Forecasting the Regulating Price in the Finnish Energy Market using the Multi-Horizon Quantile Recurrent Neural Network

In recent years there has been a large increase in available data from the electric grid in Finland. The availability of both operational as well as financial data enables exploration of forecasting energy prices using deep learning techniques. As a result this thesis implements the Multi-Horizon Quantile Recurrent Neural Network (MQRNN) to forecast the regulating price in the Finnish energy market. The forecast is a rolling window three to eight hours into the future and contains several quantiles. The results suggest that while the central location of the distribution does not change much from the spot price the tails can be long, especially the right tail. Since the model is able to capture changes in the distribution there is indication that the market contains some structure. Finally, after dicussing the results and drawing conclusions some suggestions for future improvements are presented.Energy is something we use on a daily basis. Among other things we use it to charge our phones and turn on the lights. But what determines the price of energy? In the Nordics, energy is traded like a commodity by the market. That means the price is set according to supply and demand. And while the demand is easy to forecast, determining what will be produced is harder. Forecasting production accurately has become more difficult as the share of energy produced comes from renewable energy sources, such as wind power. The wind can be unpredictable. This has led to larger price movements on the regulating market for energy. Sometimes it can be very low and sometimes is can be very high. As such this thesis forecasts the energy price in the regulating market in Finland. The model is called the Multi-Horizon Quantile Recurrent Neural Network (MQRNN). The output of the model is quantiles for a period three to eight hours into the future. Quantiles are used to describe what the randomness looks like. To accomplish this, the model uses deep learning techniques which is a field within machine learning. Deep learning requires plenty of data on which the model can optimize itself on. In recent years the amount of data in the Finnish energy grid has increased. This is something the model leverages as inputs. The results indicate the model can detect relationships between the input variables. It is reflected in the output of the model where the quantiles can change from hour to hour. The results are of importance to companies which trade energy