Skip to content

Latest commit

 

History

History
31 lines (16 loc) · 1.43 KB

File metadata and controls

31 lines (16 loc) · 1.43 KB

Time-series glossary

Autocorrelation — correlation between a series and a lagged version of itself.

Backtesting — simulating historical forecasts using only information available before each forecast origin.

Baseline — a simple reference forecast that an advanced model should improve upon.

Drift forecast — extrapolation of the average change observed between the first and last training values.

Expanding window — evaluation where the training set grows as the forecast origin moves forward.

Forecast horizon — number of future periods predicted.

Lag — a past observation used to explain or predict a later value.

Leakage — use of information during training that would not be available at prediction time.

MAE — mean absolute error; average absolute distance between actual and predicted values.

RMSE — root mean squared error; penalizes larger errors more strongly than MAE.

Seasonal naive — forecast that repeats the most recent observed seasonal cycle.

Seasonal period — number of observations in one repeating cycle, such as 12 for monthly annual seasonality.

sMAPE — symmetric mean absolute percentage error; a scale-relative percentage-style metric.

Temporal holdout — final chronological portion of a series reserved for evaluation.

Walk-forward validation — repeated forecasting from successive historical origins while preserving temporal order.