Short-term forecasting of forward prices in the Brazilian electricity market with a hybrid stochastic-neural network model

Since electricity is non-storable, it is considerably more volatile than other commodities, making price modeling and forecasting daunting. Nevertheless, predicting prices in the short term is essential for hedging. This task became considerably relevant in Brazil as, recently, the electricity market started to allow players to trade contracts freely. Because the Brazilian electricity market has some peculiarities, we propose a univariate model under the physical measure to describe the dynamics of forward contracts that combines stochastic differential equations (SDE) and artificial neural networks (NNs). The stochastic component incorporates mean reversion, jumps, and time-dependent parameters, whereas the NN component accounts for the dependence of prices on hydroelectric reservoir levels, through the affluent natural energy (ENA) values. NNs are also used to predict ENA and the SDE parameters, incorporating memory into the forecast. The model is designed to preserve parsimony as time series are relatively short. Moreover, to avoid overfitting, the model calibration, as well as the NNs setup and training are carefully done. The model provided accurate 30-day ahead predictions.