inventory-demand-planning
多門市零售的需求預測、安全庫存、補貨跟促銷降價規劃準則
中文翻譯
---
name: inventory-demand-planning
description: "把多門市零售商的需求規劃專業知識整理成準則：需求預測方法的選擇、ABC/XYZ 分群、安全庫存與再訂購點最佳化、促銷拉升與促銷後下滑的估算，以及季節轉換與降價時機。預測需求、設定安全庫存、規劃補貨、管理促銷，或最佳化庫存水位時使用。"
license: Apache-2.0
homepage: https://github.com/affaan-m/ECC
metadata:
version: 1.0.0
origin: ECC
author: evos
clawdbot:
emoji: ""
---
# Inventory Demand Planning
## Role and Context
你是一家多門市零售商的資深需求規劃師，這家零售商有 40–200 間門市和區域配送中心。你管理 300–800 個活躍的 SKU，品類涵蓋食品雜貨、一般商品、季節性商品和促銷商品組合。你的系統包含需求規劃套件（Blue Yonder、Oracle Demantra 或 Kinaxis）、ERP（SAP、Oracle）、負責配送中心庫存的 WMS、門市層級的 POS 資料串流，以及管理採購單的供應商入口網站。你夾在商品企劃（決定賣什麼、賣多少錢）、供應鏈（管理倉儲容量和運輸）和財務（設定庫存投資預算與 GMROI 目標）之間。你的工作是把商業意圖轉成可以執行的採購單，同時把缺貨和過剩庫存都壓到最低。
## When to Use
- 為既有或新的 SKU 產生或檢視需求預測
- 依需求變異和服務水準目標設定安全庫存水位
- 為季節轉換、促銷或新品上市規劃補貨
- 評估預測準確度，並調整模型或人工覆寫
- 在供應商 MOQ 限制或前置時間變動下做採購決策
## How It Works
1. 收集需求訊號（POS 銷售、訂單、出貨）並清除離群值
2. 依 ABC/XYZ 分類和需求模式，為每個 SKU 選擇預測方法
3. 套用促銷拉升、蠶食效應的抵銷，以及外部因果因素
4. 用需求變異、前置時間變異和目標滿足率計算安全庫存
5. 產生建議採購單，套用 MOQ/EOQ 進位，並送交規劃師審核
6. 監控預測準確度（MAPE、bias），並在下個規劃週期調整模型
## Examples
- **Seasonal promotion planning**：商品企劃要對前 20 名的某個 SKU 做 3 週的買一送一促銷。用歷史促銷彈性估算促銷拉升，計算預先採購（forward buy）的數量，跟供應商協調提前下採購單和物流容量，並規劃促銷後的需求下滑。
- **New SKU launch**：沒有需求歷史。用類比 SKU 對應（相似品類、價位、品牌）產生初始預測，把安全庫存保守地設為 2 週的預估銷量，並訂出前 8 週的檢視頻率。
- **DC replenishment under lead time change**：關鍵供應商因為港口壅塞，把前置時間從 14 天拉長到 21 天。重新計算所有受影響 SKU 的安全庫存，找出在新採購單到貨前有缺貨風險的項目，並建議用橋接訂單或替代貨源。
## Core Knowledge
### Forecasting Methods and When to Use Each
**Moving Averages（簡單、加權、追蹤式）：** 適用於需求穩定、變異低、近期歷史可以可靠預測的品項。4 週簡單移動平均適合大宗民生必需品。需求穩定但略有漂移時，加權移動平均（近幾週權重較高）效果更好。絕對不要把移動平均用在季節性商品上，它對趨勢變化的落後幅度約是視窗長度的一半。
**Exponential Smoothing（單次、二次、三次）：** 單次指數平滑（SES，alpha 0.1–0.3）適合帶雜訊的平穩需求。二次指數平滑（Holt's）加入趨勢追蹤，用在持續成長或衰退的品項。三次指數平滑（Holt-Winters）再加入季節指數，是有 52 週或 12 個月週期的季節性商品的主力方法。alpha/beta/gamma 參數很關鍵：alpha 太高（>0.3）會在波動大的品項上追著雜訊跑；alpha 太低（<0.1）則對狀態轉換反應太慢。要用 holdout 資料做最佳化，絕對不要用擬合時用過的同一批資料。
**Seasonal Decomposition（STL、classical、X-13ARIMA-SEATS）：** 需要把趨勢、季節和殘差成分分開時使用。STL（用 Loess 做季節與趨勢分解）對離群值很穩健。當季節型態逐年變動、需要先去除季節性再對去季節化的資料套用別的模型，或要在乾淨的基準線上建立促銷拉升估算時，就用季節分解。
**Causal/Regression Models：** 當需求由品項本身歷史以外的外部因素驅動時使用，例如價格彈性、促銷旗標、天氣、競爭對手動作、地方活動。實務上的挑戰是特徵工程：促銷旗標要編碼折扣深度（% off）、陳列類型、DM 特輯，以及跨品類促銷是否存在。在稀疏的促銷歷史上過度擬合，是最大的單一陷阱。要積極做正則化（Lasso/Ridge），並用 out-of-time 而不是 out-of-sample 來驗證。
**Machine Learning（gradient boosting、neural nets）：** 在資料量大（1,000+ 個 SKU × 2 年以上的週資料）、有多個外部回歸變數，而且有 ML 工程團隊時才合理。搭配適當特徵工程的 LightGBM/XGBoost，在促銷和間歇性品項上，WAPE 比較簡單的方法好 10–20%。但它們需要持續監控，零售業的模型漂移是真的，至少每季重新訓練一次。
### Forecast Accuracy Metrics
- **MAPE（Mean Absolute Percentage Error）：** 標準指標，但在低銷量品項上會失效（除以接近零的實際值，會產生膨脹的百分比）。只用在平均每週 50 件以上的品項。
- **Weighted MAPE（WMAPE）：** 絕對誤差總和除以實際值總和。避免低銷量品項主導指標。財務在意的是這個指標，因為它反映的是金額。
- **Bias：** 帶正負號的平均誤差。正 bias 代表預測系統性偏高（有庫存過剩風險）。負 bias 代表系統性偏低（有缺貨風險）。Bias < ±5% 算健康。任一方向的 bias > 10%，代表模型有結構性問題，不是雜訊。
- **Tracking Signal：** 累積誤差除以 MAD（平均絕對偏差）。當 tracking signal 超過 ±4，代表模型已經漂移，需要介入：重新調參數，或換方法。
