Prophet-style additive models

Best for: Business forecasting with seasonality/holidays Aliases: Prophet, Generalized Additive Model

How it works

Decomposes the series additively as $y(t)=g(t)+s(t)+h(t)+\varepsilon_t$, where $g(t)$ is a piecewise-linear (or saturating logistic) trend with changepoints, $s(t)$ models seasonality via a Fourier series, and $h(t)$ adds holiday/event effects. The components are jointly fit by Bayesian (Stan) or L-BFGS regression, with the number of Fourier terms and the changepoint prior controlling flexibility. Forecasts come from extrapolating each fitted component forward, giving interpretable, uncertainty-aware predictions robust to missing data.

When to use

Business forecasting with multiple seasonalities, holidays, and regressors where ease and interpretability matter.

Watch out

Often underperforms gradient boosting; defaults assume additive components; changepoint detection needs tuning.

Common fields

Sales · marketing · web traffic