New methods of structural break detection and an ensemble approach to analyze exchange rate volatility of Indian rupee during COVID-19

Journal: Journal of the Royal Statistical Society – Series A

Abstract: Exchange rates are arguably the most crucial macroeconomic variables to understand the dynamics of the currency market. This research is focused on understanding the behaviour of the exchange rates of Indian rupee against four other major currencies, especially when the economy is subjected to external shocks such as the ongoing pandemic. The authors aim to assess this, by proposing a novel technique of detecting if there are structural breaks, and by proposing an ensemble of multiple models which is better to analyse and forecast exchange rate volatility. In particular, they develop a methodology to detect structural breaks in multivariate time series data using the t-distributed stochastic neighbour embedding (t-SNE) technique and non-parametric spectral density estimates. By applying the proposed algorithm to the exchange rates of Indian rupee against four primary currencies, they establish that the coronavirus pandemic (COVID-19) has indeed caused a structural break in the volatility dynamics. They also identify that another break was observed during 2015. Such results are not observed if they use the benchmark techniques of structural break detection for every individual series. Next, to study the effect of the pandemic on the Indian currency market, they provide a compact and efficient way of combining three models, each with a specific objective, to explain and forecast the exchange rate volatility. They find that a forward-looking regime change makes a drop in persistence, while an exogenous shock like COVID-19 makes the market highly persistent. Their analysis shows that although all exchange rates are found to be exposed to common structural breaks, the degrees of impact vary across the four series. Finally, they develop an ensemble approach to combine predictions from multiple models in the context of volatility forecasting. Using model confidence set procedure, they show that the proposed approach improves the accuracy from benchmark models. Relevant economic explanations to their findings are provided as well.