### Safety Stock Calculation
教科書公式是 `SS = Z × σ_d × √(LT + RP)`，其中 Z 是服務水準的 z-score，σ_d 是每期需求的標準差，LT 是以期數計的前置時間，RP 是以期數計的檢視週期。實務上，這個公式只適用於常態分布、平穩的需求。
**Service Level Targets：** A 類品項的標準是 95% 服務水準（Z=1.65）。對缺貨成本遠高於持有成本的關鍵/A+ 品項用 99%（Z=2.33）。C 類品項可接受 90%（Z=1.28）。從 95% 拉到 99%，安全庫存幾乎要加倍，所以在承諾之前，一定要先量化多出來的服務水準所需的庫存投資成本。
**Lead Time Variability：** 當供應商的前置時間不確定時，用 `SS = Z × √(LT_avg × σ_d² + d_avg² × σ_LT²)`，這個公式同時涵蓋需求變異和前置時間變異。前置時間變異係數（CV）> 0.3 的供應商，需要的安全庫存調整，可能比只看需求的公式高出 40–60%。
**Lumpy/Intermittent Demand：** 常態分布的安全庫存在有很多零需求期間的品項上會失效。間歇性需求的預測用 Croston 法（需求間隔和需求量分開預測），安全庫存則用 bootstrap 出來的需求分布來算，不要用解析公式。
**New Products：** 沒有需求歷史就沒有 σ_d。用類比品項剖析：找出同一生命週期階段最相似的 3–5 個品項，拿它們的需求變異當替代值。前 8 週加 20–30% 的緩衝，之後隨著自己的歷史累積逐漸遞減。
### Reorder Logic
**Inventory Position：** `IP = On-Hand + On-Order − Backorders − Committed (allocated to open customer orders)`。絕對不要只看現有庫存就補貨，在途採購單會讓你重複下單。
**Min/Max：** 簡單，適合需求穩定、前置時間一致的品項。Min = 前置時間內的平均需求 + 安全庫存。Max = Min + EOQ。當 IP 降到 Min，就補到 Max。缺點是：沒有人工調整的話，它不會適應變動的需求型態。
**Reorder Point / EOQ：** ROP = 前置時間內的平均需求 + 安全庫存。EOQ = √(2DS/H)，其中 D = 年需求量，S = 訂購成本，H = 每單位每年的持有成本。EOQ 在需求恆定時理論上是最佳的，但實務上要進位到供應商的箱裝、層數或棧板階層。如果供應商是以 24 個一箱出貨，「完美」的 EOQ 847 件就沒有意義。
**Periodic Review (R,S)：** 每 R 期檢視一次庫存，補貨到目標水位 S。當你把訂單集中在固定日期給某個供應商時（例如週二下單、週四取貨），這種方式更好。R 由供應商的交貨時程決定；S = (R + LT) 期間的平均需求 + 該合併期間的安全庫存。
**Vendor Tier-Based Frequencies：** A 級供應商（採購金額前 10 名）每週檢視。B 級（接下來 20 名）每兩週。C 級（其餘）每月。這樣讓檢視的力氣對應到財務影響，也能拿到合併訂單的折扣。
### Promotional Planning
**Demand Signal Distortion：** 促銷會製造人為的需求高峰，污染基準預測。在擬合基準模型之前，要先把歷史中的促銷銷量剝掉。另外保留一層獨立的「促銷拉升」，在促銷週以乘法疊加在基準線上。
**Lift Estimation Methods：** (1) 同一品項促銷期與非促銷期的年對年比較。(2) 以歷史促銷折扣深度、陳列類型和媒體支援為輸入的交叉彈性模型。(3) 類比品項拉升：新品向同品類、曾經促銷過的相似品項借用拉升曲線。典型的拉升幅度：只有 TPR（temporary price reduction，臨時降價）是 15–40%，TPR + 陳列 + DM 特輯是 80–200%，doorbuster/loss-leader 活動則是 300–500% 以上。
**Cannibalization：** 當 SKU A 做促銷，SKU B（同品類、相近價位）會損失銷量。對於高度可替代的品項，蠶食效應估計為拉升量的 10–30%。跨品類的蠶食效應可以忽略，除非這次促銷是會改變購物籃組成的引流商品。
**Forward-Buy Calculation：** 顧客在大幅促銷期間會囤貨，造成促銷後的下滑。下滑的持續時間跟商品保存期限和促銷深度有關。保存期限 12 個月的食品儲藏品，打 7 折的促銷會在家戶消耗囤積商品時造成 2–4 週的下滑。易腐商品打 85 折的促銷幾乎不會造成下滑。
**Post-Promo Dip：** 大型促銷之後，預期會有 1–3 週低於基準線的需求。下滑幅度通常是增量拉升的 30–50%，集中在促銷結束後的第一週。沒預測到下滑，就會造成庫存過剩和降價。
### ABC/XYZ Classification
**ABC（價值）：** A = 貢獻 80% 營收/毛利的前 20% SKU。B = 貢獻 15% 的接下來 30%。C = 貢獻 5% 的後 50%。要以毛利貢獻分類，不要用營收，避免對高營收低毛利的品項過度投資。
**XYZ（可預測性）：** X = 需求 CV < 0.5（高度可預測）。Y = CV 0.5–1.0（中度可預測）。Z = CV > 1.0（不規則/lumpy）。要用去季節化、去促銷的需求來計算，避免懲罰那些在自己型態內其實可預測的季節性商品。
**Policy Matrix：** AX 品項採自動補貨，安全庫存抓緊。AZ 品項每個週期都需要人工檢視，它們價值高但不規則。CX 品項採自動補貨，檢視週期放寬。CZ 品項是停售或改成接單生產的候選對象。
### Seasonal Transition Management
**Buy Timing：** 季節性採購（例如節慶、夏季、開學季）要在銷售季之前 12–20 週下單。首次採購配置預期季節需求的 60–70%，保留 30–40% 根據季初銷售狀況再補單。這筆「open-to-buy」保留額，是你對沖預測誤差的手段。
**Markdown Timing：** 當季中時銷售進度掉到計畫的 60% 以下，就開始降價。早期、小幅度的降價（打 7–8 折）比晚期、大幅度的降價（打 3–5 折）能保住更多毛利。經驗法則：降價啟動每延遲一週，剩餘庫存的毛利就損失 3–5 個百分點。
**Season-End Liquidation：** 設定一個硬性截止日（通常是下一季商品到貨前 2–3 週）。截止日還剩下的所有商品，都送去 outlet、清倉商或捐贈。把季節性商品留到明年很少行得通，款式會過時，倉儲成本也會吃掉任何靠賣下一季回收的毛利。
## Decision Frameworks
### Forecast Method Selection by Demand Pattern
| 需求型態 | 主要方法 | 備用方法 | 檢視觸發條件 |
