Additive Decomposition Forecasting
Additive forecasters model a series as an interpretable sum of trend, seasonal, and holiday components with uncertainty.
Forecasting as curve fitting
A popular family of business forecasters treats prediction as fitting an interpretable additive model: y(t) = trend(t) + seasonality(t) + holidays(t) + noise. Rather than the recursive equations of ARIMA, each component is an explicit function of time, which makes the forecast easy to inspect, explain, and adjust by hand. This design targets analysts who need transparent, robust forecasts without deep time-series expertise.
The components
- Trend: a piecewise-linear or saturating growth curve with automatically placed changepoints where the slope shifts
- Seasonality: periodic patterns represented by Fourier terms, one set per period (weekly, yearly)
- Holidays and events: user-supplied dates that add bumps of learned magnitude
- Uncertainty: intervals from trend-changepoint variability and observation noise
Why practitioners like it
The model handles missing data and outliers gracefully, tolerates irregular sampling, and lets a domain expert inject known events such as promotions or closures. Because components are additive and named, a stakeholder can read off how much of a forecast comes from trend versus season versus a holiday, which supports trust and debugging in a way opaque models do not.
Strengths and cautions
This approach excels on human-scale business series with strong multi-period seasonality and known events. It is weaker when dynamics are driven by short-term autocorrelation or exogenous feedback, where ARIMA or state-space methods fit better. As always, the model must be validated with rolling-origin backtesting and compared against simple baselines; its interpretability is a feature, not a substitute for honest evaluation.
Automatic changepoint placement can overreact to noise, so the changepoint flexibility is a key regularization dial.