The hardest forecasting decision is not which model to use!
When you have many entities to forecast, don't just build a better model. Build a better routing system.
The Problem
In a multi-entity time series forecasting problem, I spent more time on routing than on model tuning. The series were too heterogeneous to treat uniformly:
- Some were decelerating
- Some were accelerating
- Some were near-zero
- Some were noisy and ambiguous
Applying the same model to all would produce structurally wrong forecasts. The same model optimized for an accelerating trajectory will systematically overforecast a decelerating one.
The Classification Layer
Before fitting anything, I built a classification layer:
- Compute an elasticity score across two rolling windows
- Classify each entity into a trajectory regime
- Route to a different forecasting strategy per regime
Decelerating
Pick the most conservative model
Accelerating
Pick the median prediction for robustness
Ambiguous
Let recent holdout validation error decide
Key Takeaway
Design insight: Most forecasting pipelines invest heavily in model selection but apply it uniformly. Before asking which model is best, ask what state your series is in. Build a routing system, not just a better model.