Análise comparativa de modelos lineares para previsão do preço do açúcar

Abstract

The time series forecasting process is based on the assumption that future values are predicted from past observations and other inputs. It is one of the oldest predictive analysis techniques and has deep statistical bases and is widely used in organizational and research environments. A common approach to time series analysis involves evaluating various modeling techniques on a dataset and looking at how they explain past behavior. In this sense, a problem of relevance to the present day is to understand the movements in world sugar prices. Understanding the price dynamics between the domestic and international markets is of great importance for the strategic planning of the sugar and ethanol sector. Thus, the present study aims to analyze linear time series forecasting models to predict the price of sugar. For this, four databases were used, Brazil, United States, World and European Union. The data were applied using 7 models, namely, Autoregressive (AR), Moving Averages (MA), Autoregressive and Moving Averages (ARMA), Integrated Autoregressive Moving Averages (ARIMA), Simple Exponential Smoothing (SES), Exponential Smoothing with Trend (HOLT) and Exponential smoothing with trend and seasonality (HOLT WINTERS) and at the end, combinations of different models were performed based on the average of the predictors' outputs. For the present research, quantitative measurements of accuracy were used to evaluate the prediction models: Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percent Error (MAPE). It can be seen that for 1-step-ahead forecasts, the ARIMA family models estimated the sugar price more accurately. As for the 3, 6, and 12 steps forward predictions, the HOLT WINTERS family were the models that presented the best performance.