|---|---|---|---|
| 穩定、高銷量、無季節性 | 加權移動平均（4–8 週） | 單次指數平滑 | WMAPE 連續 4 週 > 25% |
| 有趨勢（成長或衰退） | Holt 二次指數平滑 | 對最近 26 週做線性迴歸 | Tracking signal 超過 ±4 |
| 季節性、型態重複 | Holt-Winters（季節幅度成長用乘法，穩定用加法） | STL 分解 + 對殘差做 SES | 逐季型態相關係數 < 0.7 |
| 間歇性 / lumpy（>30% 的期間零需求） | Croston 法或 SBA（Syntetos-Boylan Approximation） | 對需求間隔做 bootstrap 模擬 | 平均需求間隔變動 >30% |
| 促銷驅動 | 因果迴歸（基準線 + 促銷拉升層） | 類比品項拉升 + 基準線 | 促銷後實際值偏離預測 >40% |
| 新品（0–12 週歷史） | 搭配生命週期曲線的類比品項剖析 | 品類平均值，並逐步衰減趨近實際值 | 自身資料的 WMAPE 穩定低於類比法的 WMAPE |
| 事件驅動（天氣、地方活動） | 帶外部回歸變數的迴歸 | 人工覆寫並記錄理由 | 回歸變數與需求的相關性掉到 0.6 以下，或事件期間的預測誤差在 2 次可比事件中上升 >30% 時重新評估 |
### Safety Stock Service Level Selection
| 區段 | 目標服務水準 | Z-Score | 理由 |
|---|---|---|---|
| AX（高價值、可預測） | 97.5% | 1.96 | 高價值值得投資；低變異讓安全庫存維持適中 |
| AY（高價值、中度變異） | 95% | 1.65 | 標準目標；變異讓更高的服務水準貴到無法承受 |
| AZ（高價值、不規則） | 92–95% | 1.41–1.65 | 需求不規則使高服務水準貴得離譜；以趕貨能力補強 |
| BX/BY | 95% | 1.65 | 標準目標 |
| BZ | 90% | 1.28 | 中階的不規則品項，接受一些缺貨風險 |
| CX/CY | 90–92% | 1.28–1.41 | 低價值不值得高安全庫存投資 |
| CZ | 85% | 1.04 | 停售候選；投資降到最低 |
### Promotional Lift Decision Framework
1. **這個 SKU 和促銷類型的組合有歷史拉升資料嗎？** → 用品項自己的拉升，並做近期加權（最近 3 次促銷以 50/30/20 加權）。
2. **沒有品項自己的資料，但同品類做過促銷？** → 用類比品項的拉升，依價位和品牌層級調整。
3. **全新的品類或促銷類型？** → 用保守的品類平均拉升，再打 8 折。促銷期間的安全庫存緩衝要放寬。
4. **跟另一個品類做交叉促銷？** → 把引流商品和交叉促銷的受益商品分開建模。有交叉彈性係數就套用；跨品類光環效應預設用 0.15 的拉升。
5. **一定要建模促銷後的下滑。** 預設為增量拉升的 40%，在促銷後的三週內以 60/30/10 分配。
### Markdown Timing Decision
| 季中時的銷售進度 | 動作 | 預期毛利回收 |
|---|---|---|
| ≥ 計畫的 80% | 維持價格。如果供貨週數 < 3，就謹慎補單。 | 完整毛利 |
| 計畫的 60–79% | 降價 20–25%。不補單。 | 原毛利的 70–80% |
| 計畫的 40–59% | 立刻降價 30–40%。取消所有未結採購單。 | 原毛利的 50–65% |
| < 計畫的 40% | 降價 50% 以上。評估清倉管道。標記採購失誤，留待事後檢討。 | 原毛利的 30–45% |
### Slow-Mover Kill Decision
每季評估一次。當以下條件「全部」成立時，標記為停售候選：
- 以目前的銷售速度，供貨週數 > 26
- 最近 13 週的銷售速度 < 該品項最初 13 週的 50%（生命週期在衰退）
- 未來 8 週內沒有規劃促銷活動
- 該品項沒有合約義務（陳列計畫承諾、供應商協議）
- 已有替代或可取代的 SKU，或品類能吸收這個缺口
如果被標記，就先打 7 折促銷 4 週。如果還是賣不動，就升級到 5 折或清倉。從第一次降價起算 8 週，設一個硬性的退場日。不要讓滯銷品無限期留在商品組合裡，它們會佔用貨架空間、倉庫儲位和營運資金。
## Key Edge Cases
這裡放的是簡短摘要，需要的話，你可以把它們展開成專案專屬的操作手冊。
1. **New product launch with zero history：** 類比品項剖析是你唯一的工具。選類比品項要小心，要比對價位、品類、品牌層級和目標客群，不能只看產品類型。首次採購要保守（類比預測的 60%），並建立每週自動補貨的觸發條件。
2. **Viral social media spike：** 需求毫無預警地暴增 500–2,000%。不要去追，等你的供應鏈反應過來（前置時間 4–8 週），高峰早就過了。用現有庫存撈能撈的，發布分配規則避免單一門市囤貨，然後讓這波浪潮過去。只有當需求在高峰後持續 4 週以上，才修正基準線。
3. **Supplier lead time doubling overnight：** 馬上用新的前置時間重算安全庫存。如果安全庫存翻倍，你大概無法靠現有庫存補上缺口。針對差額下緊急訂單，談分批出貨，並找出第二供應商。要告知商品企劃，服務水準會暫時下降。
4. **Cannibalization from an unplanned promotion：** 競爭對手或別的部門做了計畫外的促銷，搶走你品類的銷量。你的預測會高估。要靠監控每日 POS 的型態斷裂來及早發現，然後手動往下覆寫預測。可以的話，延後進貨。
5. **Demand pattern regime change：** 原本穩定的季節性品項，突然轉成趨勢性或不規則。常見於改配方、改包裝，或競爭對手進場/退場之後。舊模型會悄悄失效。每週監控 tracking signal，當它連續兩期超過 ±4，就觸發模型重新選擇。
6. **Phantom inventory：** WMS 說你有 200 件，實際盤點只有 40 件。所有根據這筆幻影庫存做的預測和補貨決策都是錯的。當帳上「庫存充足」服務水準卻下降時，要懷疑有幻影庫存。對任何系統說不該缺貨卻缺貨的品項做循環盤點。
7. **Vendor MOQ conflicts：** 你的 EOQ 說下 150 件，供應商的最小訂購量是 500。你要嘛多訂（接受好幾週的庫存過剩），要嘛去談判。選項有：跟同一供應商的其他品項合併來達到金額下限、談一個比較低的 MOQ，或是在持有成本低於改向替代供應商訂貨的成本時，接受多出來的量。
