Update Forecast bug

streamlit_app.py

@@ -96,9 +96,11 @@ LANG = {
         "fc_month_customers": "Forecast Customers (this month)",
         "fc_month_revenue":   "Forecast Revenue (this month, est.)",
         "fc_month_basis":     "Revenue is estimated as forecast customers × trailing 3-month Rev/Head per branch.",
-        "fc_month_basis_full":"Copper Buffet uses the model-based daily forecast (Prediction × Rev/Head per branch). "
-                              "Tiew Copper has no per-day forecast, so its current-month tiles are estimated as the "
-                              "trailing 3-month average of monthly Customers and Revenue per branch.",
+        "fc_month_basis_full":"This month total combines actual customers and revenue for days that have already "
+                              "passed (from kpi_daily) with projections for remaining days. Copper Buffet's "
+                              "remaining days use the model's per-day prediction × trailing 3-month Rev/Head per "
+                              "branch. Tiew Copper's remaining days are projected at the trailing 90-day average "
+                              "per day-of-week, so weekdays and weekends are weighted separately.",
         "fc_header":          "Copper Buffet — Forecast & Bookings",
         "fc_caption":         "Forecasted customer counts and confirmed bookings for upcoming "
                               "service dates. Data is captured only for Copper Buffet.",
@@ -222,9 +224,11 @@ LANG = {
         "fc_month_customers": "พยากรณ์จำนวนลูกค้า (เดือนนี้)",
         "fc_month_revenue":   "พยากรณ์รายได้ (เดือนนี้, ประมาณการ)",
         "fc_month_basis":     "ประมาณการรายได้จาก: พยากรณ์จำนวนลูกค้า × รายได้ต่อหัวเฉลี่ย 3 เดือนล่าสุดของแต่ละสาขา",
-        "fc_month_basis_full":"คอปเปอร์บุฟเฟ่ต์ใช้พยากรณ์รายวันจากโมเดล (พยากรณ์ลูกค้า × รายได้ต่อหัวต่อสาขา) "
-                              "เตี่ยวคอปเปอร์ไม่มีพยากรณ์รายวัน จึงประมาณค่าของเดือนนี้จาก "
-                              "ค่าเฉลี่ยลูกค้าและรายได้ต่อเดือน 3 เดือนล่าสุดของแต่ละสาขา",
+        "fc_month_basis_full":"ยอดรวมเดือนนี้รวมข้อมูลจริงของวันที่ผ่านมาแล้ว (จาก kpi_daily) "
+                              "กับการประมาณการสำหรับวันที่เหลือ คอปเปอร์บุฟเฟ่ต์ใช้พยากรณ์รายวันจากโมเดล "
+                              "× รายได้ต่อหัวเฉลี่ย 3 เดือนล่าสุดของแต่ละสาขา ส่วนเตี่ยวคอปเปอร์ "
+                              "ใช้ค่าเฉลี่ย 90 วันล่าสุดตามวันในสัปดาห์สำหรับวันที่เหลือ "
+                              "(วันธรรมดาและวันหยุดสุดสัปดาห์จะถูกถ่วงน้ำหนักแยกกัน)",
         "fc_header":          "คอปเปอร์บุฟเฟ่ต์ — พยากรณ์และการจอง",
         "fc_caption":         "พยากรณ์จำนวนลูกค้าและการจองที่ยืนยันแล้วสำหรับวันที่บริการในอนาคต "
                               "ข้อมูลมีเฉพาะของคอปเปอร์บุฟเฟ่ต์เท่านั้น",
@@ -2149,75 +2153,127 @@ with tab_forecast:
         st.subheader(t("fc_header"))
         st.caption(t("fc_caption"))
 
-        # ── This Month forecast — per-restaurant customer + revenue ─────
-        # Copper Buffet uses fact_predictions (model-based daily prediction
-        # × trailing 3-month Rev/Head per branch). Tiew Copper has no
-        # per-day forecast in the dataset, so it falls back to the
-        # trailing 3-month average of monthly Customers / Revenue per
-        # branch — a reasonable "what we usually do" baseline.
+        # ── This Month forecast — actual MTD + projection for remaining ─
+        # For days that have already passed, use real customers + revenue
+        # from kpi_daily. For days that haven't happened yet:
+        #   • Copper Buffet — model-based per-day prediction from
+        #     fact_predictions × trailing 3-month Rev/Head per branch.
+        #   • Tiew Copper — trailing 3-month average daily rate ×
+        #     remaining days (no per-day model exists for Tiew).
         st.markdown(f"**{t('fc_month_title')}**")
         _now = pd.Timestamp(_dt.now().date())
         _month_start = _now.replace(day=1)
         _month_end   = (_month_start + pd.offsets.MonthEnd(0)).normalize()
 
