Exchange rate volatility is said to exemplify the economic health of a country. Exchange rate break points (known as structural breaks) have a momentous impact on the macroeconomy of a country. Nonetheless, this country study makes use of both unsupervised and supervised machine learning algorithms to classify structural changes as regime shifts in real exchange rates in South Africa. Weekly data for the period January 2003-June 2020 are used. To these data we apply both non-linear principal component analysis and Markov-switching generalized autoregressive conditional heteroscedasticity. The former approach is used to reduce the dimensionality of the data using an orthogonal linear transformation by preserving the statistical variance of the data, with the proviso that a new trait is non-linearly independent, and it identifies the number of regime switches that are to be used in the Markov-switching model. The latter is used to partition the variance in each regime by allowing an estimation of multiple break transitions. The transition breakpoints estimates derived from this machine learning approach produce results that are comparable to other methods on similar system sizes. Application of these methods shows that the machine learning approach can also be employed to identify structural changes as a regimeswitching process. During times of financial crisis, the growing concern over exchange rate volatility, including its adverse effects on employment and growth, broadens the debates on exchange rate policies. Our results should help the South African monetary policy committee to anticipate when exchange rates will pick up and be prepared for the effects of periods of high exchange rates.
machine learning Markov-switching process non-linear principal component SouthAfrica