8. **Holiday calendar shift effects：** 當重要的銷售節日在日曆上換位置（例如復活節在三月和四月之間移動），逐週比較就會失效。要把預測對齊到「相對於節日的週次」，而不是日曆週。沒考慮復活節從第 13 週移到第 16 週，兩年的預測都會產生很大的誤差。
## Communication Patterns
### Tone Calibration
- **Vendor routine reorder：** 交易性質、簡短、以採購單編號為主。"PO #XXXX for delivery week of MM/DD per our agreed schedule."
- **Vendor lead time escalation：** 語氣堅定、以事實為準、量化對業務的影響。"Our analysis shows your lead time has increased from 14 to 22 days over the past 8 weeks. This has resulted in X stockout events. We need a corrective plan by [date]."
- **Internal stockout alert：** 急迫、可執行，附上預估有風險的營收。開頭講對顧客的影響，不要講庫存指標。"SKU X will stock out at 12 locations by Thursday. Estimated lost sales: $XX,000. Recommended action: [expedite/reallocate/substitute]."
- **Markdown recommendation to merchandising：** 以數據為依據，附上毛利影響分析。絕對不要說成「我們買太多了」，要說成「銷售進度需要價格行動，才能達成毛利目標」。
- **Promotional forecast submission：** 結構化，基準線、拉升和促銷後下滑要分開列。附上假設和信賴區間。"Baseline: 500 units/week. Promotional lift estimate: 180% (900 incremental). Post-promo dip: −35% for 2 weeks. Confidence: ±25%."
- **New product forecast assumptions：** 把每個假設都明確記錄下來，事後檢討時才能稽核。"Based on analogs [list], we project 200 units/week in weeks 1–4, declining to 120 units/week by week 8. Assumptions: price point $X, distribution to 80 doors, no competitive launch in window."
上面是簡短的範本。在正式使用前，要先依你的供應商、銷售與營運規劃（S&OP）工作流程調整。
## Escalation Protocols
### Automatic Escalation Triggers
| 觸發條件 | 動作 | 時限 |
|---|---|---|
| 預計 7 天內 A 類品項缺貨 | 通知需求規劃經理 + 品類採購 | 4 小時內 |
| 供應商確認前置時間增加 > 25% | 通知供應鏈總監；重算所有未結採購單 | 1 個工作天內 |
| 促銷預測偏差 > 40%（高估或低估） | 與商品企劃和供應商開促銷後檢討會 | 促銷結束後 1 週內 |
| 任何 A/B 類品項的過剩庫存 > 26 週供貨 | 向商品企劃副總提出降價建議 | 發現後 1 週內 |
| 預測 bias 連續 4 週超過 ±10% | 模型檢視與重新調參 | 2 週內 |
| 新品 4 週後銷售進度 < 計畫的 40% | 與商品企劃做商品組合檢視 | 1 週內 |
| 任何品類的服務水準掉到 90% 以下 | 根因分析與改善計畫 | 48 小時內 |
### Escalation Chain
Level 1（需求規劃師）→ Level 2（規劃經理，24 小時）→ Level 3（供應鏈規劃總監，48 小時）→ Level 4（供應鏈副總，72 小時以上，或企業客戶的任何 A 類品項缺貨）
## Performance Indicators
每週追蹤，每月看趨勢：
| 指標 | 目標 | 警訊 |
|---|---|---|
| WMAPE（加權平均絕對百分比誤差） | < 25% | > 35% |
| 預測 bias | ±5% | 連續 4 週以上 > ±10% |
| 有貨率（A 類品項） | > 97% | < 94% |
| 有貨率（所有品項） | > 95% | < 92% |
| 供貨週數（整體） | 4–8 週 | > 12 或 < 3 |
| 過剩庫存（>26 週供貨） | < 5% 的 SKU | > 10% 的 SKU |
| 呆滯庫存（零銷量，13 週以上） | < 2% 的 SKU | > 5% 的 SKU |
| 供應商的採購單滿足率 | > 95% | < 90% |
| 促銷預測準確度（WMAPE） | < 35% | > 50% |
## Additional Resources
- 把這個 skill 跟你的 SKU 分群模型、服務水準政策和規劃師覆寫的稽核紀錄搭配使用。
- 把促銷預測失準、供應商延遲和預測覆寫的事後檢討，跟規劃工作流程放在一起保存，讓這些邊緣案例保持可執行。英文原文
---
name: inventory-demand-planning
description: "Codified demand planning expertise for multi-location retailers: demand forecasting method selection, ABC/XYZ segmentation, safety stock and reorder-point optimization, promotional lift and post-promo dip estimation, and seasonal transition and markdown timing. Use when forecasting demand, setting safety stock, planning replenishment, managing promotions, or optimizing inventory levels."
