Week 7 | Session 1: AI-ML for Demand Forecasting & Demand Planning
Course: Supply Chain Digitization — Module 3: Analytics in SCM
Session Agenda
Section titled “Session Agenda”Why Demand Planning Matters — The Cost of Forecast Errors
Section titled “Why Demand Planning Matters — The Cost of Forecast Errors”Demand planner’s goal: Forecasted demand ≈ Actual demand. Any deviation in either direction creates cost.
| Scenario | What Happens | Immediate Impact | Long-term Impact |
|---|---|---|---|
| Actual > Forecast (Under-forecasting) | Demand exceeds available stock. Cannot fulfil all orders. | Lost sales. Customers go to competitors. | Customer churn. Revenue loss. Market share erosion. |
| Actual < Forecast (Over-forecasting) | More stock produced/ordered than demand. Excess inventory accumulates. | Inventory holding cost. Spoilage if perishable. | Tied-up capital. Warehouse cost. Markdowns to clear stock. |
| Actual = Forecast (Accurate forecast) | Production and procurement aligned with real demand. | No stockouts. No excess inventory. | Efficient SC. Happy customers. Optimal working capital. |
- Perishable products: Over-forecasting is especially costly — spoilage adds physical loss on top of holding cost
- Lost sales: Harder to quantify than excess inventory — but more damaging long-term through customer churn
Why AI-ML is Needed — 4 Vs of Big Data in Demand Context
Section titled “Why AI-ML is Needed — 4 Vs of Big Data in Demand Context”Traditional forecasting tools: Moving average, exponential smoothing, ARIMA — good for stable trends and seasonality. Problem: In today’s environment, demand data has all 4 problematic Vs simultaneously → traditional tools fail to handle this complexity.
Big Data vs. Traditional Forecasting
Section titled “Big Data vs. Traditional Forecasting”| V | Big Data Characteristic | How It Appears in Demand Data | Why Traditional Forecasting Fails |
|---|---|---|---|
| V1 Volume | Huge data | Orders placed 24/7 via omni-channel → millions of data points continuously generated | Not built to process and learn from millions of continuous data points in real-time |
| V2 Variety | Multiple data types | Customer review as text | Image |
| V3 Velocity | Real-time data | Orders, cancellations, cart additions happening 24/7 in real time | Batch-based traditional forecasting too slow to respond to real-time shifts |
| V4 Variability | Unstable patterns | Customer preferences shift unpredictably due to trends, promotions, policies | Capture seasonality/trend but cannot model social-media-driven swings |
AI-ML Models for Demand Forecasting — Overview
Section titled “AI-ML Models for Demand Forecasting — Overview”Key AI-ML models used for demand forecasting
Section titled “Key AI-ML models used for demand forecasting”| Model | Output Type | Key Assumption / Feature | Demand Forecasting Use Case |
|---|---|---|---|
| Linear Regression | Continuous | Relationship between dependent and independent variable is LINEAR | Predict units sold based on: historical sales + pricing + promotions. Simple baseline. |
| Logistic Regression | Categorical (probability) | Dependent variable is binary or categorical — probability of an event | Predict probability of successful adoption of a new product launch or promotion response. |
| Decision Tree | Categorical or Continuous | Rule-based splits. Handles non-linear relationships. | Identify KEY DRIVERS of demand variation (price, season). Generates interpretable business rules. |
| Random Forest | Categorical or Continuous | Ensemble of multiple decision trees. Reduces overfitting. | More robust demand forecasting when single tree overfits. Handles noisy data better. |
| Boosting Algorithms (XGBoost, LightGBM) | Categorical or Continuous | Sequential ensemble: each model corrects errors of the previous one. | Dynamic forecasting — adapts to changing conditions (competitor pricing, economy). Widely used in industry. |
| Neural Networks (LSTM, CNN) | Continuous or Categorical | Captures highly complex, non-linear patterns. Handles structured AND unstructured data. | Best for unstructured data (text reviews, image feedback) and complex time-series (LSTM). |
Deep Dive into Key Models
Section titled “Deep Dive into Key Models”- Linear Regression: Simple, but underperforms when demand has sharp, non-proportional responses to changes.
- Logistic Regression: Links marketing spend to expected demand uplift.
- Decision Tree: Identifies WHICH factors drive demand variation most. Limitation: May overfit to training data.
- Random Forest: Averages multiple trees → more stable, less sensitive to noise. Trade-off: Less interpretable.
- Boosting Algorithms: Most widely used in competitive ML and industry demand planning teams today — fast, accurate, scalable.
- Neural Networks: LSTM is widely used for demand time-series to capture long-term dependencies. Trade-off: Hardest to interpret.
External Factors — What Must Be Monitored Continuously
Section titled “External Factors — What Must Be Monitored Continuously”Demand planner’s challenge: Internal data alone is not enough. External factors constantly shift demand patterns and must be incorporated into the model.
| External Factor | How It Affects Demand | How AI-ML Incorporates It |
|---|---|---|
| Competitor promotions | Customer shifts to competitor → your demand drops | Boosting models ingest promotion data as a feature and adjust forecast |
| Competitor pricing | Lower price → demand shifts. Loyal customers switch. | Price elasticity modelled. Model predicts demand reduction and adjusts plan. |
| Overall economy / GDP | Economy growing → demand rises. Contracting → spending limited. | Macroeconomic indicators included as features in boosting/regression models |
| Government policy changes | Regulations, taxes, or subsidies suddenly shift demand | Treated as dummy variables or structural breaks in time-series models |
| Social media trends | Viral product → demand spike. Negative coverage → drop. | NLP on social media posts feeds into neural network models to detect sentiment early |
How to Select the Right Model — Framework
Section titled “How to Select the Right Model — Framework”- Define the output type: Is demand a number (continuous)? → Regression models. Is it a category? → Classification models.
- Assess data type available: Structured numerical data only → Linear regression, Decision tree, Random forest, Boosting | Unstructured (text, image, audio) → Neural networks
- Check demand pattern: Stable trend + seasonality → Traditional tools may suffice | High variability, non-linear → AI-ML needed
- Consider interpretability need: Need to explain to management? → Decision tree (most interpretable) | Accuracy over interpretability? → Random forest, XGBoost, LSTM
- Benchmark models: Run multiple models → compare error metrics (MAE, RMSE, MAPE) → select model with lowest error on test data
- Monitor and retrain: Demand patterns evolve. Re-train model periodically or set up continuous learning pipeline to stay current.
Session Summary
Section titled “Session Summary”- Why demand planning: Forecast ≠ Actual → either lost sales (under) or inventory cost (over). Accurate forecast = revenue protection + cost control.
- Why AI-ML needed: 4 Vs of demand data (Volume, Variety, Velocity, Variability) exceed what traditional forecasting tools can handle.
- 6 AI-ML models: Linear regression | Logistic regression | Decision tree | Random forest | Boosting (XGBoost, LightGBM) | Neural networks (RNN, LSTM, CNN)
- External factors to monitor: Competitor promotions + pricing | GDP / economy | Government policy | Social media trends
- Model selection: Match model to output type, data type, pattern complexity, interpretability need, and error benchmark.
- Next session: Case study — apply an AI-ML model to a demand forecasting problem. Introduce error metrics (MAE, RMSE, MAPE).