An Ensemble learning of volatility spillovers among emerging currency markets

Abstract

The accurate prediction of volatility spillovers enables policy makers and traders to create more efficient investment and hedging strategies. There is an overall lack of research that focuses on the shocks that the Kenyan currency receives from other currency markets. Therefore, this paper estimates the degree of volatility spillovers from select exchange rates to the KES/USD using a GARCH-LSTM (General Auto- Regressive Conditional Heteroskedasticity-Long Short Term Memory) model. The study focuses on daily exchange rate data from 2017 to 2024. This timeline captures the recent significant events that have affected emerging economies: BREXIT, the COVID pandemic, the Russia -Ukraine war, and the increase of Federal Reserve interest rates. The selected currencies are the BRICS currencies: the South African Rand, the Brazilian Real, the Chinese Yuan, the Russian Ruble, the Indian Rupee, and the two dominant Western currencies: the Euro, and the British Pound. Daily volatilities are estimated by the GARCH(General Auto-Regressive Conditional Heteroskedasticity model and then forwarded to the LSTM (Long Short Term Memory) model which captures dependencies in the non-linear and complex volatility patterns. The results of the GARCH-LSTM ensemble model show that for the full period (2017-2024), the KES/USD incurred the highest degree of volatility spillover from the INR/USD at 1.3131 and RUB/USD at 1.1609. While the ZAR/USD transferred the least volatility at 0.0098. CNY/USD followed closely at 0.0137. Therefore, the study recommends the implementation of hedging strategies, especially during trade with the Indian and Russian markets as they demonstrate the highest contagion risk to the KES/USD exchange rate. In addition, among the forex markets studied, the South African and Chinese markets are safer to trade as they pose the least risk of contagion to the KES / USD exchange rate. Besides, research centers should focus more on the application of ensemble learning systems in macroeconomic issues. This is because they give more layered information about financial time series than traditional GARCH models. For example, in this paper, the GARCH-LSTM model outputs a close estimate of the degree of shock transmission, while a GARCH model could only give us the correlation between the volatilities.