Abstract:
In several businesses, forecasting is needed to predict expenses, future revenue, and profit margin. As such, accurate forecasting is pivotal to the success of those businesses. Due to the effects of different exogenous factors, such as economic fluctuations, and weather conditions, time series is susceptible to nonlinearity and complexity, making accurate predictions difficult. This work proposes a machine-learning-based time series forecasting technique to improve the precision and computation performance of an induced time series forecasting model. The proposed technique, a multi-population particle swarm optimization-based nonlinear time series predictive model, decomposes a predictive task into three sub-tasks: observation window optimization, predictive model induction task, and forecasting horizon prediction. Each sub-task is optimized by a particle swarm optimization sub-swarm in which the sub-swarms are executed in parallel. The proposed technique was experimentally evaluated on fifteen Electric load time series using root mean square error, mean absolute percentage error, and computation time as performance measures. The results obtained show that the proposed technique was effective to induce a forecasting model of improved predictive and computation performance to outperform the benchmark techniques on all datasets. Also, the proposed algorithm was competitive with state-of-the-art techniques. The future direction of this work will consider an empirical analysis of the search and solution spaces of the proposed technique and perform a fitness landscape analysis.
