Machine-learning techniques and short-term combination forecasting of industrial production

• Autor(a) principal: Ferreira, Marcolino
• Data de Publicação: 2018
• Tipo de documento: Artigo de conferência
• Idioma: eng
• Título da fonte: Repositório Institucional do FGV (FGV Repositório Digital)
• Texto Completo: https://hdl.handle.net/10438/29319
• topic: Forecasting combination; Machine learning; Industrial production; Time series; Random forest; Combinação de previsões; Produção industrial; Séries temporais; Economia; Oferta e procura

The aim of this study was to develop short-term forecasts of the industrial production index in Brazil. Forecasts are made using five different methodologies: SARIMA, regressions, a structural, a dynamic factor models and decision trees. The random forest method had the best accuracy and was markedly superior to the other techniques. The univariate models had the worst performance during the period studied. Forecast combination was effective in reducing the one-step-ahead error. For the month-overmonth variation, for example, the RMSE, which varied between 1.27 and 7.57 for the individual models, was reduced to 0.85 for one of the combinations.