STOP 22 / 31Waves for seasons. ARIMA for the rest.
Layout
STOP 22 · More models

Waves for seasons. ARIMA for the rest.

EXPLORE~6 min
Back to route map
  • Dynamic harmonic regression combines Fourier features for repeating seasonal shapes with ARIMA errors for remaining temporal dependence.
  • It is useful for long or multiple seasonal periods.
Dynamic harmonic regression: concepts and Northstar examples. Illustrative Northstar examples, not measured business results.
ConceptWhat it meansNorthstar example
Split the jobs.
  • Regression handles weekly and yearly waves, event windows, and known drivers.
  • The ARIMA error model handles autocorrelation left after those effects.

Fourier and event features explain Northstar's repeating calendar shape; ARIMA errors handle remaining serial dependence after that regression.

Choose seasonal complexity.
  • The number of Fourier pairs controls smoothness.
  • Tune it without leaking the evaluation period; holdout checks matter even when an information criterion is used.

Northstar selects the number of seasonal waves inside validation rather than choosing the order that best fits the held-out Black Friday period.

Seasonal coefficients are typically fixed.
  • Standard Fourier regression assumes a repeating seasonal shape with fixed coefficients.
  • If the pattern evolves, consider adaptations or alternative models.
  • A fitted standard DHR seasonal shape repeats with fixed coefficients.
  • If Northstar's weekday habits evolve, that needs an explicit adaptation or another approach.
FIG 22.1

Waves for seasons. ARIMA for the rest.

WHAT THIS SHOWS

Fourier terms describe Northstar's repeating annual shape. A regression alone assumes the remaining errors are unrelated; DHR adds an ARIMA error model for serial dependence. Toggle the AR correction and compare the held-out forecast.

The question it answers: How does dynamic harmonic regression split a seasonal shape from the autocorrelation left behind?

Northstar annual forecast: Fourier regression with and without autocorrelated errors$70k$85k$100k$115kForecast originLast 120 training daysToday+90 days
Actual sales · evaluation onlyFourier regression forecastFourier + AR(1) errors

Held-out absolute error by horizon · lower is better

Absolute forecast error by held-out day for Fourier regression and AR-corrected forecasts−$2k$4k$10k$15kDay 1Day 45Day 90
Fourier regression errorWith AR(1) error correction
Residual ACF at lag 1 · regression0.83
Residual ACF at lag 1 · after AR correction0.01
Estimated AR coefficient ρ0.84
Held-out MAE · selected$4,069
MAE · regression only$4,742
MAE · with AR correction$4,069

The AR(1) correction uses the last training residual and lets its expected effect decay at rate ρ. It reduces the held-out error and the lag-one residual correlation in this constructed example.

Fine print & method

Constructed daily sales: three Fourier pairs for a smooth annual shape plus an AR(1) error process with ρ = 0.85. Fourier coefficients and the single AR coefficient are estimated only on the first 730 days; the final 90 days are held out. This is a deliberately small DHR demonstration, not a full ARIMA/SARIMAX fit, model-selection procedure, or causal analysis. The displayed ACF after correction is the lag-one correlation of prewhitened residuals. Real data need seasonal-order, stationarity, event, and predictor checks.

CHECK YOURSELF · CHOOSE AN ANSWER

What is the ARIMA component modeling here?

Your learning note

Private · just for you