What is Multi-Horizon Forecasting?

Definition: Multi-horizon forecasting predicts values at multiple future time points simultaneously—not just tomorrow, but tomorrow, next week, and next month in a single model pass. Different business decisions require different horizons: inventory needs 2-4 week forecasts, capacity planning needs 3-6 month forecasts.

Why Single-Horizon Falls Short

Single-horizon models optimize for one specific lookahead. A model optimized for 1-day forecasts performs poorly at 7-day horizon because error patterns differ. Training separate models per horizon is expensive and creates inconsistencies—the 7-day forecast might exceed the sum of daily forecasts. Multi-horizon models produce coherent forecasts across all horizons.

Uncertainty Grows with Horizon

Forecast accuracy degrades as horizon increases. 1-day forecasts might achieve 5% MAPE, 7-day forecasts 15%, 30-day forecasts 30%. Multi-horizon models should output uncertainty estimates (prediction intervals) that widen appropriately with horizon. Decision-makers need both point forecasts and confidence bounds.

Key Insight: Different horizons serve different decisions. Short-horizon high-accuracy forecasts drive operational decisions (staffing, logistics). Long-horizon lower-accuracy forecasts drive strategic decisions (capacity investment). One model serving both requires explicit multi-horizon design.

Business Applications

Demand forecasting: retailers need 1-day (replenishment), 2-week (promotions), 3-month (seasonal ordering) forecasts. Energy: 1-hour (dispatch), 1-day (scheduling), 1-week (maintenance) horizons. Finance: 1-day (trading), 1-month (budgeting), 1-year (planning) forecasts. Each application has natural horizon granularities matching decision cycles.