Name: Marco Otoya
Institution: Universidad de Costa Rica
This article shows a comparison between the traditional frequentist models, used for the analysis and forecasts of time series, and the statistical models of Bayesian approach. For the above, an application of the Bayesian forecast models to electricity sales for the Costa Rican Electricity Institute (ICE) to the residential sector of Costa Rica, considering the period 2012-2017.
Sales of electricity to the residential sector reflect the electricity demand of this sector towards the ICE, the study of its behavior through the analysis of time series is fundamental in terms of planning of demand, planning of future investments and in tariff aspects. The analysis of sales allows, for example, identify seasonal patterns and make appropriate forecasts for decision making.
The analysis of time series and, in particular, the development of forecasts is done through the application of a wide variety of techniques, from the simplest methods such as moving averages (Single, Double, Holt-Winters no seasonal, additive or multiplicative ), to more sophisticated techniques such as Box-Jenkins methods and neural network models (Anwar Rahman, et al 2012). Econometric forecasting models usually follow several principles, keep the model as simple as possible, use as much data as possible, use theory (not data) as a guide to select the causal variable (Ramanathan R. Et al. , 1997); this is an important difference with respect to Bayesian analysis.
In the case of estimates of electricity demand, different approaches have been identified in the literature, among which we can mention multiple regression models, the Box-Jenkins models and neural networks. Although Bayesian methods are widely known in statistical modeling, they have been used much less frequently to predict electricity demand (Anwar Rahman, et al 2012).