Time-series data prediction based on reconstruction of missing samples and selective ensembling of FIR neural networks

This paper considers the problem of time-series forecasting by a selective ensemble neural network when the input data are incomplete. Five fill-in methods, viz. cubic smoothing spline interpolation, EM (Expectation maximization), regularized EM, average EM, and average regularized EM, are simultaneously employed in a first step for reconstructing the missing values of time-series data. A set of complete data from each individual fill-in method is used to train a finite impulse response (FIR) neural network to predict the time series. The outputs from individual network are combined by a selective ensemble method in the second step. Experimental results show that the prediction made by the proposed method is more accurate than those predicted by neural networks without a fill-in process or by a single fill-in process.