Something old, something new : a hybrid approach with ARIMA and LSTM to increase portfolio stability

2020

In this thesis we seek to examine how modern forecasting approaches can improve estimations

of stock pair correlations, and derived from this, contribute to making portfolios more stable.

Volatility of financial markets have experienced increases due to the ongoing global pandemic.

This amplifies the issues that investors face when assessing the risk related to their

investments. We construct a hybrid model consisting of an ARIMA component to explain the

linear tendencies of correlation, and a Long Short-Term Memory component to explain the

non-linear tendencies. Our approach is populated by data from constituents of Oslo Stock

Exchange ranging a time span from 2006 through the third quarter of 2020. Our results indicate

that modern approaches to forecasting accrue stronger predictive performances than the

conventional methods. Across all test periods our proposed hybrid model achieves an RMSE

of 0.186 compared to an average benchmark RMSE of 0.237. However, the implications of

these findings are ambiguous as the increase in predictive performance cannot be said to

definitively outweigh the increase in cost of implementation. Our thesis contributes to the

existing literature by exhibiting the untapped potential of how modern approaches to

forecasting can improve accuracy of quantitative inputs for decision making.