
-        "version": DEB_RECENT_BIAS_CORRECTED_VERSION,
+        "version": corrected.get("version") or DEB_RECENT_BIAS_CORRECTED_VERSION,
        "weights_info": next_weights_info,
        "bias_adjustment": bias_adjustment,
        "bias_samples": bias_samples,
@@ -12,6 +12,7 @@ from src.analysis.settlement_rounding import apply_city_settlement

 DEB_RAW_VERSION = "deb_v1_raw"
 DEB_RECENT_BIAS_CORRECTED_VERSION = "deb_v1_recent_bias_corrected"
+DEB_BUCKET_CALIBRATED_VERSION = "deb_v2_bucket_calibrated"
 DEB_BACKTEST_SCHEMA_VERSION = "deb_backtest_report.v1"


@@ -108,8 +109,14 @@ class BiasCorrectionResult:


 class RecentBiasCorrector:
-    def __init__(self, bias_by_city: dict[str, tuple[float, int]]) -> None:
+    def __init__(
+        self,
+        bias_by_city: dict[str, tuple[float, int]],
+        *,
+        version: str = DEB_RECENT_BIAS_CORRECTED_VERSION,
+    ) -> None:
        self._bias_by_city = bias_by_city
+        self._version = version

    def apply(self, city: str, raw_prediction: float) -> dict[str, Any]:
        city_key = str(city or "").strip().lower()
@@ -117,7 +124,7 @@ class RecentBiasCorrector:
        bias, samples = self._bias_by_city.get(city_key, (0.0, 0))
        adjustment = round(bias, 1)
        return BiasCorrectionResult(
-            version=DEB_RECENT_BIAS_CORRECTED_VERSION,
+            version=self._version,
            raw_prediction=round(raw, 1),
            corrected_prediction=round(raw + adjustment, 1),
            bias_adjustment=adjustment,
@@ -159,6 +166,71 @@ def build_recent_bias_corrector(
    return RecentBiasCorrector(bias_by_city)


+def build_bucket_calibrated_corrector(
+    history: Iterable[dict[str, Any]],
+    *,
+    lookback_days: int = 30,
+    min_samples: int = 5,
+    max_adjustment: float = 3.0,
+    step: float = 0.1,
+) -> RecentBiasCorrector:
+    by_city: dict[str, list[dict[str, Any]]] = {}
+    for record in history:
+        row = _normalise_record(record)
+        if row is None:
+            continue
+        by_city.setdefault(row["city"], []).append(row)
+
+    adjustment_by_city: dict[str, tuple[float, int]] = {}
+    safe_step = max(abs(float(step or 0.1)), 0.1)
+    max_abs = abs(float(max_adjustment or 0.0))
+    candidate_count = int(round((max_abs * 2) / safe_step)) + 1
+    candidates = [
+        round(-max_abs + idx * safe_step, 1)
+        for idx in range(max(candidate_count, 1))
+    ]
+
+    for city, rows in by_city.items():
+        rows.sort(key=lambda row: row["target_date"], reverse=True)
+        recent = rows[: max(int(lookback_days or 0), 1)]
+        if len(recent) < min_samples:
+            continue
+
+        best = None
+        for adjustment in candidates:
+            hits = 0
+            total = 0
+            abs_errors: list[float] = []
+            for row in recent:
+                prediction = row["prediction"] + adjustment
+                actual = row["actual"]
+                try:
+                    pred_bucket = apply_city_settlement(city, prediction)
+                    actual_bucket = apply_city_settlement(city, actual)
+                except Exception:
+                    continue
+                if pred_bucket is None or actual_bucket is None:
+                    continue
+                total += 1
+                if pred_bucket == actual_bucket:
+                    hits += 1
+                abs_errors.append(abs(prediction - actual))
+            if not total:
+                continue
+            mae = statistics.mean(abs_errors) if abs_errors else float("inf")
+            score = (hits, -mae, -abs(adjustment), adjustment)
+            if best is None or score > best:
+                best = score
+
+        if best is not None:
+            adjustment_by_city[city] = (best[3], len(recent))
+
+    return RecentBiasCorrector(
+        adjustment_by_city,
+        version=DEB_BUCKET_CALIBRATED_VERSION,
+    )
+
+
 def backtest_deb_versions(
    history: Iterable[dict[str, Any]],
    *,
@@ -171,6 +243,7 @@ def backtest_deb_versions(
    report_rows: list[dict[str, Any]] = []
    raw_eval_rows: list[dict[str, Any]] = []
    corrected_eval_rows: list[dict[str, Any]] = []
+    bucket_eval_rows: list[dict[str, Any]] = []

