A hybrid framework of deep learning and traditional time series models for exchange rate prediction

The shortcoming of standalone deep learning and time series model necessitates the formulation of the hybrid of the two techniques. This study provides a framework for the hybrid of a traditional time series model and a deep learning model and implemented to fit exchange rate time series data. The hybrid of SARIMA-LSTM was formulated and compared with three deep learning models such as the Gated recurrent units (GRUs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) and two traditional time series models namely the Autoregressive Integrated Moving Average (ARIMA) and the Seasonal ARIMA (SARIMA). First, the models were fitted to simulated time series and secondly to exchange rate of South African ZAR relative to United States Dollars (USD), the European Euro (EUR), and the Chinese Yuan (CNY). The study showed that the SARIMA-LSTM is more robust and outperforms other models based on the forecast accuracy metrics, therefore recommend for modelling financial time series data. The study showed that more losses are expected than gains for the ZAR/USD, and the same cannot be said for the ZAR/EUR and ZAR/CNY. The South African government should take measures to mitigate against losses of ZAR /USD. This study aligns with the sustainable development goals (SDGs 8) on growth and stable economy where trade and investments thrive.