-        def _cb_month_forecast() -> tuple[int, float]:
-            """Copper Buffet — predict from fact_predictions × Rev/Head."""
-            _mp = fact_predictions.copy() if not fact_predictions.empty else pd.DataFrame()
-            if _mp.empty:
-                return 0, 0.0
-            _mp["Date"] = pd.to_datetime(_mp["Date"], errors="coerce")
-            _mp = _mp[(_mp["Date"] >= _month_start) & (_mp["Date"] <= _month_end)]
-            if "Restaurant" in _mp.columns:
-                _mp = _mp[_mp["Restaurant"] == "Copper Buffet"]
-            if sel_branches and "Branch" in _mp.columns:
-                _mp = _mp[_mp["Branch"].isin(sel_branches)]
-            if _mp.empty:
-                return 0, 0.0
-            # Latest snapshot per (Date, Branch).
-            if "Date_Diff" in _mp.columns:
-                _mp = _mp.assign(_a=_mp["Date_Diff"].abs()) \
-                         .sort_values("_a") \
-                         .drop_duplicates(["Date", "Branch"], keep="first") \
-                         .drop(columns="_a")
-            cust = int(_mp["Prediction"].sum())
-
-            # Per-branch trailing 3-month Rev/Head from kpi_monthly.
-            rev = 0.0
-            if not kpi_monthly.empty and {"Restaurant", "Branch", "Rev_Per_Head", "Year", "Month"}.issubset(kpi_monthly.columns):
-                hist = kpi_monthly[kpi_monthly["Restaurant"] == "Copper Buffet"].sort_values(["Year", "Month"])
-                rph_map = (
-                    hist.groupby("Branch").tail(3)
-                        .groupby("Branch")["Rev_Per_Head"].mean().to_dict()
-                )
-                fallback = sum(rph_map.values()) / len(rph_map) if rph_map else 0.0
-                for _br, _cust in _mp.groupby("Branch")["Prediction"].sum().items():
-                    rev += float(_cust) * rph_map.get(_br, fallback)
-            return cust, rev
-
-        def _avg_month_forecast(restaurant: str) -> tuple[float, float]:
-            """Trailing 3-month avg of monthly Customers and Revenue,
-            summed across the branches the sidebar is filtered to."""
-            if kpi_monthly.empty:
+        def _actual_mtd(restaurant: str) -> tuple[float, float]:
+            """Sum of actual Customers + Revenue from kpi_daily for days
+            in the current month that are strictly before today."""
+            if kpi_daily.empty:
                 return 0.0, 0.0
-            df = kpi_monthly[kpi_monthly.get("Restaurant", "") == restaurant].copy()
-            if sel_branches and "Branch" in df.columns:
-                df = df[df["Branch"].isin(sel_branches)]
-            if df.empty or not {"Year", "Month"}.issubset(df.columns):
+            kd = kpi_daily.copy()
+            if "Date" in kd.columns:
+                kd["Date"] = pd.to_datetime(kd["Date"], errors="coerce")
+            if "Restaurant" in kd.columns:
+                kd = kd[kd["Restaurant"] == restaurant]
+            if sel_branches and "Branch" in kd.columns:
+                kd = kd[kd["Branch"].isin(sel_branches)]
+            kd = kd[(kd["Date"] >= _month_start) & (kd["Date"] < _now)]
+            if kd.empty:
                 return 0.0, 0.0
-            # One row per (Year, Month) — sum across whichever branches survived.
-            monthly_totals = (
-                df.groupby(["Year", "Month"], as_index=False)
+            cust = float(kd["Customers"].sum()) if "Customers" in kd.columns else 0.0
+            rev  = float(kd["Revenue"].sum())   if "Revenue"   in kd.columns else 0.0
+            return cust, rev
+
+        def _cb_month_forecast() -> tuple[float, float]:
+            """Copper Buffet — actual MTD + per-day prediction for remaining."""
+            actual_cust, actual_rev = _actual_mtd("Copper Buffet")
+            pred_cust = 0.0
+            pred_rev  = 0.0
+            if not fact_predictions.empty:
+                _mp = fact_predictions.copy()
+                _mp["Date"] = pd.to_datetime(_mp["Date"], errors="coerce")
+                # Only days from today onwards within current month.
+                _mp = _mp[(_mp["Date"] >= _now) & (_mp["Date"] <= _month_end)]
+                if "Restaurant" in _mp.columns:
+                    _mp = _mp[_mp["Restaurant"] == "Copper Buffet"]
+                if sel_branches and "Branch" in _mp.columns:
+                    _mp = _mp[_mp["Branch"].isin(sel_branches)]
+                if not _mp.empty:
+                    # Latest snapshot per (Date, Branch).
