For sizing South African government bonds, monthly growth in the rand's real effective exchange rate matters at least as much per standard deviation as geopolitical risk. Moodley and Lawrence call the geopolitical index "the largest predictor of bond return volatility." Putting their coefficients on a common scale gives the exchange rate at least equal billing.

Four outside inputs to a volatility model

The paper's potential use on a desk is position sizing. Moodley and Lawrence fit monthly log returns for the All-Bond Index (ALBI) and four maturity buckets from IRSS, covering January 2005 to September 2025, 249 observations. Their model is EGARCH: last month's standardised shock feeds into the log of conditional variance. Four covariates enter that variance equation directly. The SARB supplies percentage growth rates for CPI, the long-term interest rate and the real effective exchange rate (REER). Caldara and Iacoviello's South African geopolitical risk index (GPRSA) enters in levels.

The economic argument starts with a small open economy dependent on foreign portfolio funding. A stronger rand and calmer politics should steady those flows and quieten bonds. If the covariates shift variance, a desk with a risk budget for SA government bonds could use them to adjust exposure before realised volatility catches up.

The authors select models in-sample. For each index, they fit GARCH, GJR-GARCH and EGARCH with normal, Student-t and GED errors, giving nine candidates. The Schwarz criterion chooses EGARCH throughout. It selects GED errors for the ALBI (4.456893), and Student-t for the 1-3yr (2.289773) and 3-7yr (3.765703) indices.

Only three series make it through the preliminary tests. Pre-estimation ARCH-LM tests reject no-heteroskedasticity for the 1-3yr (F 19.50, p 0.0000) and 3-7yr (F 30.87, p 0.0000) indices. For the ALBI, rejection comes only at 10% (F 2.82, p 0.0613). The authors exclude the 7-12yr (p 0.2704) and over-12yr (p 0.2131) buckets, following their test. Yet the over-12yr bucket has the sample's widest return dispersion: a standard deviation of 3.09 and a range of -13.53 to 12.27. The discussion nevertheless calls the geopolitical coefficients "associated with all bond maturities".

Across all three models, REER growth reduces variance and GPRSA increases it. Inflation and rate growth yield no evidence of an effect, with every p above 0.16. The inflation variable measures the percent change of an inflation rate; its minimum is -88.77 and its kurtosis is 90.4. That makes its null result difficult to interpret either way. The asymmetry term is positive and significant throughout, ranging from 0.504 to 0.573. The authors interpret this as rallies increasing volatility more than sell-offs, reversing the usual leverage pattern. The abstract describes an effect "on return", though the estimate concerns volatility.

How much weight should geopolitical risk carry?

The paper's p-values give the exchange-rate result firmer support. For REER growth, the coefficients are -0.0807 for the ALBI (p 0.0155), -0.0778 for 1-3yr (p 0.0204) and -0.1062 for 3-7yr (p 0.0021). The corresponding GPRSA estimates are 5.004 (p 0.0576), 4.963 (p 0.0795) and 5.480 (p 0.0453).

Two of the three geopolitical estimates clear only 10%. REER clears 5% throughout, reaching 1% in one case. These estimates follow a nine-way specification search for each index. We found no adjustment for testing three indices against four regressors. The discussion treats foreign exchange rates and geopolitical risk together as external drivers; geopolitics becomes the largest predictor only in the conclusion.

Causality is a further stretch. As written, the variance equations use all four regressors without lags. The month's REER change therefore appears alongside that month's bond variance. A capital-flow shock could weaken the rand and unsettle bonds simultaneously, generating a negative REER coefficient without either variable causing the other. The table estimates conditional associations.

A coefficient of five, a standard deviation of 0.037

Units explain why 5.0 looks so large beside -0.08. GPRSA ranges from 0 to 0.198 and has a standard deviation of 0.0373. REER growth, measured as a percentage change, has a standard deviation of 2.92.

Multiplying each coefficient by its regressor's sample standard deviation closes that apparent gap. A one-standard-deviation increase in GPRSA adds 0.187, 0.185 and 0.204 to conditional log variance for the ALBI, 1-3yr and 3-7yr indices. Those changes mean variance increases of 20.5%, 20.3% and 22.7%. A one-standard-deviation increase in REER growth subtracts 0.236, 0.227 and 0.310 from log variance. Variance falls by 21.0%, 20.3% and 26.7%, respectively. Volatility changes by roughly half those percentages. These calculations are ours, using the authors' table and holding the rest of the variance equation fixed.

REER growth has the larger effect in log-variance units across all three indices: 0.236 against 0.187, 0.227 against 0.185, and 0.310 against 0.204. Reverse the REER move and a one-standard-deviation fall would increase ALBI variance by about 26.6%, compared with 20.5% for a GPRSA increase.

Nothing here supports putting geopolitics first.

I would change my view if lagged GPRSA retained its size and produced a tighter p-value while lagged REER faded.

Where is the variance persistence?

The conclusion says bond shocks "persist over time". The GARCH term in the variance equation offers little support: 0.0407 for the ALBI (p 0.7335), -0.1369 for 3-7yr (p 0.3296), and -0.2904 for 1-3yr (p 0.0292). The last estimate is negative and significant. The discussion refers to "significant and positive GARCH coefficients for the three- to seven-year bond index", matching the mean-equation term of 0.817358. Alongside it is a nearly offsetting mean-equation 'ARCH' term of -0.752833. The ALBI has the same pattern, 0.7596 against -0.8835.

None of the three variance equations shows positive persistence. Conditional variance consequently consists mostly of the shock terms and the four regressors. The covariates take on forecasting work that GARCH dynamics would normally do.

Keep it on the watchlist

The diagnostics come out clean. Post-estimation ARCH-LM F-statistics are 0.0598, 0.0352 and 0.1288 (p 0.81, 0.85 and 0.72). Every Nyblom stability statistic falls below the 10% critical value of 0.353 except the 3-7yr mean intercept. Its statistic, 0.417, remains below the 5% value of 0.470. Clean residuals across 249 in-sample months establish fit. Forecasting value requires a separate test, and we did not find an out-of-sample evaluation or a portfolio exercise in the paper.

We lacked the South African series needed to test this ourselves. The authors state that their data are not publicly available. Using US bond ETFs would test a different market's mechanism.

A one-month lag and a rolling out-of-sample variance forecast against plain EGARCH would settle the question. REER growth would earn a place in a rand-bond risk model if it continued to cut forecast variance by about a fifth per standard deviation out of sample. Until then, this remains an in-sample description. Its conclusion puts geopolitics ahead of an exchange-rate effect that is at least as large and better measured.