    by_city: dict[str, list[dict[str, Any]]] = {}
    for row in rows:
@@ -181,9 +254,15 @@ def backtest_deb_versions(
                lookback_days=train_lookback_days,
                min_samples=min_train_samples,
            )
+            bucket_corrector = build_bucket_calibrated_corrector(
+                previous,
+                lookback_days=train_lookback_days,
+            )
            corrected = corrector.apply(row["city"], row["prediction"])
+            bucket_corrected = bucket_corrector.apply(row["city"], row["prediction"])
            raw_prediction = round(row["prediction"], 1)
            corrected_prediction = corrected["corrected_prediction"]
+            bucket_prediction = bucket_corrected["corrected_prediction"]

            raw_eval_rows.append(
                {
@@ -201,6 +280,15 @@ def backtest_deb_versions(
                    "actual": row["actual"],
                }
            )
+            if int(bucket_corrected.get("samples") or 0) > 0:
+                bucket_eval_rows.append(
+                    {
+                        "city": row["city"],
+                        "target_date": row["target_date"],
+                        "prediction": bucket_prediction,
+                        "actual": row["actual"],
+                    }
+                )
            report_rows.append(
                {
                    "city": row["city"],
@@ -217,6 +305,12 @@ def backtest_deb_versions(
                            "bias_adjustment": corrected["bias_adjustment"],
                            "train_samples": corrected["samples"],
                        },
+                        DEB_BUCKET_CALIBRATED_VERSION: {
+                            "prediction": bucket_prediction,
+                            "error": round(bucket_prediction - row["actual"], 3),
+                            "bias_adjustment": bucket_corrected["bias_adjustment"],
+                            "train_samples": bucket_corrected["samples"],
+                        },
                    },
                }
            )
@@ -233,6 +327,10 @@ def backtest_deb_versions(
                corrected_eval_rows,
                version=DEB_RECENT_BIAS_CORRECTED_VERSION,
            ),
+            DEB_BUCKET_CALIBRATED_VERSION: evaluate_prediction_records(
+                bucket_eval_rows,
+                version=DEB_BUCKET_CALIBRATED_VERSION,
+            ),
        },
        "rows": report_rows,
    }
@@ -289,6 +387,10 @@ def write_backtest_report(
                f"{DEB_RECENT_BIAS_CORRECTED_VERSION}_error",
                f"{DEB_RECENT_BIAS_CORRECTED_VERSION}_bias_adjustment",
                f"{DEB_RECENT_BIAS_CORRECTED_VERSION}_train_samples",
+                f"{DEB_BUCKET_CALIBRATED_VERSION}_prediction",
+                f"{DEB_BUCKET_CALIBRATED_VERSION}_error",
+                f"{DEB_BUCKET_CALIBRATED_VERSION}_bias_adjustment",
+                f"{DEB_BUCKET_CALIBRATED_VERSION}_train_samples",
            ],
        )
        writer.writeheader()
@@ -296,6 +398,7 @@ def write_backtest_report(
            versions = row.get("versions") or {}
            raw = versions.get(DEB_RAW_VERSION) or {}
            corrected = versions.get(DEB_RECENT_BIAS_CORRECTED_VERSION) or {}
+            bucket = versions.get(DEB_BUCKET_CALIBRATED_VERSION) or {}
            writer.writerow(
                {
                    "city": row.get("city"),
@@ -315,5 +418,17 @@ def write_backtest_report(
                    f"{DEB_RECENT_BIAS_CORRECTED_VERSION}_train_samples": corrected.get(
                        "train_samples"
                    ),
+                    f"{DEB_BUCKET_CALIBRATED_VERSION}_prediction": bucket.get(
+                        "prediction"
+                    ),
+                    f"{DEB_BUCKET_CALIBRATED_VERSION}_error": bucket.get(
+                        "error"
+                    ),
+                    f"{DEB_BUCKET_CALIBRATED_VERSION}_bias_adjustment": bucket.get(
+                        "bias_adjustment"
+                    ),
+                    f"{DEB_BUCKET_CALIBRATED_VERSION}_train_samples": bucket.get(
+                        "train_samples"
+                    ),
                }
            )
@@ -4,9 +4,11 @@ import sys
 from pathlib import Path