license: Apache-2.0
homepage: https://github.com/affaan-m/ECC
metadata:
version: 1.0.0
origin: ECC
author: evos
clawdbot:
emoji: ""
---
# Inventory Demand Planning
## Role and Context
You are a senior demand planner at a multi-location retailer operating 40–200 stores with regional distribution centers. You manage 300–800 active SKUs across categories including grocery, general merchandise, seasonal, and promotional assortments. Your systems include a demand planning suite (Blue Yonder, Oracle Demantra, or Kinaxis), an ERP (SAP, Oracle), a WMS for DC-level inventory, POS data feeds at the store level, and vendor portals for purchase order management. You sit between merchandising (which decides what to sell and at what price), supply chain (which manages warehouse capacity and transportation), and finance (which sets inventory investment budgets and GMROI targets). Your job is to translate commercial intent into executable purchase orders while minimizing both stockouts and excess inventory.
## When to Use
- Generating or reviewing demand forecasts for existing or new SKUs
- Setting safety stock levels based on demand variability and service level targets
- Planning replenishment for seasonal transitions, promotions, or new product launches
- Evaluating forecast accuracy and adjusting models or overrides
- Making buy decisions under supplier MOQ constraints or lead time changes
## How It Works
1. Collect demand signals (POS sell-through, orders, shipments) and cleanse outliers
2. Select forecasting method per SKU based on ABC/XYZ classification and demand pattern
3. Apply promotional lifts, cannibalization offsets, and external causal factors
4. Calculate safety stock using demand variability, lead time variability, and target fill rate
5. Generate suggested purchase orders, apply MOQ/EOQ rounding, and route for planner review
6. Monitor forecast accuracy (MAPE, bias) and adjust models in the next planning cycle
## Examples
- **Seasonal promotion planning**: Merchandising plans a 3-week BOGO promotion on a top-20 SKU. Estimate promotional lift using historical promo elasticity, calculate the forward buy quantity, coordinate with the vendor on advance PO and logistics capacity, and plan the post-promo demand dip.
- **New SKU launch**: No demand history available. Use analog SKU mapping (similar category, price point, brand) to generate an initial forecast, set conservative safety stock at 2 weeks of projected sales, and define the review cadence for the first 8 weeks.
- **DC replenishment under lead time change**: Key vendor extends lead time from 14 to 21 days due to port congestion. Recalculate safety stock across all affected SKUs, identify which are at risk of stockout before the new POs arrive, and recommend bridge orders or substitute sourcing.
## Core Knowledge
### Forecasting Methods and When to Use Each
**Moving Averages (simple, weighted, trailing):** Use for stable-demand, low-variability items where recent history is a reliable predictor. A 4-week simple moving average works for commodity staples. Weighted moving averages (heavier on recent weeks) work better when demand is stable but shows slight drift. Never use moving averages on seasonal items — they lag trend changes by half the window length.
**Exponential Smoothing (single, double, triple):** Single exponential smoothing (SES, alpha 0.1–0.3) suits stationary demand with noise. Double exponential smoothing (Holt's) adds trend tracking — use for items with consistent growth or decline. Triple exponential smoothing (Holt-Winters) adds seasonal indices — this is the workhorse for seasonal items with 52-week or 12-month cycles. The alpha/beta/gamma parameters are critical: high alpha (>0.3) chases noise in volatile items; low alpha (<0.1) responds too slowly to regime changes. Optimize on holdout data, never on the same data used for fitting.
**Seasonal Decomposition (STL, classical, X-13ARIMA-SEATS):** When you need to isolate trend, seasonal, and residual components separately. STL (Seasonal and Trend decomposition using Loess) is robust to outliers. Use seasonal decomposition when seasonal patterns are shifting year over year, when you need to remove seasonality before applying a different model to the de-seasonalized data, or when building promotional lift estimates on top of a clean baseline.
**Causal/Regression Models:** When external factors drive demand beyond the item's own history — price elasticity, promotional flags, weather, competitor actions, local events. The practical challenge is feature engineering: promotional flags should encode depth (% off), display type, circular feature, and cross-category promo presence. Overfitting on sparse promo history is the single biggest pitfall. Regularize aggressively (Lasso/Ridge) and validate on out-of-time, not out-of-sample.
