Gradient Boosting for Time Series

Best for: Feature-based forecasting

How it works

Reframes forecasting as supervised regression: the series is embedded into feature vectors $x_t=[y_{t-1},y_{t-2},\dots,y_{t-l}]$ of lagged values, rolling statistics, calendar dummies, and exogenous variables, with target $y_{t+h}$. A boosted ensemble $F_M(x_t)=\sum_{m=1}^{M}h_m(x_t)$ is then built stage-wise, each tree fitting the gradient of the loss at the previous prediction. Powerful for nonlinear interactions among engineered features, but it needs careful lag/window engineering and leakage-free validation.

Common fields

Demand forecasting · pricing · operations