Demand Forecasting Across Industries
"Demand forecasting" is not a monolith. The algorithmic choices, data dependencies, and mathematical loss functions fundamentally change depending on the physical nature of the product and its economic lifecycle.
This article contrasts four highly distinct domains: Fresh Food, Fashion, Consumer Electronics, and Heavy Machinery (Spare Parts).
1. Fresh Food: The Battle Against Physical Decay
- The Core Problem: High physical perishability (1–7 day shelf life) and sensitivity to daily exogenous factors.
- Data Dependencies: Intra-day sales, hyper-local weather APIs, competitor promotions, and strict batch-level expiry data.
- Algorithmic Approach: Models (like LightGBM or Temporal Fusion Transformers) are trained with asymmetric loss functions (e.g., pinball loss). Because the cost of over-forecasting (100% waste) is radically different from under-forecasting (lost margin), the model explicitly biases the forecast.
- Key Distinction: Forecast errors instantly materialize as physical waste. See FreshFoodDemandForecasting.
2. Fashion & Apparel: Trend Volatility and Markdown Optimization
- The Core Problem: "Zero-history" forecasting. Every season introduces entirely new SKUs with no historical baseline. Products don't physically decay, but their economic value decays rapidly as trends shift.
- Data Dependencies: Attribute-level analysis (color, fabric, cut, style), social media sentiment, and early in-season sell-through velocity.
- Algorithmic Approach:
- Pre-season: Machine learning algorithms cluster new items with historical "look-alikes" based on attributes to generate initial buy quantities.
- In-season (Markdown Optimization): The focus shifts from predicting absolute demand to quantifying price elasticity. The model calculates how volume will react to a 10% vs. 30% discount. Adaptive algorithms trigger "clear-as-you-go" markdowns early in the lifecycle to protect margins and clear inventory before the season ends, avoiding deep, profit-eroding blanket discounts at the end of life.
3. Consumer Electronics: Cannibalization and Hierarchies
- The Core Problem: Extremely complex Bills of Materials (BOM), long manufacturing lead times, and rapid innovation cycles that cause old products to instantly obsolete new ones.
- Data Dependencies: Product launch calendars, component availability, and promotional cycles.
- Algorithmic Approach: Hierarchical forecasting. A forecast is not just generated for a "laptop," but must be mathematically reconciled down to the display panel, processor, and battery.
- Key Distinction: The models must explicitly account for cannibalization. If the iPhone 16 is launched, the algorithm must predict the exact decay curve of iPhone 15 demand, preventing massive write-downs of obsolete stock.
4. Heavy Machinery & Spare Parts: Intermittent Demand
-
The Core Problem: Intermittent, "lumpy" demand. An airplane part might not be needed for 18 months, and then suddenly three are needed in one week. Traditional models like Simple Exponential Smoothing (SES) or ARIMA fail completely, oscillating wildly or predicting a constant fraction (e.g., "0.1 parts per month"), which is physically meaningless.
-
Data Dependencies: Installed base data (how many machines are currently active in the field) and sensor-based telemetry (predictive maintenance).
-
Algorithmic Approach: Croston's Method and its variants (SBA, TSB).
Croston's method separates the forecast into two independent SES updates:
- Demand Size (z): How many parts are needed when an order occurs.
- Inter-demand Interval (p): How many periods pass between orders.
The final forecast per period (f_t) is the ratio of the estimated size to the estimated interval:
$f_t = \frac{\hat{z}_t}{\hat{p}_t}$
Variants like the Teunter-Syntetos-Babai (TSB) method improve this by updating the probability of demand occurring at every time step, allowing the forecast to properly decay toward zero if a machine type is retired from the field.
Summary Matrix
| Industry | Primary Constraint | Key Algorithm / Approach | Forecasting Horizon |
|---|
| Fresh Food | Physical Spoilage | Asymmetric Loss (Quantile Regression) | Intra-day / Daily |
| Fashion | Trend Obsolescence | Attribute Clustering & Markdown Elasticity | Seasonal |
| Electronics | Component Complexity | Hierarchical Reconciliation & Cannibalization | Mid-Term (Months) |
| Spare Parts | Intermittent Demand | Croston's Method / Installed Base | Long-Term (Years) |