Modelling and forecastng sustainable development tme series via single and combined approaches
Stochastic processes. Single predictor. Combined approach. Optimisation. Sustainable development.
The last two decades of the twentieth century have marked the world with
changes in the Social, Economic, and Environmental fields. As a result, one of the most
highlighted discussions has been the quest for sustainable development. In fact, this new way
of seeing the world and the humans points out the close and complex relationship among the
Social, Environmental, and Economic dimensions, the so-called triple bottom line for
sustainability. Therefore, the development of methodologies focused on monitoring,
forecasting, and controlling variables underlying these dimensions has been encouraged. In
this way, the issue of time series modelling and forecasting becomes paramount. This
dissertation aims to present a framework for dealing with sustainable development time
series. Specifically, the problem of offering parsimonious predictors is taken into account.
Three single approaches (Autoregressive integrated moving average -ARIMA, Exponential
Smoothing -ETS, and Artificial Neural Network -ANN) and four combined approaches
(Simple average -SA, Simple median -SM, Minimal Variance -MV, and artificial neural
networks -cANN) are considered. An optimisation process based on Simulated annealing
(SA) and Bayesian information criterion (BIC) was developed in order to define the
parsimonious ANN and cANN structures. To test the proposed approach, twelve time series
regarding sustainable development variables, four to each dimension of the triple bottom line,
are investigated. The Brazilian Amazonian deforestation, S&P500 Index and the incidence of
mumps in New York city are some examples. As a result, it is claimed that time series
involving environmental, economic, and social issues have different characteristics and thus
might demand different formalisms for modelling and forecasting exercises. Further, single
models can not be the best methods, at a first glance. In turn, the proposed approach to
promote parsimonious ANN and cANN, via SA by the minimisation of BIC, have showed
good performance for both single and combined modelling.