Time Series Forecasting • Multi-horizon ForecastingEasy⏱️ ~2 min
What is Multi-Horizon Forecasting?
Multi-horizon forecasting predicts a vector of future values for a single time series in one unified model. Instead of predicting just tomorrow's sales, you predict the next 7 days, or 28 days, or even 60 minutes at 5 minute intervals. Each prediction horizon represents a different point in the future, and the entire sequence is generated together rather than one step at a time.
The key difference from single step forecasting is that multi-horizon systems must explicitly model how uncertainty grows over time. A forecast for tomorrow might be accurate within 5%, but a forecast for 30 days out could have 30% error. Production systems output probabilistic forecasts, typically as quantiles (like 10th, 50th, 90th percentile) rather than single point estimates. For example, Amazon's retail planning uses quantile forecasts where the 90th percentile (0.9 quantile) tells you the inventory level that will cover demand 90% of the time, preventing stockouts while balancing holding costs.
Consider a retail demand system forecasting 5 million SKUs across 20 regions, 28 days ahead. Each forecast produces 7 quantiles (0.05, 0.1, 0.2, 0.5, 0.8, 0.9, 0.95). That's 100 million series times 28 horizons times 7 quantiles, generating 19.6 billion numbers (roughly 157 GB) per forecast run. This output feeds into replenishment systems, pricing engines, and workforce planning.
In contrast, a real-time marketplace like Uber forecasts the next 60 minutes in 5 minute increments for 10,000 zones per city. With 100 cities, each update computes 12 million outputs, all within a 300 ms latency budget to drive surge pricing and driver incentives.
💡 Key Takeaways
•Produces a vector of future predictions (7 days, 28 days, 60 minutes) in a single model run, not one prediction at a time
•Must explicitly represent growing uncertainty. A retail system's horizon 1 might have 5% error, horizon 28 reaches 30% error
•Output is probabilistic (quantiles or distributions), not point estimates. Amazon uses 0.9 quantile forecasts for inventory that prevent stockouts 90% of the time
•Massive scale in production. A 5 million SKU retail system generates 19.6 billion numbers (157 GB) per nightly forecast run with 7 quantiles over 28 horizons
•Real-time marketplaces like Uber forecast 60 minutes ahead at 5 minute granularity for 10,000 zones per city, computing 12 million outputs every 1 to 5 minutes within 300 ms
•Feeds directly into operational systems including replenishment, pricing, surge control, and workforce scheduling
📌 Examples
Amazon retail: 5 million SKUs, 28 day horizon, 7 quantiles (0.05 to 0.95), nightly batch run produces 157 GB of forecasts for inventory planning
Uber marketplace: 10,000 zones per city, 100 cities, 12 horizons at 5 minute intervals, generates 12 million outputs every 1 to 5 minutes under 300 ms latency for surge pricing
Airbnb pricing: Forecasts booking demand 90 days ahead with daily granularity and 5 quantiles to guide dynamic pricing for millions of listings