Definition
A forecast combination blends predictions from different models for the same target, horizon, and forecast origin. A simple average is one starting point; estimated weights are another. Combining models differs from adding demand across products or time periods.
Why It Matters
- A replenishment team may receive conflicting SKU forecasts from seasonal and promotion-aware models.
- The Commerce Intelligence OS framework needs a documented way to turn competing predictions into a planning input.
How It Works
- Align units, SKU scope, forecast origin, and horizon before combining model outputs.
- Start with a transparent average and compare it with individual models on historical future windows. Estimate any weights using only data available before each evaluation window.
- Record missing-model behavior so an unavailable component does not silently change the planning rule.
- Evaluate the resulting forecast and its uncertainty separately. Dependence between model errors matters; averaging interval endpoints is not a general uncertainty solution.
Ecommerce Example
Context: Illustrative example: three models predict 80, 100, and 120 units for the same product next week.
Recommended move: Their equal-weight mean is 100 units. Compare this rule with each component across earlier planning windows before adopting it.
Why it matters: The numbers show the arithmetic only. They do not establish that buying 100 units is optimal or that iKawn has achieved a forecast improvement.
iKawn Framework
Align
The iKawn framework attaches forecasts to one commercial target.
Combine
Keep component models and the blend rule traceable.
Evaluate
Compare retained demand outcomes with each candidate forecast.
Decide
Connect the selected forecast to inventory constraints and merchant review.
Concise Summary
Combine predictions only after aligning their targets. Validate the blend against its component models and keep the planning rule reproducible.