🧠 Case Study — Basic Forecasting: Walmart, Amazon, Thế Giới Di Động

Trong buổi học này, chúng ta đã nắm Forecasting fundamentals — time series components, methods (Naïve, SMA, ETS, Holt-Winters), evaluation metrics (MAE, MAPE, RMSE). Bây giờ hãy xem 3 công ty áp dụng forecasting như thế nào — từ Walmart dự báo demand cho 11,000+ stores, Amazon forecast inventory cho next-day delivery, và Thế Giới Di Động forecast khuyến mãi cho 2,000+ cửa hàng.

---

Case Study 1: Walmart — Demand Forecasting cho 11,000+ Stores

Bối cảnh

Walmart (2025): retailer lớn nhất thế giới — $648B revenue, 11,500+ stores tại 19 quốc gia, 2.1 triệu nhân viên. Mỗi store trung bình bán 120,000+ SKUs. Mỗi ngày, Walmart cần quyết định: nhập bao nhiêu, của mặt hàng nào, vào store nào?

Quy mô forecast problem: 11,500 stores × 120,000 SKUs = 1.38 tỷ forecast points mỗi tuần. Đây là bài toán forecasting lớn nhất thế giới — không ai làm bằng Excel!

Thách thức:

• Weather sensitivity: Mưa → bán nhiều dù, áo mưa; nắng → bán kem, nước giải khát. Walmart cần forecast weather impact cho từng store, từng ngày
• Promotion effect: 10,000+ promotions đồng thời — mỗi promotion thay đổi demand pattern
• Holiday spikes: Thanksgiving, Christmas, Super Bowl — demand tăng 300-500% một số categories
• Local factors: Store ở Miami ≠ store ở Minneapolis — demographics, income, preferences khác nhau

Hệ thống Forecasting

Phase 1: Traditional Statistical Methods (2000-2015)

Walmart ban đầu dùng Holt-Winters Exponential Smoothing — phương pháp chính cho demand forecasting:

Component	Method	Parameters
Base demand	Holt-Winters (additive)	α = 0.15, β = 0.02, γ = 0.10
Seasonal adjustment	Multiplicative seasonal factors	52 weekly factors
Promotion uplift	Regression adjustment	Promotion type × discount depth
Weather adjustment	Manual rules	Temperature thresholds per category

Quy trình forecast:

1. Historical data: 3 năm weekly sales per store-SKU
2. Holt-Winters decompose: trend + seasonality
3. Seasonal factors: 52 tuần, tính multiplicative index
4. Promotion calendar: overlay promotion uplift factors
5. Weather: adjust ±5-15% based on forecast temperature
6. Final demand = HW_forecast × seasonal_factor × promo_uplift × weather_adj

Accuracy (Phase 1):

Metric	Value	Context
Overall MAPE	22%	Across all store-SKU combinations
Top 1000 SKUs MAPE	12%	High-volume, stable items
Long-tail SKUs MAPE	45%	Low-volume, intermittent demand
Promotion periods MAPE	35%	Demand spikes hard to predict
Holiday weeks MAPE	28%	Extreme spikes, variable patterns

Vấn đề: MAPE 22% overall ≈ $3.5 billion/năm excess inventory cost. CFO nói: "1% improvement in forecast accuracy = $150M savings."

Phase 2: Enhanced Statistical + External Data (2015-2022)

Walmart thêm external signals vào Holt-Winters baseline:

Signal	Data Source	Impact
Weather	The Weather Company (IBM)	Temperature, precipitation → ±10-20% demand shift
Events	Local event calendars	Concerts, sports games → foot traffic
Economic indicators	Bureau of Labor Statistics	Unemployment rate, consumer confidence
Competitor pricing	Web scraping + market data	Price matching → demand shift
Social media	Twitter/X trends	Viral products → sudden demand spikes

New accuracy targets:

Metric	Phase 1	Phase 2	Improvement
Overall MAPE	22%	14%	-36%
Top 1000 SKUs	12%	7%	-42%
Promotion MAPE	35%	18%	-49%
Holiday MAPE	28%	15%	-46%

Financial impact: MAPE giảm 8 percentage points → $1.2B/năm savings từ giảm excess inventory + giảm stockouts.

