Feature values, which may be multi-dimensional, collected over successive time slices, are efficiently processed for use, for example, in known adaptive learning functions and event detection. A Markov chain in a recursive function to calculate imputed values for data points by use of a "nearest neighbor" matrix. Only data for the time slices currently required to perform computations must be stored. Earlier data need not be retained. A data selector, referred to herein for convenience as a window driver, selects successive cells of appropriate adjacent values in one or more dimensions to comprise an estimation set. The window driver effectively indexes tables of data to efficiently deliver input data to the matrix. In one form, feature inputs are divided into subgroups for parallel, pipelined processing.