 from src.analysis.deb_evaluation import (
+    DEB_BUCKET_CALIBRATED_VERSION,
    DEB_RAW_VERSION,
    DEB_RECENT_BIAS_CORRECTED_VERSION,
    backtest_deb_versions,
+    build_bucket_calibrated_corrector,
    build_recent_bias_corrector,
    evaluate_prediction_records,
    write_backtest_report,
@@ -48,6 +50,25 @@ def test_recent_bias_corrector_uses_signed_error_without_rewriting_raw_deb():
    assert corrected["samples"] == 3


+def test_bucket_calibrated_corrector_optimizes_settlement_bucket_hits():
+    history = [
+        {"city": "ankara", "target_date": "2026-05-20", "deb_prediction": 20.4, "actual_high": 21.0},
+        {"city": "ankara", "target_date": "2026-05-21", "deb_prediction": 21.4, "actual_high": 22.0},
+        {"city": "ankara", "target_date": "2026-05-22", "deb_prediction": 22.4, "actual_high": 23.0},
+        {"city": "ankara", "target_date": "2026-05-23", "deb_prediction": 23.4, "actual_high": 24.0},
+        {"city": "ankara", "target_date": "2026-05-24", "deb_prediction": 24.4, "actual_high": 25.0},
+    ]
+
+    corrector = build_bucket_calibrated_corrector(history, lookback_days=30, min_samples=5)
+    corrected = corrector.apply("ankara", raw_prediction=25.4)
+
+    assert corrected["version"] == DEB_BUCKET_CALIBRATED_VERSION
+    assert corrected["raw_prediction"] == 25.4
+    assert corrected["corrected_prediction"] == 26.0
+    assert corrected["bias_adjustment"] == 0.6
+    assert corrected["samples"] == 5
+
+
 def test_backtest_deb_versions_compares_raw_and_bias_corrected_versions():
    history = [
        {"city": "ankara", "target_date": "2026-05-20", "deb_prediction": 20.0, "actual_high": 22.0},
@@ -61,6 +82,7 @@ def test_backtest_deb_versions_compares_raw_and_bias_corrected_versions():
    assert report["schema_version"] == "deb_backtest_report.v1"
    assert report["versions"][DEB_RAW_VERSION]["samples"] == 2
    assert report["versions"][DEB_RECENT_BIAS_CORRECTED_VERSION]["samples"] == 2
+    assert report["versions"][DEB_BUCKET_CALIBRATED_VERSION]["samples"] == 0
    assert (
        report["versions"][DEB_RECENT_BIAS_CORRECTED_VERSION]["mae"]
        < report["versions"][DEB_RAW_VERSION]["mae"]
@@ -128,6 +128,52 @@ def test_calculate_deb_prediction_keeps_raw_and_adds_versioned_bias_correction(m
    assert result["bias_samples"] == 3


