Lag Features for Time Series Forecasting

A tabular model can forecast a time series when each training row represents a forecast origin and its features contain information available at that origin. Lag features are shifted values such as y_(t-1), y_(t-7) or y_(t-24). They provide recent history and seasonal repeats as predictors for a future target y_t or y_(t+h). Rolling statistics summarize a window of past values, while calendar features encode known timing patterns.

Alignment is the central risk. To forecast y_t at time t, features must not include y_t or later values. A seven-day rolling average must end at t-1 if the target is y_t. A common error is computing a rolling mean before shifting, which includes the target observation. Another is random train-test splitting, where future rows can train a model evaluated on earlier periods and where overlapping windows share near-identical information.

Consider a hypothetical weekly-demand forecast. Features may include demand one week ago, four weeks ago, a trailing four-week mean, weekday and planned holiday indicator. Each value's availability should be traced: holiday calendars are known, while actual weather may not be known unless a weather forecast is used. A feature can be historically correlated yet unavailable at decision time.

Use time-ordered validation such as rolling or expanding windows, with forecast horizons matched to use. A gap may be needed when labels arrive with delay or features span periods that overlap a test window. Fit scalers and imputers only within each training fold. For multi-step forecasting, decide whether to forecast recursively using prior predictions, predict each horizon directly, or use a strategy with future-known covariates. Evaluate against simple seasonal baselines and monitor how performance changes over time. Lag-feature learning provides flexible nonlinear prediction, but its reliability depends on disciplined information timing and the stability of the sequence process.