Modeling and Forecasting the Ghanaian Cedi’s Exchange Rate Volatility Against Major Global and Regional Currencies: Evidence from the Bank of Ghana Interbank Market

Abstract

This study models and forecasts the daily volatility of the Ghanaian cedi (GHS) against eight major and regional trading-partner currencies (the United States dollar, the pound sterling, the euro, the Swiss franc, the Japanese yen, the Chinese yuan, the South African rand, and the Nigerian naira) using symmetric GARCH (1,1) and asymmetric EGARCH (1,1) specifications from the generalized autoregressive conditional heteroskedasticity family. Daily interbank reference rates were taken from the Bank of Ghana’s published historical interbank foreign-exchange records for 1 December 2025 to 8 September 2026 (189 dates, 186 usable daily log returns on a common sample). All eight-return series exhibited negative skewness and heavy tails (raw kurtosis between 6.6 and 22.8, against a normal benchmark of 3). Augmented Dickey–Fuller tests did not reject a unit root in the price levels at the 1% level for any currency and did not reject it at the 5% level for five of the eight; the euro, Swiss franc and yen levels rejected at 5% but not at 1%. Log returns were stationary for all eight currencies. The ARCH-LM test rejected the null of no ARCH effects at the 5% level for every series. In the GARCH (1,1) models the ARCH coefficient was significant at 5% for six of the eight currencies (not for the Swiss franc or the yen) and the GARCH coefficient was significant for none. Estimates lay at or on the boundary of the parameter space, with α + β equal to one for the dollar, pound, euro, yuan and naira and between 0.31 and 0.80 for the franc, yen and rand, so the fitted GARCH models behave like ARCH (1) processes. The EGARCH (1,1) specification had the lower Akaike Information Criterion (AIC) for all eight currencies, and the leverage parameter was negative and statistically significant for seven of the eight, indicating higher subsequent volatility after cedi-appreciation shocks than after cedi-depreciation shocks of equal size; because the eight series share the cedi leg and the largest shocks are common across pairs, this evidence rests on a small number of episodes. Residual diagnostics showed no significant remaining autocorrelation for the preferred models, with one borderline exception. Forecast accuracy was evaluated out of sample using one-step-ahead variance forecasts for the final 40 trading days, re-estimated on an expanding window and compared with historical-variance, exponentially weighted moving average and random-walk benchmarks. EGARCH (1,1) had the lowest QLIKE loss for seven of the eight currencies and significantly outperformed the historical-variance benchmark for the dollar, yuan, and naira, but historical variance or GARCH (1,1) had lower squared-error loss, and EGARCH (1,1) produced occasional extreme forecasts, so its forecasting advantage is not robust. The short estimation sample is an important limitation: the Bank of Ghana page offers a much longer history (at least for the US dollar), and extending the sample is the main priority for future work.