EcoSta 2022: Start Registration
View Submission - EcoSta2022
A0565
Title: LSTM in varying regimes: How to combine hidden Markov and machine learning models for financial risk management Authors:  Christina Erlwein-Sayer - University of Applied Sciences HTW Berlin (Germany) [presenting]
Stefanie Grimm - Fraunhofer Institute of Industrial Mathematics ITWM (Germany)
Abstract: Time series analysis and machine learning techniques are combined to forecast financial time series. In our hidden Markov model, hidden market regimes are detected and filtered out from observed financial time series. These states are subsequently incorporated into a long-short-term memory neural network (LSTM). Through this, we develop a model to forecast corporate credit spreads over changing market regimes within an LSTM; switching regimes are included as a feature to the neural network. This HMM-LSTM model is calibrated to corporate credit spreads from three European countries. The performance of the LSTM is analysed and compared to the accuracy of an LSTM without regime-switching information. Furthermore, we propose an HMM-LSTM mixture of experts' model, where regime-switching information acts as a gating function to activate a neural network. Applications of this approach to time series forecasting in electricity markets are shown. Our findings show that in most cases the LSTM performance is improved when regime information is added.