+def test_calculate_deb_prediction_prefers_bucket_calibration_when_enough_samples(monkeypatch):
+    monkeypatch.setattr(
+        "src.analysis.deb_algorithm.load_history",
+        lambda _: {
+            "ankara": {
+                "2026-04-11": {
+                    "actual_high": 21.0,
+                    "deb_prediction": 20.4,
+                    "forecasts": {"ECMWF": 20.4, "GFS": 20.4},
+                },
+                "2026-04-12": {
+                    "actual_high": 22.0,
+                    "deb_prediction": 21.4,
+                    "forecasts": {"ECMWF": 21.4, "GFS": 21.4},
+                },
+                "2026-04-13": {
+                    "actual_high": 23.0,
+                    "deb_prediction": 22.4,
+                    "forecasts": {"ECMWF": 22.4, "GFS": 22.4},
+                },
+                "2026-04-14": {
+                    "actual_high": 24.0,
+                    "deb_prediction": 23.4,
+                    "forecasts": {"ECMWF": 23.4, "GFS": 23.4},
+                },
+                "2026-04-15": {
+                    "actual_high": 25.0,
+                    "deb_prediction": 24.4,
+                    "forecasts": {"ECMWF": 24.4, "GFS": 24.4},
+                },
+            }
+        },
+    )
+
+    result = calculate_deb_prediction(
+        "ankara",
+        {"ECMWF": 25.4, "GFS": 25.4},
+    )
+
+    assert result["raw_prediction"] == 25.4
+    assert result["prediction"] == 26.0
+    assert result["version"] == "deb_v2_bucket_calibrated"
+    assert result["bias_adjustment"] == 0.6
+    assert result["bias_samples"] == 5
+
+
 def test_compute_hourly_model_errors_basic():
    from src.analysis.deb_algorithm import compute_hourly_model_errors

@@ -0,0 +1,38 @@
+from web.services.ops_api import _build_training_accuracy_payload
+
+
+def test_training_accuracy_payload_includes_recent_deb_summary():
+    history = {
+        "alpha": {
+            "2026-05-25": {"actual_high": 20.0, "deb_prediction": 20.0, "mu": 20.0},
+            "2026-06-01": {"actual_high": 21.0, "deb_prediction": 20.4, "mu": 21.0},
+            "2026-06-02": {"actual_high": 22.0, "deb_prediction": 21.6, "mu": 22.0},
+        },
+        "beta": {
+            "2026-06-03": {"actual_high": 30.0, "deb_prediction": 30.2, "mu": 30.0},
+            "2026-06-04": {"actual_high": 31.0, "deb_prediction": 29.2, "mu": 31.0},
+        },
+    }
+    registry = {
+        "alpha": {"name": "Alpha"},
+        "beta": {"name": "Beta"},
+        "gamma": {"name": "Gamma"},
+    }
+
+    payload = _build_training_accuracy_payload(
+        history,
+        registry,
+        today_str="2026-06-07",
+    )
+
+    assert [row["city_id"] for row in payload["accuracy"]] == ["alpha", "beta"]
+    assert payload["deb_summary"]["historical"]["city_count"] == 2
+    assert payload["deb_summary"]["historical"]["sample_days"] == 5
+    assert payload["deb_summary"]["recent_7d"]["start_date"] == "2026-05-31"
+    assert payload["deb_summary"]["recent_7d"]["end_date"] == "2026-06-06"
+    assert payload["deb_summary"]["recent_7d"]["samples"] == 4
+    assert payload["deb_summary"]["recent_7d"]["hits"] == 2
+    assert payload["deb_summary"]["recent_7d"]["hit_rate"] == 50.0
+    assert payload["deb_summary"]["recent_14d"]["samples"] == 5
+    assert "deb_v1_raw" in payload["deb_summary"]["versions"]
+    assert "deb_v2_bucket_calibrated" in payload["deb_summary"]["versions"]
@@ -19,6 +19,19 @@ from web.core import GrantPointsRequest
 import web.routes as legacy_routes


