Estimation of Prediction Intervals for Time Series by Neural Network Constructed with Particle Swarm Optimization

Antonio Fabrício Guimarães de Sousa¹; Helaine Cristina Moraes Furtado¹; Anderson Alvarenga de Moura Meneses¹

¹ Universidade Federal do Oeste do Pará

Resumo

In the present work, Particle Swarm Optimization (PSO) is applied to determine the weight connections of an Artificial Neural Network (ANN) trained to estimate Prediction Intervals (PIs) of time series forecasting. Three case studies are implemented to evaluate the performance of PSO and PIs, in terms of accuracy (coverage probability) and efficacy (width). Results indicate that the utilized method reveals efficiency in constructing high quality PIs in a simpler and faster manner. According to them, the algorithm built narrow PIs, with a (mean±st.dev.)% for the coverage probability of (85.85±1.01)%, (92.77±1.50)% and (87.63±1.15)%, to the first, second and third. And it took 0.022 seconds on average for PI construction time for all cases.

Palavras-chave: Prediction Interval; Artificial Neural Network; Particle Swarm Optimization; Time Series Forecasting