Demand forecasting predicts how much of a product or service will be needed in the future. It drives inventory planning, staffing, pricing, and supply chain decisions. Getting it wrong means either stockouts (lost sales) or overstock (wasted capital).

Where It Is Used

Demand forecasting is critical across industries:

Retail: SKU-level forecasting for inventory optimization
E-commerce: Warehouse stocking, delivery capacity planning
Manufacturing: Production scheduling, raw material ordering
Energy: Load forecasting for grid management
Transportation: Ride demand, flight booking predictions
Hospitality: Hotel occupancy, restaurant reservations

Products: Amazon fulfillment, Uber surge pricing, Walmart inventory, Starbucks staffing.

Typical Data You Will See

Aspect	Typical Values
Granularity	Hourly, daily, weekly
History length	2-5 years
Entities	1K to 1M SKUs/locations
Exogenous features	10-50 variables
Forecast horizon	1 day to 52 weeks

Common table structure:

date | sku_id | location_id | sales_qty | price | promo_flag | stock_level

External data sources: - Calendar features: holidays, events, weekdays - Weather data: temperature, precipitation - Marketing: promotions, campaigns, ad spend - Economic indicators: GDP, consumer confidence - Competitor actions: pricing, launches

What You Do in Practice

Define forecast scope - What entities, what horizon, what frequency?
Gather historical data - Sales, inventory, stockouts, external factors
Handle data issues - Missing days, zeros vs stockouts, outliers
Identify patterns - Trend, seasonality, holiday effects, promotions
Build hierarchy - Product categories, regions, aggregation levels
Train baseline model - Simple exponential smoothing or seasonal naive
Add ML models - Gradient boosting with lag features, or deep learning
Reconcile forecasts - Ensure hierarchy consistency (store forecasts sum to region)
Evaluate properly - Time-based splits, multiple horizons, business metrics
Deploy and monitor - Track accuracy by segment, retrain on schedule

Baseline First

Two baselines every forecasting project needs:

Seasonal naive: Predict same value as same period last year
Captures yearly seasonality
No tuning required
Often surprisingly competitive
Simple exponential smoothing: Weighted average of recent observations
Captures level and trend
Three parameters to tune
Fast to compute

If your ML model cannot beat seasonal naive by 10%+, investigate data quality before adding complexity.

Common Models and When to Choose

Model	Best For	Limitations
Exponential Smoothing (ETS)	Single series, clear patterns	No external features
ARIMA/SARIMA	Statistical analysis, confidence intervals	Complex tuning, slow
Prophet	Business data with holidays, multiple seasonalities	Can overfit, slow at scale
XGBoost with lag features	Many features, complex interactions	Requires careful feature engineering
LightGBM	Large scale, fast training	Same as XGBoost
DeepAR	Many related series, learned patterns	Requires significant data
Temporal Fusion Transformer	Complex multi-horizon, interpretability	Computationally expensive
N-BEATS	Strong benchmark, no feature engineering	Black box

Industry default: Start with ETS or Prophet for baselines, move to gradient boosting or deep learning for production.

Feature Engineering Patterns

Lag features: - Sales 1, 7, 14, 28, 364 days ago - Rolling mean/std over 7, 28, 90 days - Difference from same period last year

Calendar features: - Day of week, month, quarter (cyclical encoding) - Holiday indicators (with lead/lag effects) - Payday indicators, school holidays

Promotional features: - Promo active flag - Discount depth - Days since/until promotion - Promo type encoding

External features: - Weather (temperature, rain probability) - Events (sports, concerts, local events) - Economic indicators

Hierarchical features: - Category-level trends - Store-level seasonality - Cross-series patterns

Evaluation and Metrics

Primary metrics:

Metric	Formula	Use Case
MAPE	Mean Absolute Percentage Error	Easy to interpret, but fails on zeros
WMAPE	Weighted MAPE	Handles volume differences
RMSE	Root Mean Squared Error	Penalizes large errors
MAE	Mean Absolute Error	Robust to outliers
Bias	Mean Error (signed)	Detect systematic over/under forecasting

Business metrics: - Service level: % of demand fulfilled - Inventory turns: How fast stock rotates - Stockout rate: % of time out of stock - Overstock cost: Capital tied in excess inventory

Pitfalls: - Evaluating only on aggregate (hides poor SKU-level performance) - Using MAPE when zeros are common - Not evaluating at the horizon that matters for decisions - Ignoring bias (systematic over or under forecasting)

Failure Modes and Debugging Checklist

[ ] Data leakage: Using future information (e.g., actual promo results)
[ ] Stockout confusion: Zeros are stockouts, not low demand
[ ] Outlier distortion: Black Friday, COVID effects dominate model
[ ] Seasonality mismatch: Model trained on wrong seasonal period
[ ] Promotion blindness: Model does not capture promo effects
[ ] New product cold start: No history for new SKUs
[ ] Hierarchy inconsistency: Store forecasts do not sum to total
[ ] Regime change: Fundamental demand shift not captured

Production Checklist

Before deployment: - [ ] Backtesting across multiple time periods - [ ] Performance acceptable across all hierarchy levels - [ ] Cold start strategy for new products - [ ] Outlier handling defined - [ ] Forecast intervals generated (not just point forecasts)

After deployment: - [ ] Daily accuracy monitoring by segment - [ ] Bias tracking (directional errors) - [ ] Automatic alert on accuracy degradation - [ ] Weekly/monthly retraining schedule - [ ] Feedback loop with actual sales data - [ ] Human override capability for special events

Tools and Stack

Category	Tools
Statistical models	statsmodels, Prophet, statsforecast
ML models	XGBoost, LightGBM, scikit-learn
Deep learning	PyTorch, GluonTS, Darts, NeuralForecast
Scalable forecasting	Spark, Ray, Dask
Visualization	Plotly, matplotlib
Production	Airflow, MLflow, cloud ML platforms

Example Mini-Case

Scenario: Grocery retailer needs daily SKU-level forecasts for 10K products across 200 stores.

Data: 3 years daily sales, promotions, holidays, weather.

Approach: 1. Baseline: Seasonal naive (same day last year) - WMAPE 35% 2. Feature engineering: Lag features, promo indicators, holiday effects 3. Model: LightGBM with 28-day forecast horizon 4. Hierarchy: Trained at SKU-store level, reconciled to category and store totals 5. Result: WMAPE 22%, 37% improvement

Production setup: - Nightly batch job computing next 28 days - Weekly model refresh - Daily accuracy dashboard - Alerts on WMAPE > 30% for any category

Business impact: 15% reduction in stockouts, 8% reduction in overstock.

Index: ML Field Guide
Time Series Anomaly Detection
Financial Time Series
Feature Engineering Fundamentals

Try it live

Everything above runs in your browser — open Demand Forecasting Simulator and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.

▶ Open Demand Forecasting Simulator simulation

Demand Forecasting