Some Solutions to the A Priori Knowledge Issue in the Short-Term Electricity Price Forecasting

The recently proposed in the energy literature approach to short-term electricity price forecasting, based on explicit accounting for the long-term price dynamic (i.e. its independent modeling), has demonstrated its efficiency in gaining forecast accuracy. But the practical implementation of this approach has certain impediments, because the "true" trend-cyclical component is unknown in most cases, while the choice of the method and the degree of smoothing of a time-series to estimate this component can only be made by experts on an a priori basis. If such choice is made incorrectly, this eliminates the mentioned advantage of this approach, and may lead to accuracy loss as compared even to less sophisticated forecasting methods. In the current research we call it the a priori knowledge issue and study some possible solutions of this problem. We show that the adaptive methods of trend estimation, which are based on different algorithms of the empirical mode decomposition, while not requiring any a priori setups, still, do not solve the studied issue. In turn, forecast combining conducted for individual models (based on different methods and degrees of time-series smoothing) allows not only to mitigate the need of making a priori choices, but also has lower forecast error and, thus, performs better than individual models. We also propose a new approach to forecast combining (based on p-values of a model confidence set (MCS)) and show that it outperforms a number of well-established classic forecast averaging schemes (simple averaging, constrained OLS, inverted root mean square errors). Finally, our research indicates that making an MCS-based trimming of the pool of models before averaging of their forecasts does not lead to lower prediction errors relative to their untrimmed averaging. Hence, conducting such trimming does not provide any extra advantages in solving the a priori knowledge issue