+                    if "Date_Diff" in _mp.columns:
+                        _mp = _mp.assign(_a=_mp["Date_Diff"].abs()) \
+                                 .sort_values("_a") \
+                                 .drop_duplicates(["Date", "Branch"], keep="first") \
+                                 .drop(columns="_a")
+                    pred_cust = float(_mp["Prediction"].sum())
+                    # Per-branch trailing 3-month Rev/Head from kpi_monthly.
+                    if not kpi_monthly.empty and {"Restaurant", "Branch", "Rev_Per_Head", "Year", "Month"}.issubset(kpi_monthly.columns):
+                        hist = kpi_monthly[kpi_monthly["Restaurant"] == "Copper Buffet"].sort_values(["Year", "Month"])
+                        rph_map = (
+                            hist.groupby("Branch").tail(3)
+                                .groupby("Branch")["Rev_Per_Head"].mean().to_dict()
+                        )
+                        fallback = sum(rph_map.values()) / len(rph_map) if rph_map else 0.0
+                        for _br, _cust in _mp.groupby("Branch")["Prediction"].sum().items():
+                            pred_rev += float(_cust) * rph_map.get(_br, fallback)
+            return actual_cust + pred_cust, actual_rev + pred_rev
+
+        def _tc_month_forecast() -> tuple[float, float]:
+            """Tiew Copper — actual MTD + per-day projection using day-of-week
+            weighted averages from the trailing 90 days.
+
+            Why day-of-week? Restaurant traffic varies sharply by DOW (weekends
+            ≫ weekdays in most cases). Averaging by DOW means a Sunday at the
+            end of the month gets projected at a Sunday-typical rate instead
+            of a "mean of every day this quarter" rate, which would massively
+            under-count weekend nights and over-count weekday nights.
+            """
+            actual_cust, actual_rev = _actual_mtd("Tiew Copper")
+            remaining = pd.date_range(_now, _month_end, freq="D")
+            if len(remaining) == 0 or kpi_daily.empty:
+                return actual_cust, actual_rev
+
+            # Trailing 90 days before the start of the current month.
+            trailing_start = _month_start - pd.Timedelta(days=90)
+            kd = kpi_daily.copy()
+            kd["Date"] = pd.to_datetime(kd["Date"], errors="coerce")
+            kd = kd[kd.get("Restaurant", "") == "Tiew Copper"]
+            if sel_branches and "Branch" in kd.columns:
+                kd = kd[kd["Branch"].isin(sel_branches)]
+            kd = kd[(kd["Date"] >= trailing_start) & (kd["Date"] < _month_start)]
+            if kd.empty:
+                return actual_cust, actual_rev
+
+            # Sum branches first to get a single per-day total, then average
+            # across days within each day-of-week bucket. DOW: Mon=0 … Sun=6.
+            daily_totals = (
+                kd.groupby("Date", as_index=False)
                   .agg(Customers=("Customers", "sum"),
                        Revenue=("Revenue", "sum"))
-                  .sort_values(["Year", "Month"])
-                  .tail(3)
             )
-            if monthly_totals.empty:
-                return 0.0, 0.0
-            return monthly_totals["Customers"].mean(), monthly_totals["Revenue"].mean()
+            daily_totals["DOW"] = daily_totals["Date"].dt.dayofweek
+            dow_avg = (
+                daily_totals.groupby("DOW", as_index=False)
+                            .agg(AvgCust=("Customers", "mean"),
+                                 AvgRev=("Revenue", "mean"))
+            )
+            # Fallback rate if a DOW has no historical samples (e.g. closed
+            # on Mondays during the trailing window).
+            fb_cust = float(daily_totals["Customers"].mean())
+            fb_rev  = float(daily_totals["Revenue"].mean())
+            dow_cust = dict(zip(dow_avg["DOW"], dow_avg["AvgCust"]))
+            dow_rev  = dict(zip(dow_avg["DOW"], dow_avg["AvgRev"]))
+
+            proj_cust = 0.0
+            proj_rev  = 0.0
+            for d in remaining:
+                dow = int(d.dayofweek)
+                proj_cust += float(dow_cust.get(dow, fb_cust))
+                proj_rev  += float(dow_rev.get(dow,  fb_rev))
+            return actual_cust + proj_cust, actual_rev + proj_rev
 
         cb_cust, cb_rev = _cb_month_forecast()
-        tc_cust, tc_rev = _avg_month_forecast("Tiew Copper")
+        tc_cust, tc_rev = _tc_month_forecast()
 
         # Copper Buffet block
         st.markdown("**Copper Buffet**")