DSAS Colloquium - Dr. Christina Erlwein-Sayer (Germany)
1170
Dr. Christina Erlwein-Sayer (Hochschule für Technikund Wirtschaft Berlin, Germany)
Invited by Dr. Rogemar Mamon
Machine learning (ML) models such as classification trees and artificial neural networks are widely applied to understand and predict patterns in financial markets. ML models are powerful tools, but must be handled with care to produce interpretable and reliable results. As data-driven decision processes made their way into risk management, estimation and model identification issues have arisen. These might result from data issues and be reinforced by model choices that are inappropriate given the prevailing market conditions.
To capture varying market conditions, we develop a model combining a long-short-term memory neural network (LSTM) with a Hidden Markov model (HMM). We detect hidden states in observed financial time series, filter states and incorporate regime information into LSTM. The forecast of corporate credit spreads in changing market regimes is developed. Furthermore, the accuracy of neural network and HMM-LSTM predictions applied to corporate credit spreads of three European countries is analysed.
We also give an outlook on an HMM-LSTM ensemble model, where regime-switching information works as a gating function to activate a neural network. We apply this to time series forecasting in electricity markets.
https://www.htw-berlin.de/hochschule/personen/person/?eid=12150