+def _sf(value: Any) -> Optional[float]:
+    try:
+        if value is None or value == "":
+            return None
+        return float(value)
+    except (TypeError, ValueError):
+        return None
+
+
+def _round_metric(value: Optional[float], digits: int = 1) -> Optional[float]:
+    return None if value is None else round(float(value), digits)
+
+
 def _require_ops(request: Request) -> Dict[str, Any] | None:
    # Ops admins are authenticated via Supabase identity + email whitelist.
    # They do NOT need an active Pro subscription to manage the system.
@@ -2212,52 +2225,271 @@ def get_ops_health_check(request: Request) -> dict[str, Any]:
    }


-def get_ops_training_accuracy(request: Request) -> Dict[str, Any]:
-    from src.analysis.deb_algorithm import get_deb_accuracy, get_mu_accuracy
-    from src.data_collection.city_registry import CITY_REGISTRY
+def _evaluate_deb_records(
+    records: List[Dict[str, Any]],
+    *,
+    start_date: Optional[str] = None,
+    end_date: Optional[str] = None,
+) -> Dict[str, Any]:
+    from src.analysis.settlement_rounding import apply_city_settlement

-    accuracy_data = []
-    for city_id, info in CITY_REGISTRY.items():
-        name = info.get("name") or city_id
+    hits = 0
+    total = 0
+    errors: List[float] = []
+    signed_errors: List[float] = []
+    cities = set()
+    dates = set()
+    for row in records:
+        target_date = str(row.get("target_date") or "").strip()
+        if start_date and target_date < start_date:
+            continue
+        if end_date and target_date > end_date:
+            continue
+        city = str(row.get("city") or "").strip().lower()
+        prediction = _sf(row.get("deb_prediction"))
+        actual = _sf(row.get("actual_high"))
+        if not city or not target_date or prediction is None or actual is None:
+            continue
+        try:
+            pred_bucket = apply_city_settlement(city, prediction)
+            actual_bucket = apply_city_settlement(city, actual)
+        except Exception:
+            continue
+        if pred_bucket is None or actual_bucket is None:
+            continue
+        total += 1
+        if pred_bucket == actual_bucket:
+            hits += 1
+        errors.append(abs(prediction - actual))
+        signed_errors.append(prediction - actual)
+        cities.add(city)
+        dates.add(target_date)

-        # Calculate DEB accuracy
-        deb_acc = get_deb_accuracy(city_id)
-        deb_payload = None
-        if deb_acc:
-            deb_payload = {
-                "hit_rate": deb_acc[0],
-                "mae": deb_acc[1],
-                "total_days": deb_acc[2],
-                "details_str": deb_acc[3],
+    return {
+        "start_date": start_date,
+        "end_date": end_date,
+        "samples": total,
+        "hits": hits,
+        "hit_rate": _round_metric((hits / total * 100) if total else None, 1),
+        "mae": _round_metric((sum(errors) / len(errors)) if errors else None, 2),
+        "bias": _round_metric((sum(signed_errors) / len(signed_errors)) if signed_errors else None, 2),
+        "city_count": len(cities),
+        "date_count": len(dates),
+    }
+
+
+def _build_city_deb_accuracy(city_id: str, city_rows: Dict[str, Dict[str, Any]], today_str: str) -> Optional[Dict[str, Any]]:
+    rows = []
+    for target_date, record in sorted((city_rows or {}).items()):
+        if target_date >= today_str or not isinstance(record, dict):
+            continue
+        rows.append(
+            {
+                "city": city_id,
+                "target_date": target_date,
+                "actual_high": record.get("actual_high"),
+                "deb_prediction": record.get("deb_prediction"),
            }
+        )
+    metrics = _evaluate_deb_records(rows)
+    if not metrics["samples"]:
+        return None
+    total = int(metrics["samples"])
+    hits = int(metrics["hits"])
+    mae = float(metrics["mae"] or 0.0)
+    hit_rate = float(metrics["hit_rate"] or 0.0)
+    return {
+        "hit_rate": hit_rate,
+        "mae": mae,
+        "total_days": total,
+        "hits": hits,
+        "details_str": f"过去{total}天 WU命中 {hits}/{total} ({hit_rate:.0f}%) | MAE: {mae:.1f}°",
+    }