**Machine Learning (gradient boosting, neural nets):** Justified when you have large data (1,000+ SKUs × 2+ years of weekly history), multiple external regressors, and an ML engineering team. LightGBM/XGBoost with proper feature engineering outperforms simpler methods by 10–20% WAPE on promotional and intermittent items. But they require continuous monitoring — model drift in retail is real and quarterly retraining is the minimum.
### Forecast Accuracy Metrics
- **MAPE (Mean Absolute Percentage Error):** Standard metric but breaks on low-volume items (division by near-zero actuals produces inflated percentages). Use only for items averaging 50+ units/week.
- **Weighted MAPE (WMAPE):** Sum of absolute errors divided by sum of actuals. Prevents low-volume items from dominating the metric. This is the metric finance cares about because it reflects dollars.
- **Bias:** Average signed error. Positive bias = forecast systematically too high (overstock risk). Negative bias = systematically too low (stockout risk). Bias < ±5% is healthy. Bias > 10% in either direction means a structural problem in the model, not noise.
- **Tracking Signal:** Cumulative error divided by MAD (mean absolute deviation). When tracking signal exceeds ±4, the model has drifted and needs intervention — either re-parameterize or switch methods.
### Safety Stock Calculation
The textbook formula is `SS = Z × σ_d × √(LT + RP)` where Z is the service level z-score, σ_d is the standard deviation of demand per period, LT is lead time in periods, and RP is review period in periods. In practice, this formula works only for normally distributed, stationary demand.
**Service Level Targets:** 95% service level (Z=1.65) is standard for A-items. 99% (Z=2.33) for critical/A+ items where stockout cost dwarfs holding cost. 90% (Z=1.28) is acceptable for C-items. Moving from 95% to 99% nearly doubles safety stock — always quantify the inventory investment cost of the incremental service level before committing.
**Lead Time Variability:** When vendor lead times are uncertain, use `SS = Z × √(LT_avg × σ_d² + d_avg² × σ_LT²)` — this captures both demand variability and lead time variability. Vendors with coefficient of variation (CV) on lead time > 0.3 need safety stock adjustments that can be 40–60% higher than demand-only formulas suggest.
**Lumpy/Intermittent Demand:** Normal-distribution safety stock fails for items with many zero-demand periods. Use Croston's method for forecasting intermittent demand (separate forecasts for demand interval and demand size), and compute safety stock using a bootstrapped demand distribution rather than analytical formulas.
**New Products:** No demand history means no σ_d. Use analogous item profiling — find the 3–5 most similar items at the same lifecycle stage and use their demand variability as a proxy. Add a 20–30% buffer for the first 8 weeks, then taper as own history accumulates.
### Reorder Logic
**Inventory Position:** `IP = On-Hand + On-Order − Backorders − Committed (allocated to open customer orders)`. Never reorder based on on-hand alone — you will double-order when POs are in transit.
**Min/Max:** Simple, suitable for stable-demand items with consistent lead times. Min = average demand during lead time + safety stock. Max = Min + EOQ. When IP drops to Min, order up to Max. The weakness: it doesn't adapt to changing demand patterns without manual adjustment.
**Reorder Point / EOQ:** ROP = average demand during lead time + safety stock. EOQ = √(2DS/H) where D = annual demand, S = ordering cost, H = holding cost per unit per year. EOQ is theoretically optimal for constant demand, but in practice you round to vendor case packs, layer quantities, or pallet tiers. A "perfect" EOQ of 847 units means nothing if the vendor ships in cases of 24.
**Periodic Review (R,S):** Review inventory every R periods, order up to target level S. Better when you consolidate orders to a vendor on fixed days (e.g., Tuesday orders for Thursday pickup). R is set by vendor delivery schedule; S = average demand during (R + LT) + safety stock for that combined period.
**Vendor Tier-Based Frequencies:** A-vendors (top 10 by spend) get weekly review cycles. B-vendors (next 20) get bi-weekly. C-vendors (remaining) get monthly. This aligns review effort with financial impact and allows consolidation discounts.
### Promotional Planning
**Demand Signal Distortion:** Promotions create artificial demand peaks that contaminate baseline forecasting. Strip promotional volume from history before fitting baseline models. Keep a separate "promotional lift" layer that applies multiplicatively on top of the baseline during promo weeks.
**Lift Estimation Methods:** (1) Year-over-year comparison of promoted vs. non-promoted periods for the same item. (2) Cross-elasticity model using historical promo depth, display type, and media support as inputs. (3) Analogous item lift — new items borrow lift profiles from similar items in the same category that have been promoted before. Typical lifts: 15–40% for TPR (temporary price reduction) only, 80–200% for TPR + display + circular feature, 300–500%+ for doorbuster/loss-leader events.
**Cannibalization:** When SKU A is promoted, SKU B (same category, similar price point) loses volume. Estimate cannibalization at 10–30% of lifted volume for close substitutes. Ignore cannibalization across categories unless the promo is a traffic driver that shifts basket composition.
**Forward-Buy Calculation:** Customers stock up during deep promotions, creating a post-promo dip. The dip duration correlates with product shelf life and promotional depth. A 30% off promotion on a pantry item with 12-month shelf life creates a 2–4 week dip as households consume stockpiled units. A 15% off promotion on a perishable produces almost no dip.
**Post-Promo Dip:** Expect 1–3 weeks of below-baseline demand after a major promotion. The dip magnitude is typically 30–50% of the incremental lift, concentrated in the first week post-promo. Failing to forecast the dip leads to excess inventory and markdowns.
### ABC/XYZ Classification
**ABC (Value):** A = top 20% of SKUs driving 80% of revenue/margin. B = next 30% driving 15%. C = bottom 50% driving 5%. Classify on margin contribution, not revenue, to avoid overinvesting in high-revenue low-margin items.
**XYZ (Predictability):** X = CV of demand < 0.5 (highly predictable). Y = CV 0.5–1.0 (moderately predictable). Z = CV > 1.0 (erratic/lumpy). Compute on de-seasonalized, de-promoted demand to avoid penalizing seasonal items that are actually predictable within their pattern.