Key Forecasting Insight: Walmart's "1-1-1 Rule"

Walmart phát hiện rule of thumb:

• 1% improvement in forecast accuracy
• = 1% reduction in safety stock
• = ~$150M savings per year

"Forecast accuracy không phải academic exercise — mỗi percentage point = hàng trăm triệu dollars," — Greg Smith, VP Supply Chain Analytics.

Lessons Learned

1. Simple methods work surprisingly well: Holt-Winters cho MAPE 12% trên top SKUs — competitive với complex ML models
2. External data matters: Weather alone improved forecast accuracy 3-5% cho weather-sensitive categories
3. Decompose before forecast: Walmart LUÔN decompose trước — hiểu trend vs season vs promotion effect
4. Forecast granularity matters: Store-level forecast chính xác hơn regional forecast → rồi allocate
5. Forecast error ≠ symmetric: Over-forecasting (excess inventory) cost khác under-forecasting (stockout) cost → loss function asymmetric

---

Case Study 2: Amazon — Inventory Forecasting cho Next-Day Delivery

Bối cảnh

Amazon (2025): e-commerce + cloud giant — $638B revenue, 400M+ products, 200+ fulfillment centers (FCs) tại 21 quốc gia. Next-day delivery (và same-day delivery tại nhiều thành phố) là competitive advantage #1 — nhưng yêu cầu inventory đúng product, đúng FC, đúng thời điểm.

Thách thức forecasting:

• 400M+ products = forecast cho mỗi ASIN (Amazon Standard Identification Number)
• 200+ FCs = forecast per ASIN per FC (not just total)
• Lead time varies: Seller ở China = 3 tuần, seller local = 2 ngày → forecast horizon khác nhau
• Long tail: 80% products bán < 10 units/tuần → intermittent demand, traditional methods fail

Hệ thống Forecasting

Multi-Tier Forecasting System

Amazon không dùng 1 model — dùng ensemble of models cho từng product tier:

Tier	Products	Volume	Method	MAPE
Tier 1: Fast movers	Top 5% ASINs	60% revenue	Holt-Winters + Deep Learning	5-8%
Tier 2: Regular	Next 20%	25% revenue	ETS (auto-selected) + Regression	12-18%
Tier 3: Slow movers	Next 35%	10% revenue	Croston's method (intermittent)	25-35%
Tier 4: Long tail	Bottom 40%	5% revenue	Safety stock rules, no forecast	N/A

Forecast Process cho Next-Day Delivery

AMAZON FORECASTING PIPELINE
━━━━━━━━━━━━━━━━━━━━━━━━━━

Step 1: DEMAND FORECAST (what will customers buy?)
├── Per ASIN: daily demand forecast, 90-day horizon
├── Method: Holt-Winters (baseline) + event uplift + promo calendar
├── Output: point forecast + prediction intervals (P10, P50, P90)
└── Update frequency: DAILY

Step 2: INVENTORY POSITIONING (where should inventory be?)
├── Model: demand forecast × service level target
├── Constraint: FC capacity, inbound lead time
├── Output: reorder points + reorder quantities per FC
└── Safety stock = f(forecast error, lead time, service level)

Step 3: INBOUND PLANNING (when to order from suppliers?)
├── Lead time: 2 days (local) to 30 days (overseas)
├── Forecast horizon must exceed lead time!
├── Order quantity: EOQ adjusted for forecast + PI
└── Output: PO (Purchase Order) per supplier per week

Step 4: PLACEMENT (which FC gets what?)
├── Customer demand by region → allocate to nearest FCs
├── Transportation cost optimization
├── Output: transfer orders between FCs
└── Goal: 95%+ items within 1-day delivery radius

Forecast Accuracy Impact

Scenario	Forecast Error	Business Impact
Perfect forecast	0%	Optimal inventory, zero waste
Over-forecast 10%	+10%	$2.8B excess inventory annually; storage costs
Under-forecast 10%	-10%	15% orders delayed → customer churn → ~$4.5B revenue risk
Amazon actual	~8% MAPE (Tier 1)	Balance: 98.5% in-stock rate, minimal excess

Key metric: Amazon targets 98.5% in-stock rate cho Tier 1 products — nghĩa là 98.5% thời gian, customer thấy "In Stock" khi search. Đạt mức này cần forecast MAPE < 10%.

Prediction Intervals — Critical cho Amazon

Amazon KHÔNG dùng point forecast cho inventory decisions — dùng quantile forecasts:

Quantile	Meaning	Usage
P10	10% chance demand ≤ value	Minimum order for slow categories
P50	Median forecast	Standard replenishment
P75	75% chance demand ≤ value	Fast movers, low stockout tolerance
P90	90% chance demand ≤ value	Critical items (Prime, Subscribe & Save)
P99	99% chance demand ≤ value	Ultra-critical (pandemic essentials)

"We don't forecast demand. We forecast the DISTRIBUTION of demand." — Werner Vogels, CTO Amazon.

Prime Day — Stress Test cho Forecasting

Amazon Prime Day 2025:

• 48-hour event, deals across all categories
• Demand tăng 300-1000% tùy category
• Forecast challenge: lần đầu tiên cho many deals → no historical pattern

Approach:

1. Baseline demand (Holt-Winters): normal week forecast
2. Promotion uplift: similar deals last year → regression predict uplift
3. Category-level adjustment: Electronics +500%, Fashion +300%, Grocery +150%
4. FC pre-positioning: 3 tuần trước Prime Day → ship inventory to optimal FCs
5. Dynamic re-forecasting: giữa Prime Day, update forecast mỗi 6 giờ

Result Prime Day 2025:

Metric	Target	Actual
In-stock rate	97%	98.2%
Demand forecast MAPE	<20%	14.5%
Next-day delivery rate	85%	88%
Excess inventory post-Prime Day	<$500M	$380M

Lessons Learned

1. Tiered approach: Không cần 1 method cho tất cả — Holt-Winters cho fast movers, Croston cho slow movers
2. Quantile forecasts > Point forecasts: Inventory decisions cần distribution, không chỉ mean
3. Daily re-forecasting: Forecast decay nhanh — daily update giữ accuracy cao
4. Pre-positioning critical: Forecast đúng nhưng inventory sai chỗ = vẫn late delivery
5. Under-forecast costlier than over-forecast: Stockout → lost customer > Excess → storage cost

---

Case Study 3: Thế Giới Di Động — Demand Forecasting cho Khuyến Mãi

Bối cảnh

Thế Giới Di Động (MWG) (2025): chuỗi bán lẻ điện tử + tiêu dùng lớn nhất Việt Nam — revenue ~120,000 tỷ VND, 2,100+ cửa hàng (Thế Giới Di Động + Điện Máy Xanh + Bách Hóa Xanh). 50,000+ SKUs electronics, 15,000+ SKUs grocery.

Thách thức forecasting:

• Khuyến mãi liên tục: TGDĐ chạy 50+ promotions/tháng — flash sale, bundle deal, seasonal campaign
• New product launches: iPhone mới, Samsung Galaxy mới — không có historical data
• Kênh online + offline: Demand shift giữa online và physical store khó predict
• Regional difference: Demand TP.HCM ≠ Hà Nội ≠ Đà Nẵng

Hệ thống Forecasting

Trước 2023: "Gut Feeling" + Excel

Aspect	Detail
Method	Category managers estimate demand dựa kinh nghiệm
Tool	Excel spreadsheet, copy paste từ năm trước
Accuracy	MAPE 32-45% (categories)
Problem	Thừa hàng → giảm giá sâu; Thiếu hàng → mất doanh số
Cost	~800 tỷ VND/năm excess inventory write-down

2023-2025: Data-Driven Forecasting System

TGDĐ build forecasting system theo 3 layers:

Layer 1: Baseline Demand (Holt-Winters)

Data: 3 năm weekly sales per store-category
Method: Holt-Winters (multiplicative seasonality)
Granularity: Store × Category × Week
Update: Weekly automated pipeline
Accuracy: MAPE 15% (baseline, no promotion)

Layer 2: Promotion Uplift Model

Historical promotions: 2,000+ past campaigns
Features:
  - Discount depth (10%, 20%, 30%, 50%)
  - Promotion type (flash sale, bundle, trade-in)
  - Duration (1 day, 3 days, 7 days)
  - Channel (online only, offline only, omni)
  - Category
  - Day of week, month

Model: Regression → predicted uplift multiplier
Output: demand_promo = baseline × uplift_multiplier

Layer 3: New Product Launch

No historical data → dùng "similar product" method:
  - iPhone 17 forecast ← benchmark from iPhone 16, 15 launch patterns
  - Samsung Galaxy S27 ← Galaxy S26, S25 patterns
  - Adjustment: market sentiment, pre-order data, Google Trends

Accuracy: MAPE 20-25% (acceptable for new launches)

Case: iPhone 16 Launch — Tháng 9/2024

Thách thức: TGDĐ cần forecast demand cho iPhone 16 series — không có historical data cho exact product. Phải quyết định nhập bao nhiêu unit cho 2,100+ cửa hàng trước launch date 2 tháng.

Forecasting approach:

Step	Method	Input	Output
1. Benchmark	iPhone 15 launch data	Week 1-12 sales pattern iPhone 15	Baseline launch curve
2. Adjust for market	Google Trends, pre-order signals	Search volume +22% vs iPhone 15 launch	Uplift factor: 1.18
3. Price sensitivity	iPhone 16 pricing vs iPhone 15	iPhone 16 Pro: +$100 vs 15 Pro	Downward adjustment: 0.92
4. Holt-Winters refit	iPhone 15 actual + adjustments	12-week forecast	Point forecast per store
5. Prediction interval	Bootstrap simulation	1000 simulations	P25, P50, P75 per week

Forecast vs Actual — iPhone 16 (tuần 1-8):

Week	Forecast (P50)	Actual	Error	Cumulative Error
Week 1	85,000	92,000	-7.6%	-7.6%
Week 2	62,000	58,000	+6.9%	-1.2%
Week 3	45,000	43,000	+4.7%	+0.6%
Week 4	38,000	40,000	-5.0%	-0.8%
Week 5	32,000	31,000	+3.2%	+0.1%
Week 6	28,000	27,500	+1.8%	+0.5%
Week 7	25,000	26,000	-3.8%	-0.3%
Week 8	22,000	21,500	+2.3%	+0.1%

Overall MAPE (8 tuần): 4.4% — outstanding cho new product launch!

Key insight: Sai lệch Week 1 là lớn nhất (-7.6%) vì hype effect khó predict. Nhưng cumulative error tự correct — forecast và actual converge từ Week 3. TGDĐ đã nhập theo P75 forecast cho Week 1-2 (để tránh stockout hype phase), rồi switch sang P50 từ Week 3.

Impact tổng thể

Metric	Before (2022)	After (2025)	Change
Overall forecast MAPE	38%	14%	-63%
Promotion forecast MAPE	45%	18%	-60%
New product launch MAPE	55%	22%	-60%
Excess inventory cost	800 tỷ VND	320 tỷ VND	-60%
Stockout rate	12%	4.5%	-63%
Inventory turnover	6.2x	8.8x	+42%

Saved: ~480 tỷ VND/năm — ROI của forecast system = 50x cost of building it.

Lessons Learned

1. Promotion là yếu tố khó forecast nhất: Baseline demand MAPE 15%, nhưng có promotion → MAPE 35-45% (old system). Cần riêng 1 layer cho promotion uplift.
2. New product launch = analog forecasting: Không có historical data → dùng "most similar product" as proxy. Adjust cho market signals.
3. Granularity trade-off: National forecast accurate nhưng useless cho store ordering. Store-level forecast noisy nhưng actionable. TGDĐ chọn: forecast ở regional × category, rồi allocate xuống store.
4. Forecast là team sport: DA build model, Category Manager adjust dựa market intel, Store Manager adjust dựa local knowledge. Final forecast = model + expert + local.
5. Start simple, improve iteratively: TGDĐ bắt đầu bằng Holt-Winters (3 tháng để deploy), rồi thêm promotion layer (6 tháng), rồi new product module (12 tháng). Không cần perfect system ngày 1.

---

Tổng kết — 3 bài học chung

Bài học	Walmart	Amazon	TGDĐ
Simple methods = strong baseline	Holt-Winters MAPE 12% (top SKUs)	ETS cho fast movers	HW baseline MAPE 15%
Prediction intervals essential	Safety stock = f(forecast error)	P10/P50/P90 quantiles	P25/P50/P75 cho inventory
Continuous re-forecasting	Weekly update	Daily update	Weekly update
Combine model + domain expertise	Weather + events overlay	Prime Day manual adjustment	Category Manager adjustment
Forecast accuracy = money	1% accuracy = $150M saved	98.5% in-stock target	480 tỷ VND saved/năm