- Long seasonal periods can require many parameters if every seasonal position gets its own state.
- Fourier terms describe smooth repeating shapes using pairs of sine and cosine waves.
| Concept | What it means | Northstar example |
|---|---|---|
| K controls flexibility. |
| Northstar can start with a few smooth annual waves, then add harmonics only if time-aware tests show that the extra seasonal flexibility helps. |
| Different periods can coexist. |
| Use weekly waves and annual waves together to describe different repeating shopping patterns in the same daily sales series. |
| Events are not just smooth waves. |
| Use a named event window for the sharp Black Friday lift instead of adding many smooth annual waves to chase its spike. |
Add waves, add flexibility
A Fourier regression represents a repeating seasonal profile as an intercept plus sine and cosine pairs. Start with the reconstruction above; then compare how the same choice behaves when Northstar has only one year of history instead of five.
The question it answers: How can a handful of waves rebuild a whole year's shape, and when do extra waves start fitting noise?
1 · Rebuild the yearly shape by adding waves
The waves being added · first 3 of 6
Added together · fitted annual shape
2 · With little data, extra harmonics fit noise
Fitted to 3 years of weekly sales
Next year · forecast against the known shape
More history is stabilizing the fit. Keep checking next-year error, not only the in-sample line.
Fine print & method
Constructed weekly sales with a known annual shape containing four harmonic pairs, plus deterministic Gaussian noise. Fits are ordinary least squares on past weeks only; the next-year score compares each fitted shape with the noise-free known shape. The first panel fits the average weekly profile over three years. With one year, six pairs overfit; with five, the true fourth wave is estimated better while spurious higher waves shrink. This demonstrates model selection, not a production forecast. A moving event such as Black Friday still needs an explicit event feature; a fixed annual wave cannot know its date.