Fig. 2

The SMAP procedure. The SMAP procedure is illustrated for the target vector at t = 6 and for e = 3. The other vectors in E are now regarded as neighbors of this target. Each neighbor has an associated future value of x at time t + 1. The set of neighbors and their associated values is used to construct a local linear model with 3 parameters (c₁, c₂, c₃). These parameters only apply to a neighborhood of the target vector. When constructing this linear model, each neighbor is assigned a weight depending upon its distance to the target in embedding E, using a Gaussian function w. The width of w depends upon parameter θ. The target’s future value (x₇) is computed using the linear model with scalar values from the target vector. Thus, a separate model and prediction are generated for each vector in E. The prediction performance is computed by comparing the predicted values with the observed values