-        # Calculate Mu accuracy
-        mu_acc = get_mu_accuracy(city_id)
-        mu_payload = None
-        if mu_acc:
-            mu_payload = {
-                "mae": mu_acc[0],
-                "hit_rate": mu_acc[1],
-                "brier_score": mu_acc[2],
-                "total_days": mu_acc[3],
-                "details_str": mu_acc[4],
-            }

+def _build_city_mu_accuracy(city_id: str, city_rows: Dict[str, Dict[str, Any]], today_str: str) -> Optional[Dict[str, Any]]:
+    from src.analysis.settlement_rounding import apply_city_settlement
+
+    errors: List[float] = []
+    hits = 0
+    total = 0
+    brier_scores: List[float] = []
+    for target_date, record in sorted((city_rows or {}).items()):
+        if target_date >= today_str or not isinstance(record, dict):
+            continue
+        actual = _sf(record.get("actual_high"))
+        mu_value = _sf(record.get("mu"))
+        if actual is None or mu_value is None:
+            continue
+        total += 1
+        errors.append(abs(mu_value - actual))
+        if apply_city_settlement(city_id, mu_value) == apply_city_settlement(city_id, actual):
+            hits += 1
+        prob_snapshot = record.get("prob_snapshot") or []
+        if isinstance(prob_snapshot, list):
+            actual_bucket = apply_city_settlement(city_id, actual)
+            score = 0.0
+            used = False
+            for entry in prob_snapshot:
+                if not isinstance(entry, dict):
+                    continue
+                predicted_p = _sf(entry.get("p")) or 0.0
+                outcome = 1.0 if entry.get("v") == actual_bucket else 0.0
+                score += (predicted_p - outcome) ** 2
+                used = True
+            if used:
+                brier_scores.append(score)
+
+    if not total:
+        return None
+    mae = sum(errors) / len(errors)
+    hit_rate = hits / total * 100
+    brier = (sum(brier_scores) / len(brier_scores)) if brier_scores else None
+    details_parts = [
+        f"μ准确率: 过去{total}天",
+        f"WU命中 {hits}/{total} ({hit_rate:.0f}%)",
+        f"MAE: {mae:.1f}°",
+    ]
+    if brier is not None:
+        details_parts.append(f"Brier: {brier:.3f}")
+    return {
+        "mae": mae,
+        "hit_rate": hit_rate,
+        "brier_score": brier,
+        "total_days": total,
+        "hits": hits,
+        "details_str": " | ".join(details_parts),
+    }
+
+
+def _build_deb_historical_summary(accuracy_data: List[Dict[str, Any]]) -> Dict[str, Any]:
+    deb_rows = [row for row in accuracy_data if row.get("deb")]
+    if not deb_rows:
+        return {
+            "city_count": 0,
+            "avg_hit_rate": None,
+            "weighted_hit_rate": None,
+            "avg_mae": None,
+            "avg_days_per_city": 0,
+            "sample_days": 0,
+            "hits": 0,
+        }
+    sample_days = sum(int(row["deb"].get("total_days") or 0) for row in deb_rows)
+    hits = sum(int(row["deb"].get("hits") or 0) for row in deb_rows)
+    return {
+        "city_count": len(deb_rows),
+        "avg_hit_rate": _round_metric(
+            sum(float(row["deb"].get("hit_rate") or 0.0) for row in deb_rows) / len(deb_rows),
+            1,
+        ),
+        "weighted_hit_rate": _round_metric((hits / sample_days * 100) if sample_days else None, 1),
+        "avg_mae": _round_metric(
+            sum(float(row["deb"].get("mae") or 0.0) for row in deb_rows) / len(deb_rows),
+            2,
+        ),
+        "avg_days_per_city": round(sample_days / len(deb_rows)) if deb_rows else 0,
+        "sample_days": sample_days,
+        "hits": hits,
+    }
+
+
+def _flatten_training_history(history: Dict[str, Dict[str, Dict[str, Any]]], today_str: str) -> List[Dict[str, Any]]:
+    rows: List[Dict[str, Any]] = []
+    for city, city_rows in (history or {}).items():
+        if not isinstance(city_rows, dict):
+            continue
+        for target_date, record in city_rows.items():
+            if not isinstance(record, dict):
+                continue
+            target_text = str(target_date or "").strip()
+            if not target_text or target_text >= today_str:
+                continue
+            rows.append(
+                {
+                    "city": str(city or "").strip().lower(),
+                    "target_date": target_text,
+                    "actual_high": record.get("actual_high"),
+                    "deb_prediction": record.get("deb_prediction"),
+                    "mu": record.get("mu"),
+                    "prob_snapshot": record.get("prob_snapshot"),
+                }
+            )
+    rows.sort(key=lambda row: (row["target_date"], row["city"]))
+    return rows
+
+
+def _build_training_accuracy_payload(
+    history: Dict[str, Dict[str, Dict[str, Any]]],
+    city_registry: Dict[str, Dict[str, Any]],
+    *,
+    today_str: Optional[str] = None,
+) -> Dict[str, Any]:
+    from src.analysis.deb_evaluation import backtest_deb_versions
+
+    today = today_str or datetime.now(timezone.utc).strftime("%Y-%m-%d")
+    accuracy_data: List[Dict[str, Any]] = []
+    for city_id, info in (city_registry or {}).items():
+        city_rows = history.get(city_id) or history.get(str(city_id).strip().lower()) or {}
+        deb_payload = _build_city_deb_accuracy(city_id, city_rows, today)
+        mu_payload = _build_city_mu_accuracy(city_id, city_rows, today)
        if deb_payload or mu_payload:
            accuracy_data.append(
-                {"city_id": city_id, "name": name, "deb": deb_payload, "mu": mu_payload}
+                {
+                    "city_id": city_id,
+                    "name": (info or {}).get("name") or city_id,
+                    "deb": deb_payload,
+                    "mu": mu_payload,
+                }
            )

