Estudo do desempenho da combinação de preditores baseados em cópulas e máquinas de vetor de suporte para séries temporais úteis ao desenvolvimento sustentável

Abstract

The 21th century has emphasized discussions regarding the sustainable development. In this way, it has been paramount the use of control indicators aimed to monitor, understand, diagnostic, and forecast variables of interest. Thus, time series modelling and forecasting exercises are important tools for decision making with respect to plans and actions regarding the triple bottom line of the sustainable development: environment, economy, and society. The statistical-computing modelling has become attractive by means of the development of technics for analyzing and predicting the behavior of phenomena changing through the time, considering the respective past observations. Among the alternatives, one must highlight the combination of predictors. Usually, they have been more accurate and efficient in statistical terms in comparison with single models. Therefore, approaches like copulas and support vector regression (SVR) seem useful, though classical formalisms, such as simple average, median, and minimal variance combinators, have presented good results. This work aims to compare the performance of the aforementioned combination strategies when forecasting time series useful for the sustainable development. Thus, twelve time series are taken into account, as well as four single modeling and four combination modeling formalisms. For all cases, optimized models are provided via optimization methods. The methodology is based on: i) to build single models for each time series, considering artificial neural nets, autoregressive and moving average, exponential smoothing, and SVR; ii) to build the combinators; and iii) to evaluate the performance of the models, according to a number of stablished quality metrics as well as measures proposed in the work. The results show that copula combinators are more promising for series useful for sustainable development when compared to SVR combinators. Still, even in the face of combined models, the simplest models showed good results. Parsimony and overfitting training may have compromised the performance of individual machine learning models, such as SVR and ANN, thus affecting the performance of the combiners.