Neural networks and ARMA-GARCH models for foreign exchange risk measurement and assessment

Abstract

Market turnover levels and liquidity changes across various territories significantly influence currency prices, leading to continuous fluctuations. Consequently, traders and investors constantly seek strategies to mitigate exchange rate risks. This study aimed to measure and assess foreign exchange risk utilizing Neural Networks and ARMA-GARCH models. Data on five leading currencies, covering the period from 6 January 2016 to 28 June 2024 were sourced from the National Bank of Rwanda. Specifically, the study employed the long-short-term memory (LSTM) model, a type of recurrent neural network, to evaluate the riskiness of asset currencies. The estimated volatilities were compared with those derived from traditional ARCH-GARCH models. Notably, the LSTM model yielded lower root mean square error values compared to the ARMA-GARCH models, demonstrating superior accuracy in forecasting currency volatilities. The findings indicate that EGP and KES are riskier than USD, EUR, and GBP.

Keywords

• ARMA-GARCH
• Recurrent neural network
• LSTM
• Exchange rate
• Risk measurement
• Econometrics