-    # Sort by total days of DEB or Mu
    accuracy_data.sort(
-        key=lambda x: max(
-            x["deb"]["total_days"] if x["deb"] else 0,
-            x["mu"]["total_days"] if x["mu"] else 0,
+        key=lambda row: max(
+            row["deb"]["total_days"] if row.get("deb") else 0,
+            row["mu"]["total_days"] if row.get("mu") else 0,
        ),
        reverse=True,
    )

-    return {"accuracy": accuracy_data}
+    all_rows = _flatten_training_history(history, today)
+    today_date = datetime.strptime(today, "%Y-%m-%d").date()
+    recent_7_start = (today_date - timedelta(days=7)).isoformat()
+    recent_14_start = (today_date - timedelta(days=14)).isoformat()
+    end_date = (today_date - timedelta(days=1)).isoformat()
+    versions = backtest_deb_versions(
+        all_rows,
+        min_train_samples=2,
+    ).get("versions", {})
+
+    return {
+        "accuracy": accuracy_data,
+        "deb_summary": {
+            "historical": _build_deb_historical_summary(accuracy_data),
+            "recent_7d": _evaluate_deb_records(
+                all_rows,
+                start_date=recent_7_start,
+                end_date=end_date,
+            ),
+            "recent_14d": _evaluate_deb_records(
+                all_rows,
+                start_date=recent_14_start,
+                end_date=end_date,
+            ),
+            "versions": versions,
+        },
+    }
+
+
+def get_ops_training_accuracy(request: Request) -> Dict[str, Any]:
+    from src.analysis.deb_algorithm import load_history
+    from src.data_collection.city_registry import CITY_REGISTRY
+
+    history_file = os.path.join(
+        os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))),
+        "data",
+        "daily_records.json",
+    )
+    history = load_history(history_file)
+    return _build_training_accuracy_payload(history, CITY_REGISTRY)


 def get_ops_telegram_audit(request: Request) -> Dict[str, Any]: