Course: Supply Chain Digitization — Module 3: Analytics in SCM
Case study context — FMCG demand planning problem
Data setup — 1000 retailers, 7 independent variables, 1 target
Regression tree vs. classification tree — key difference
Regression tree output — all 7 nodes explained
How accuracy improves as nodes are added
4 business rules derived from the model
Worked prediction — Retailer A
Support interpretation
Next session: how to build regression tree in Python
Setting: Large FMCG company with retailer network across India (South/East/West/North)Problem: Demand planning head unsatisfied with forecast accuracy. Gap between actual and forecasted demand leads to either excess inventory OR lost sales.Trigger: Manager attended AI-ML training → understood that AI-ML is needed given how quickly demand patterns are changingGoal: Develop a better demand forecasting model to reduce the gap between actual and predicted demand → lower inventory holding cost + fewer lost salesModel chosen: Regression Tree (decision tree with continuous output) — predicts the ORDER QUANTITY for each retailer
★ Exam tip: Know why regression tree was chosen: (1) demand is a continuous variable (number of units) → regression tree not classification tree. (2) Non-linear relationships between size, age, promotions and demand. (3) Interpretable business rules for demand planners.
Aspect Classification Tree Regression Tree Target variable (Y) type Categorical (e.g. Fail / Not Fail) Continuous (e.g. Order quantity in units) Prediction at leaf node Majority class label (e.g. “Failed”) Mean (ȳ) of all Y values in that node Splitting criterion Gini index or Entropy (impurity-based) Mean Squared Error (MSE) / Variance reduction Used in this course for Predictive maintenance (Machine failure) Retailer order quantity prediction
Dataset: 1000 retailers (serial numbers 0 to 999). Each row = one retailer in one week.Train-test split: 700 observations for training (70%) | 300 for testing (30%)Target (Y): Order Quantity — continuous number of units ordered by that retailer in that week
# Variable Type / Unit Why It Matters for Demand X1 Region Categorical (South/East/West/North) Geographic demand patterns vary significantly by region. X2 Balanced Credit Amount Continuous (₹ Lakhs) Amount the retailer still owes the FMCG company. High outstanding balance → may affect their ordering behaviour. X3 Location Categorical (Urban / Semi-urban / Rural) Urban retailers typically serve higher footfall → higher demand expected. X4 Age of Retailer Continuous (Years) Older retailers have stronger customer relationships, loyal customer base → higher footfall. X5 Size of Retail Store Continuous (‘000 sq ft) Larger store → more products displayed → higher customer footfall. X6 Promotional Offer Binary (1 = offered, 0 = not offered) Promotions directly boost demand. A promotional week → significantly higher orders placed. X7 Number of Holidays Count (0, 1, 2, 3…) More holidays → more shopping occasions → higher demand in that week. Y Order Quantity (Target) Continuous (Units ordered) How many units the retailer orders in that week.
Model used: Regression tree | Training data: 700 obs | Tree depth: 2Variables selected by model: Size of store | Promotional offer | Age of store → model ignored Region, Balance Credit, Location, Holidays as less discriminatory
Node Type Conditions to Reach This Node Split Variable & Threshold Predicted Demand (ȳ) Obs. Support 0 (Root) Internal All training data Size ≤ 30.5K sq ft 2270 (baseline) 700 100% 1 Internal Size of the store ≤ 30.5K sq ft Promotion (0 or 1) 1902 612 87% 2 Internal Size of the store > 30.5K sq ft Age (≤ 17.5 yrs threshold) 4829 88 13% 3 (Leaf) LEAF Size ≤ 30.5K sq ft, Promotion = 0 STOP 943 198 28% 4 (Leaf) LEAF Size ≤ 30.5K sq ft, Promotion = 1 STOP 2360 414 59% 5 (Leaf) LEAF Size > 30.5K sq ft, Age ≤ 17.5 yrs STOP 2887 56 8% 6 (Leaf) LEAF Size > 30.5K sq ft, Age > 17.5 yrs STOP 8227 32 5%
★ In a regression tree, the leaf node prediction = mean (ȳ) of all Y values that fall into that node. Unlike classification tree which uses majority class. The more refined the node, the closer ȳ is to individual retailers’ actual demand.
Stage Information Used Predicted Demand Interpretation Node 0 (No info) None — random pick from 700 retailers 2270 (ȳ for all) Baseline: simple average of all 700 retailers’ orders. After Split 1 (Node 1) Store size ≤ 30.5K sq ft 1902 (↓ from 2270) Adding size info refines prediction for small stores. Now know they order less than average. After Split 1 (Node 2) Store size > 30.5K sq ft 4829 (↑ from 2270) Large stores order much more. Prediction jumps to 4829. After Split 2 (Node 4, Leaf) Size ≤ 30.5K + Promotion = 1 2360 (promo effect: +458) Same small store but running a promotion → demand surges. After Split 2 (Node 6, Leaf) Size > 30.5K + Age > 17.5 yrs 8227 (old + large = highest) Large AND old store = highest predicted demand. Loyal customer base + large display space = best combination.
Core logic: More relevant information about a retailer → more refined group it falls into → mean of that group is closer to its actual demand.
Rule Condition 1 (Store Size) Condition 2 (Promotion or Age) Retailer Profile Predicted Demand Support R1 (Node 3) Size ≤ 30.5K sq ft Promotion = 0 (No offer) Small store, no promotion → low demand week 943 units 28% R2 (Node 4) Size ≤ 30.5K sq ft Promotion = 1 (Offer given) Small store BUT promotion running → demand boosted 2360 units 59% R3 (Node 5) Size > 30.5K sq ft Age ≤ 17.5 yrs (Relatively new) Large but newer store — not yet established strong customer base 2887 units 8% R4 (Node 6) Size > 30.5K sq ft Age > 17.5 yrs (Old, established) Large AND old store → loyal customers, high footfall → highest demand 8227 units 5%
Given data for Retailer A: Region: West | Balance Credit: ₹10 lakh | Location: Urban | Age: 12 years | Store size: 8,000 sq ft (8K) | Promotional offer: Yes (1) | Number of holidays: 3
Check store size: 8K sq ft ≤ 30.5K sq ft → go to Node 1 (left branch)
Check promotion: Promotion = 1 → go to Node 4 (right branch of Node 1)
Result: Node 4 → Predicted demand = 2360 units
Support = % of training observations that fall into that leaf node.
Node 4 (59% support): Most reliable — covers majority of small-store + promotion retailers.Node 6 (5% support): Lowest confidence — only 32 of 700 are large + old stores.
Yes: Splitting nodes further with additional features → lower mean squared error within each leaf → predictions get closer to individual retailer actual demand.Trade-off: Deeper tree = better training accuracy BUT higher risk of overfitting. Same stopping criteria apply as in classification tree (Max depth limit, Min observations, Min variance reduction).
Case: FMCG company demand planning head → uses regression tree to predict retailer order quantities.Regression tree vs. classification tree: Continuous target → leaf predicts mean (ȳ). Splitting criterion = variance reduction (not Gini).Model selects 3 of 7 features: Store size | Promotional offer | Age of store4 leaf nodes → 4 business rules: 943 (small, no promo) | 2360 (small + promo, 59% support) | 2887 (large + young) | 8227 (large + old — highest demand)Retailer A prediction: Size = 8K sq ft (≤ 30.5K) | Promotion = 1 → Node 4 → Predicted demand = 2360 unitsNext session: Build regression tree in Python + demand forecast error metrics (MAE, RMSE, MAPE)