How does the concept of embedding dimensions apply to time series forecasting?
Instruction: Explain the theory of embedding dimensions and its relevance in constructing predictive models for time series.
Updated
Example Answer
A delay embedding turns a scalar series into vectors such as the current value, the value one delay earlier, and the value two delays earlier. The embedding dimension is the number of coordinates; the delay is their spacing. Those are separate choices.
Takens-style results motivate reconstructing certain smooth deterministic dynamics under specific assumptions. They do not guarantee that arbitrary noisy business data contains a recoverable attractor or that a larger lag vector improves forecasts. False-nearest-neighbor analysis can suggest a dimension, while delayed mutual information can help suggest a delay. Both can be misleading with limited or noisy data. I would also compare predictive performance across chronological validation windows and justify the representation against simpler lag features.
Make it your own
Specify the vector dimension and spacing, and explain whether you seek state reconstruction or better forecasting.
Why this works
It keeps the theoretical motivation separate from empirical model selection.
Follow-up to practice
Why does delayed mutual information not select the dimension by itself?