Expediting Prediction Accuracy with Exploration and Incorporation of Virtual Data

Achieving high accuracy in time series forecasting (TSF) is challenging; particularly for series with inadequate training data or insignificant correlation among data. Modelling such time series with ANN based systems which are data driven and necessitate abundant training samples, leads to poor generalization ability thus, inferior accuracy. In such circumstances, the time series may be enriched with few more virtual data points (VDPs) explored from existing training samples keeping the corelation factor in mind. Incorporation of such VDPs intensify the volume of training data and helps the model in extracting the inherent nonlinearity efficiently. This article proposes an extreme learning with Rao algorithm based VDP (ELRA-VDP) generation and incorporation approach to expediate TSF accuracy of ANN model. Extreme learning method (ELM) considers random hidden node parameters and apply a generalized inverse function on hidden node outputs to determines the output node parameters. Inclusion of random hidden node parameters may reduce the performance of ELM therefore, we used Rao algorithm (RA) to optimize the initial parameters and construct an efficient learning scheme, i.e., ELRA. Given a training sample, the ELRA estimates the VDPs as well as forecasts the next data in an alternative manner. The parameter-free optimization mechanism of RA and lower complexity with faster learning ability of ELM makes ELRA computationally efficient. TSF problems are then conducted with and without VDPs. Simulation studies show that ELRA-VDP method yields better predictions than others with a lowest MAPE value of 0.1746 and quite beneficial for datasets having a lesser amount of data.