We cannot give a clear answer!

Study the result by
plot 'whatisit.dat.ll' w li, 'whatisit.dat.ll' u 1:3 w li
The the first curve shows the average forcast error as a fuinction of the neighbourhood size, indicating that there is enhanced predictability by the local model, suggesting nonlinearity. But, as the second curve shows, only a few percent of all points can be predicted this way, i.e., most of the points do not have sufficiently many neighbours for these small neighbourhood sizes. When we require predictability for all points with an identical neigbourhood size, we see that the global linear AR-model works best.

The local linear predictor will nevertheless will be better than the global model, since it minimizes the neighbourhood size for every data point.

Notice that a clear application such as prediction can be optimized by an optimization of the embedding parameters. It is, however, time consuming, since it has to be done by hand.

Since the nature of the data set is yet unclear, we suggest to use a surrogate data test.