Modified Artificial Bee Colony Algorithm for Tuning Simple LSTM for Multivariate Time-Series Forecasting

The family of deep artificial neural networks has lately acquired interest within the literature for prediction of data structured as time series. The subgroup of recurrent neural networks is better suited to problem prediction compared to other sorts deep neural networks. The most elementary variant of deep recurrent neural networks is basic recurrent neural networks based on the count of required parameters. Because of their basic structure, these neural networks can be used to address predicting difficulties if properly trained. Unfortunately, due to exploding or vanishing gradient concerns, training basic recurrent neural networks is challenging. This research contributes by presenting a novel training technique based on the artificial bee colony metaheuristics, which is given the task to execute training of the artificial neural networks. The proposed approach was tested on a multivariate prediction problem and has shown promising results. The experiments were executed on the Dow Jones stock exchange dataset by performing the forecast for one step ahead. This suggested technique outscored similar approaches that were analyzed in terms of R2, MSE, and RMSE metrics.