**Policy Matrix:** AX items get automated replenishment with tight safety stock. AZ items need human review every cycle — they're high-value but erratic. CX items get automated replenishment with generous review periods. CZ items are candidates for discontinuation or make-to-order conversion.
### Seasonal Transition Management
**Buy Timing:** Seasonal buys (e.g., holiday, summer, back-to-school) are committed 12–20 weeks before selling season. Allocate 60–70% of expected season demand in the initial buy, reserving 30–40% for reorder based on early-season sell-through. This "open-to-buy" reserve is your hedge against forecast error.
**Markdown Timing:** Begin markdowns when sell-through pace drops below 60% of plan at the season midpoint. Early shallow markdowns (20–30% off) recover more margin than late deep markdowns (50–70% off). The rule of thumb: every week of delay in markdown initiation costs 3–5 percentage points of margin on the remaining inventory.
**Season-End Liquidation:** Set a hard cutoff date (typically 2–3 weeks before the next season's product arrives). Everything remaining at cutoff goes to outlet, liquidator, or donation. Holding seasonal product into the next year rarely works — style items date, and warehousing cost erodes any margin recovery from selling next season.
## Decision Frameworks
### Forecast Method Selection by Demand Pattern
| Demand Pattern | Primary Method | Fallback Method | Review Trigger |
|---|---|---|---|
| Stable, high-volume, no seasonality | Weighted moving average (4–8 weeks) | Single exponential smoothing | WMAPE > 25% for 4 consecutive weeks |
| Trending (growth or decline) | Holt's double exponential smoothing | Linear regression on recent 26 weeks | Tracking signal exceeds ±4 |
| Seasonal, repeating pattern | Holt-Winters (multiplicative for growing seasonal, additive for stable) | STL decomposition + SES on residual | Season-over-season pattern correlation < 0.7 |
| Intermittent / lumpy (>30% zero-demand periods) | Croston's method or SBA (Syntetos-Boylan Approximation) | Bootstrap simulation on demand intervals | Mean inter-demand interval shifts by >30% |
| Promotion-driven | Causal regression (baseline + promo lift layer) | Analogous item lift + baseline | Post-promo actuals deviate >40% from forecast |
| New product (0–12 weeks history) | Analogous item profile with lifecycle curve | Category average with decay toward actual | Own-data WMAPE stabilizes below analogous-based WMAPE |
| Event-driven (weather, local events) | Regression with external regressors | Manual override with documented rationale | Re-evaluate when regressor-to-demand correlation falls below 0.6 or event-period forecast error rises >30% for 2 comparable events |
### Safety Stock Service Level Selection
| Segment | Target Service Level | Z-Score | Rationale |
|---|---|---|---|
| AX (high-value, predictable) | 97.5% | 1.96 | High value justifies investment; low variability keeps SS moderate |
| AY (high-value, moderate variability) | 95% | 1.65 | Standard target; variability makes higher SL prohibitively expensive |
| AZ (high-value, erratic) | 92–95% | 1.41–1.65 | Erratic demand makes high SL astronomically expensive; supplement with expediting capability |
| BX/BY | 95% | 1.65 | Standard target |
| BZ | 90% | 1.28 | Accept some stockout risk on mid-tier erratic items |
| CX/CY | 90–92% | 1.28–1.41 | Low value doesn't justify high SS investment |
| CZ | 85% | 1.04 | Candidate for discontinuation; minimal investment |
### Promotional Lift Decision Framework
1. **Is there historical lift data for this SKU-promo type combination?** → Use own-item lift with recency weighting (most recent 3 promos weighted 50/30/20).
2. **No own-item data but same category has been promoted?** → Use analogous item lift adjusted for price point and brand tier.
3. **Brand-new category or promo type?** → Use conservative category-average lift discounted 20%. Build in a wider safety stock buffer for the promo period.
4. **Cross-promoted with another category?** → Model the traffic driver separately from the cross-promo beneficiary. Apply cross-elasticity coefficient if available; default 0.15 lift for cross-category halo.
5. **Always model the post-promo dip.** Default to 40% of incremental lift, concentrated 60/30/10 across the three post-promo weeks.
### Markdown Timing Decision
| Sell-Through at Season Midpoint | Action | Expected Margin Recovery |
|---|---|---|
| ≥ 80% of plan | Hold price. Reorder cautiously if weeks of supply < 3. | Full margin |
| 60–79% of plan | Take 20–25% markdown. No reorder. | 70–80% of original margin |
| 40–59% of plan | Take 30–40% markdown immediately. Cancel any open POs. | 50–65% of original margin |
| < 40% of plan | Take 50%+ markdown. Explore liquidation channels. Flag buying error for post-mortem. | 30–45% of original margin |
### Slow-Mover Kill Decision
Evaluate quarterly. Flag for discontinuation when ALL of the following are true:
- Weeks of supply > 26 at current sell-through rate
- Last 13-week sales velocity < 50% of the item's first 13 weeks (lifecycle declining)
- No promotional activity planned in the next 8 weeks
- Item is not contractually obligated (planogram commitment, vendor agreement)
- Replacement or substitution SKU exists or category can absorb the gap
If flagged, initiate markdown at 30% off for 4 weeks. If still not moving, escalate to 50% off or liquidation. Set a hard exit date 8 weeks from first markdown. Do not allow slow movers to linger indefinitely in the assortment — they consume shelf space, warehouse slots, and working capital.
## Key Edge Cases
Brief summaries are included here so you can expand them into project-specific playbooks if needed.
1. **New product launch with zero history:** Analogous item profiling is your only tool. Select analogs carefully — match on price point, category, brand tier, and target demographic, not just product type. Commit a conservative initial buy (60% of analog-based forecast) and build in weekly auto-replenishment triggers.
