Predicting over time.
WHAT MAKES IT DIFFERENT
Observations are ordered and dependent The future must not inform the past Patterns repeat seasonally
WHAT COMPONENTS EXIST
Trend Seasonality, possibly several cycles Cyclical variation Noise
WHAT CLASSICAL METHODS PROVIDE
Well-understood models for trend and seasonality, with few parameters.
WHERE THEY REMAIN STRONG
Single series, regular intervals, limited history.
WHAT MACHINE LEARNING APPROACHES REQUIRE
Turning the series into a supervised problem with lagged features.
WHAT LAG FEATURES ARE
Previous values, at chosen offsets.
WHAT ELSE TO INCLUDE
Rolling aggregates Calendar features Known future values, such as holidays and planned events
WHAT TO BE CAREFUL WITH
Computing rolling features across the split boundary Using future information in any aggregate
WHAT VALIDATION MUST BE
Forward-looking: training on earlier periods, testing on later.
WHAT MULTI-STEP FORECASTING REQUIRES
Deciding whether to predict each step separately or recursively.
WHAT RECURSIVE PREDICTION ACCUMULATES
Error, growing with horizon.
WHAT TO ALWAYS COMPARE AGAINST
The naive forecast: the last value, or the value a season ago.
