Assuming a good embedding and additive noise, the traditional approach to time-series embedding prediction has been to predict pointwise by (usually linear) regression of the k-nearest neighbors; no good mathematics has been previously developed to appropriately select the model (where to truncate Taylor's series) to balance the conflict between noise fluctuations of a small k, and large k data needs fitting many parameters of a high ordered model. We present a systematic approach to: (1) select the statistically significant neighborhood for a fixed (usually linear) model, (2) give an unbiased estimate of predicted mean response together with a statement of quality of the prediction in terms of confidence bands.
展开▼