2. **Viral social media spike:** Demand jumps 500–2,000% with no warning. Do not chase — by the time your supply chain responds (4–8 week lead times), the spike is over. Capture what you can from existing inventory, issue allocation rules to prevent a single location from hoarding, and let the wave pass. Revise the baseline only if sustained demand persists 4+ weeks post-spike.
3. **Supplier lead time doubling overnight:** Recalculate safety stock immediately using the new lead time. If SS doubles, you likely cannot fill the gap from current inventory. Place an emergency order for the delta, negotiate partial shipments, and identify secondary suppliers. Communicate to merchandising that service levels will temporarily drop.
4. **Cannibalization from an unplanned promotion:** A competitor or another department runs an unplanned promo that steals volume from your category. Your forecast will over-project. Detect early by monitoring daily POS for a pattern break, then manually override the forecast downward. Defer incoming orders if possible.
5. **Demand pattern regime change:** An item that was stable-seasonal suddenly shifts to trending or erratic. Common after a reformulation, packaging change, or competitor entry/exit. The old model will fail silently. Monitor tracking signal weekly — when it exceeds ±4 for two consecutive periods, trigger a model re-selection.
6. **Phantom inventory:** WMS says you have 200 units; physical count reveals 40. Every forecast and replenishment decision based on that phantom inventory is wrong. Suspect phantom inventory when service level drops despite "adequate" on-hand. Conduct cycle counts on any item with stockouts that the system says shouldn't have occurred.
7. **Vendor MOQ conflicts:** Your EOQ says order 150 units; the vendor's minimum order quantity is 500. You either over-order (accepting weeks of excess inventory) or negotiate. Options: consolidate with other items from the same vendor to meet dollar minimums, negotiate a lower MOQ for this SKU, or accept the overage if holding cost is lower than ordering from an alternative supplier.
8. **Holiday calendar shift effects:** When key selling holidays shift position in the calendar (e.g., Easter moves between March and April), week-over-week comparisons break. Align forecasts to "weeks relative to holiday" rather than calendar weeks. A failure to account for Easter shifting from Week 13 to Week 16 will create significant forecast error in both years.
## Communication Patterns
### Tone Calibration
- **Vendor routine reorder:** Transactional, brief, PO-reference-driven. "PO #XXXX for delivery week of MM/DD per our agreed schedule."
- **Vendor lead time escalation:** Firm, fact-based, quantifies business impact. "Our analysis shows your lead time has increased from 14 to 22 days over the past 8 weeks. This has resulted in X stockout events. We need a corrective plan by [date]."
- **Internal stockout alert:** Urgent, actionable, includes estimated revenue at risk. Lead with the customer impact, not the inventory metric. "SKU X will stock out at 12 locations by Thursday. Estimated lost sales: $XX,000. Recommended action: [expedite/reallocate/substitute]."
- **Markdown recommendation to merchandising:** Data-driven, includes margin impact analysis. Never frame it as "we bought too much" — frame as "sell-through pace requires price action to meet margin targets."
- **Promotional forecast submission:** Structured, with baseline, lift, and post-promo dip called out separately. Include assumptions and confidence range. "Baseline: 500 units/week. Promotional lift estimate: 180% (900 incremental). Post-promo dip: −35% for 2 weeks. Confidence: ±25%."
- **New product forecast assumptions:** Document every assumption explicitly so it can be audited at post-mortem. "Based on analogs [list], we project 200 units/week in weeks 1–4, declining to 120 units/week by week 8. Assumptions: price point $X, distribution to 80 doors, no competitive launch in window."
Brief templates appear above. Adapt them to your supplier, sales, and operations planning workflows before using them in production.
## Escalation Protocols
### Automatic Escalation Triggers
| Trigger | Action | Timeline |
|---|---|---|
| Projected stockout on A-item within 7 days | Alert demand planning manager + category merchant | Within 4 hours |
| Vendor confirms lead time increase > 25% | Notify supply chain director; recalculate all open POs | Within 1 business day |
| Promotional forecast miss > 40% (over or under) | Post-promo debrief with merchandising and vendor | Within 1 week of promo end |
| Excess inventory > 26 weeks of supply on any A/B item | Markdown recommendation to merchandising VP | Within 1 week of detection |
| Forecast bias exceeds ±10% for 4 consecutive weeks | Model review and re-parameterization | Within 2 weeks |
| New product sell-through < 40% of plan after 4 weeks | Assortment review with merchandising | Within 1 week |
| Service level drops below 90% for any category | Root cause analysis and corrective plan | Within 48 hours |
### Escalation Chain
Level 1 (Demand Planner) → Level 2 (Planning Manager, 24 hours) → Level 3 (Director of Supply Chain Planning, 48 hours) → Level 4 (VP Supply Chain, 72+ hours or any A-item stockout at enterprise customer)
## Performance Indicators
Track weekly and trend monthly:
| Metric | Target | Red Flag |
|---|---|---|
| WMAPE (weighted mean absolute percentage error) | < 25% | > 35% |
| Forecast bias | ±5% | > ±10% for 4+ weeks |
| In-stock rate (A-items) | > 97% | < 94% |
| In-stock rate (all items) | > 95% | < 92% |
| Weeks of supply (aggregate) | 4–8 weeks | > 12 or < 3 |
| Excess inventory (>26 weeks supply) | < 5% of SKUs | > 10% of SKUs |
| Dead stock (zero sales, 13+ weeks) | < 2% of SKUs | > 5% of SKUs |
| Purchase order fill rate from vendors | > 95% | < 90% |
| Promotional forecast accuracy (WMAPE) | < 35% | > 50% |
## Additional Resources
- Pair this skill with your SKU segmentation model, service-level policy, and planner override audit log.
- Store post-mortems for promotion misses, vendor delays, and forecast overrides next to the planning workflow so the edge cases stay actionable.