Prediction of time series using different types of forecasting methods enhanced with a meta-learning approach
Hugo Ernesto Betancourt Benavides, Carlos Andrés Brito González · 2020
The selection of a prediction model for a time series with irregular behaviour generally requires the help of a human expert; in many cases it is not possible to have this expert due to time or cost He/She represents. An intelligent machine can select the prediction model by analyzing the performance of different models, this strategy is called meta-learning, this work used Feed-Forward neural networks and decision trees to make the meta-learning models. To make this strategy possible it is necessary to select models based on different principles, this way we can provide variability to the list of methods. This work proposes the used of predictors based on neural network such as FFNN and LSTM, also, the inclusion of ARIMA a classical statistical method used in literature. It is important to maintain methods on the list with high performance, which provide information relevant to the learning process. This was possible with the used of Genetic Algorithms that improve neural network models. The results that were obtained show that the use of a meta-learning strategy allows the reduction of computational costs without reducing the performance of the prediction.