Title: Regime shifts in LSTM models for spot prices Authors:  Christina Erlwein-Sayer - University of Applied Sciences HTW Berlin (Germany) [presenting]
Tilman Sayer - Advanced Logic Analytics (United Kingdom)
Florian Schirra - Fraunhofer ITWM (Germany)
Stefanie Grimm - Fraunhofer Institute of Industrial Mathematics ITWM (Germany)
Abstract: Electricity spot prices show volatile periods and frequently occurring jumps over time. A prediction of day-ahead spot prices relies on suitable modelling paradigms to capture these changing dynamics. A regime-switching HMM is developed, which drives a decision-making process consisting of long-term memory models (LSTM) for specific time periods. Market regimes are adaptively filtered out from the data set and utilized to split the spot price series. This leads to the n-state HMM-LSTM, which is trained on split regime-specific daily prices. Weighted LSTM estimates lead to day-ahead predictions. Combining LSTM with filtered Markov chain probabilities increases the interpretability of predictions. Each activated LSTM is dependent on the filtered state of the underlying market. The model is applied to an extensive data set of German spot prices.