Using machine learning to improve hedging of power prices in the Nordic market : a study of how predictions of the Nordic system price can be used for Norwegian hydropower producer’s hedging strategies

Abstract

The prices in the Nordic power market are characterized by high volatility. This creates a demand for securing future power prices. Large hydropower producers use a variety of instruments to predict price changes, and sign derivatives contracts to secure prices for parts of their production. In this thesis, we examined how the introduction of machine learning, in the form of power price predictions, can contribute to risk management for hydropower producers. More specifically, we focused on the following research question: How can predictions of the Nordic system price using machine learning methods enhance decision support for hydropower producers when trading medium-term power derivatives? To answer this question, we predicted the yearly, quarterly and monthly Nordic system price for 2018. Predictions of each price was made using the programming language R, with historical data from 2013 to 2018 retrieved through open sources and Datastream. We applied eight different machine learning methods, namely linear regression with backwards selection, ridge regression, lasso regression, partial least squares, regression trees, random forests, boosting and support vector regression. In addition, we generated forecasts using ARIMA and NNAR models. To replicate how the decision-making processes of traders would be in real life, the predicted prices by the three best-performing models on data prior to 2018 were compared to contract prices at Nasdaq Commodities. Based on the comparison we determined which futures contracts should be purchased. The answer to the research question is that machine learning models have great potential to enhance the decision support for hydropower producers when trading power derivatives. Compared to a strategy of securing all prices through futures contracts, using the predictions of the estimated models to decide whether to purchase the contracts led to the same or a higher gain. To mitigate the risk associated with the models and the market in general, the predictions made by the models should be used in combination with existing information and forecasts. The risks associated with the models should also be incorporated into the general risk management strategy. Keywords – Machine learning, risk management, futures contracts, hydropower producers