CDS risk premia forecasting with multi-featured deep RNNs: An application on BR[I]CS countries

Abstract

Using state-of-the-art recurrent neural network architectures, this study attempts to predict credit default swap risk premia for BR[I]CS countries as accurately as possible. In the time series setting, these recurrent neural networks are ELMAN, NARX, GRU, and LSTM RNNs, considering local and global features. The predictive power of each architecture is compared, and the results differ depending on the country. NARX RNN was the best predictor for Brazil and South Africa in various settings. Meanwhile, ELMAN RNN produces more accurate results in China, whereas Russia’s long short-term memory RNN achieves the best predictors among other countries’ RNNs.

Keywords

• Credit default swap premium
• Prediction
• Time series
• Recurrent neural networks
• Deep learning