An ARIMA-LSTM Correlation Coefficient Based Hybrid Model for Portfolio Management of Dhaka Stock Exchange

Abstract

This paper proposes the forecasting of correlation coefficients of the Dhaka Stock Exchange (DSE) market assets required for portfolio optimization using an ARIMA-LSTM hybrid model. DSE dataset contains a mix of linear and nonlinear data, where linearity means a direct relationship between the dependent and the independent variables and vice-versa for non-linear data. In relatable papers we came across, either linearity or non-linearity was handled within the dataset but none dealt with the combination of both. Our proposed model encompasses both linearity and non-linearity within the datasets of the DSE assets. This cannot be accomplished using other conventional statistical models. We have filtered the linear components in the datasets using the ARIMA model and passed the residuals obtained onto the LSTM model which deals with the non-linear components and random errors. We have compared the empirical results of this model with several other traditional statistical models used in portfolio management namely the Single Index model, Constant Correlation model, and Historical Model. We have also predicted the correlation coefficients using the ARIMA model to see how one of the models in our hybrid performs individually. The experimental results demonstrate that the hybrid model outperforms the other models in terms of accuracy and indicates that the ARIMA-LSTM hybrid model can be an effective way of predicting correlation coefficients required for portfolio optimization.