Explanatory variable selection for forecasting extreme prices in electricity markets: application to the Spanish case

Abstract

During the last years, non-dispatchable generation has increased its participation in electricity markets worldwide. As a result, the occurrence of extremely low and extremely high prices is becoming more and more common. In this new context, the risks to which market stakeholders are exposed have grown considerably. Therefore, trying to predict these extreme events is crucial for all market agents. For this purpose, it is essential first to delimit the bounds that define extreme prices and second to find the variables that contribute to explaining their occurrence. This article provides a satisfactory guideline for the selection of the variables and the models that better explain and predict extreme prices, including a systematic set of rules to define the thresholds for extreme prices in a given data set. Such a selection should be based both on statistical techniques and on in-depth knowledge of the electricity market. To this aim, some of the most relevant variable selection techniques are compared: AIC, BIC, Backward Methods and Sensibility Analysis. These techniques are applied to logistic regression and multilayer perceptrons models. The success of the methodology used, which is based on the extensive use of statistic diagnostic tests, is checked in the Spanish electricity market.During the last years, non-dispatchable generation has increased its participation in electricity markets worldwide. As a result, the occurrence of extremely low and extremely high prices is becoming more and more common. In this new context, the risks to which market stakeholders are exposed have grown considerably. Therefore, trying to predict these extreme events is crucial for all market agents. For this purpose, it is essential first to delimit the bounds that define extreme prices and second to find the variables that contribute to explaining their occurrence. This article provides a satisfactory guideline for the selection of the variables and the models that better explain and predict extreme prices, including a systematic set of rules to define the thresholds for extreme prices in a given data set. Such a selection should be based both on statistical techniques and on in-depth knowledge of the electricity market. To this aim, some of the most relevant variable selection techniques are compared: AIC, BIC, Backward Methods and Sensibility Analysis. These techniques are applied to logistic regression and multilayer perceptrons models. The success of the methodology used, which is based on the extensive use of statistic diagnostic tests, is checked in the Spanish electricity market.