Statistical Time Series Forecasting
Classical methods such as exponential smoothing and Croston's method that project a single series forward from its own level, trend, and seasonality.
No AIestablishedstatistical
Statistical time series forecasting is the default starting point whenever a decision needs a number for the next few periods and there is a history of that number to project.
It belongs before any machine learning model, both because it often performs as well and because it is the baseline that tells you whether a more complex model has earned its cost.
It stops being enough when the drivers of the series are external events that its own history cannot reveal, or when individual series are too short or too sparse to model on their own.
General strengths
- Fast enough to fit a separate model to every series in a very large population on ordinary hardware.
- Each forecast decomposes into level, trend, and seasonality that a planner can inspect and override.
- Hard to beat on short or noisy histories, which is why it is the benchmark any more complex model has to clear.
General limitations
- Learns only from the series' own past, so it cannot borrow strength from similar series or use external drivers without extension.
- Treats the recorded history as true, so anything that suppressed or inflated the record is projected forward as if it were real.
- Struggles with series that are mostly zeros, which need specialised variants rather than the standard methods.
Applied to business problems
Each of these is written for its own problem. The technique is shared, the solution is not.