JSE Top 40 falls as rand weakens across all seven quantiles

A weaker rand tends to accompany a fall in the JSE Top 40 within the same month. Mkandla and Chipepa find that relationship at all seven estimated quantiles, from the 5th to the 95th. Earlier quantile-regression studies, largely covering Asian markets, already find a negative exchange-rate effect on equities that grows stronger in the tails. This paper contributes a joint assessment of return quantiles, persistence, downside risk and tail dependence across the same 215 months of South African data. The USD/ZAR relationship measures same-month exposure. The paper's own lagged model finds no one-month lead.

A monthly panel, examined four ways

The study covers the South African stock market through the JSE Top 40 of the Johannesburg Stock Exchange. The paper describes it as Africa's largest and most liquid exchange. Its monthly log returns, expressed in percent, cover August 2007 to July 2025 for the JSE Top 40, USD/ZAR, Brent and gold. Each macro variable has an economic channel into equities. The rand affects import costs and foreign flows; oil affects South Africa's import bill and serves as a proxy for global demand. Gold enters because South Africa is a large producer. The authors expect these exposures to change with market conditions, so quantile methods let them examine the tails alongside the median. The result is a risk map for South African equity exposure, without an attached trading rule.

Their four findings are straightforward. USD/ZAR has a negative effect at all seven quantiles, while return persistence is weak. Past VaR and past losses drive downside risk. Only the JSE-Brent pair shows copula tail dependence.

The analysis starts with a quantile regression of JSE returns on contemporaneous USD/ZAR, Brent and gold returns. It covers seven quantiles and uses 1,000 bootstrap replications for standard errors. Next comes a quantile autoregression with exogenous variables (QARX). This relates the current month's JSE return to the previous month's JSE, rand, oil and gold returns at the 5th, 50th and 95th quantiles.

The downside-risk model is an Asymmetric Slope CAViaR, which updates the 5% VaR using the previous month's VaR and separate slopes for positive and negative returns. Three variants introduce different combinations of lagged USD/ZAR, Brent and gold returns. Finally, the authors fit pairwise copulas to rank-transformed residuals after ARMA filtering. Gold alone receives a GARCH filter, following an ARCH-LM p of 0.0174. AIC determines the copula choice for each pair.

The USD/ZAR coefficient ranges from -0.490 at the median to -0.693 at the 95th quantile, with significance at 1% throughout. Brent's coefficient is positive and significant at four of seven quantiles; it is 0.246 at the 5th. Gold reaches significance only at the 10th quantile, with a coefficient of -0.264. In the VaR process, the lag coefficient is 0.4822. Returns of either sign push VaR deeper: a point of loss lowers it by 1.0775, compared with 0.3040 for a point of gain. Losses therefore move it about 3.5 times as much. Kendall's tau is -0.3422 for USD/ZAR, 0.1746 for Brent and 0.0383 for gold.

What does the USD/ZAR loading tell a trader?

The headline regression uses contemporaneous regressors throughout. Its lower-tail USD/ZAR coefficient of -0.549 describes the book's exposure during a rand selloff. It cannot establish the direction of causation: foreign equity flows and the currency move together.

The QARX addresses forecasting directly. For lagged USD/ZAR, the p-values across the three quantiles are 0.619, 0.583 and 0.170. The authors acknowledge that the quantile-regression relationships need not translate into significant one-month-ahead relationships. Yet the abstract says exchange rate movements exert a consistently negative effect on stock returns. That wording implies more causality than the evidence supports.

The QARX also contains a mislabelled result. JSE itself supplies the only significant lag, at the 95th quantile: -0.406, with p = 0.001. The paper calls this persistence. A negative upper-tail coefficient indicates reversal after strong months.

Lagged macro returns barely improve JSE VaR

The improvement is confined to the in-sample comparison. Baseline CAViaR records 11 violations, a rate of 0.0512, and average quantile loss of 0.5954. Adding the full macro set produces 10 violations at 0.0465 and loss of 0.5838. Both pass Kupiec (p = 0.9379 and 0.8124) and conditional coverage (0.2705 and 0.1882). The authors describe the loss differences as "relatively small" and conclude that macro variables supply limited additional information. We agree.

The comparison has limited reach. We did not find a holdout period. Estimation and backtesting appear to use the same 215 months, where three extra parameters should reduce loss almost automatically. We found no standard errors for the CAViaR parameters either. And Nelder-Mead offers no guarantee against local optima.

At 0.0700, the macro version's independence p is nearer rejection than the baseline's 0.1063. This hints at greater clustering of its violations. The abstract also attributes downside risk to past shocks "rather than contemporaneous macroeconomic variables", although the extended CAViaR uses lagged covariates.

Eleven months underpin the 5% tail estimates

The authors identify the binding constraint: roughly 11 observations occupy the 5% lower tail. Every claim of tail evidence depends on that count.

For JSE and Brent, the abstract describes tail dependence as "moderate" and "symmetric". The estimate is 0.1157 in either tail, and the bootstrap interval is [0.0000, 0.2366]. It includes zero. The authors warn about an interval near zero for gold, yet give Brent no equivalent warning despite its interval also including zero. Nor does symmetry constitute a finding: the Student-t copula imposes equal tails by construction.

The rand result runs into the corresponding problem. The selected Gaussian copula forces tail dependence to zero. Its AIC is -70.32, against -68.32 for the Student-t, a difference of exactly 2.00 that equals the penalty for one extra parameter. The degrees of freedom purchased no likelihood improvement. The abstract's conclusion of no asymptotic tail dependence thus comes largely from the selected family's label. The data never got to vote.

The paper treats gold honestly. It reports a Clayton lower-tail estimate of 0.0006 and cautions against interpreting it as economic dependence. But all the gold AIC values are positive (Clayton 0.46, Gaussian 1.46). Had the candidate set included an independence copula, its AIC of zero would have won. For all four series, the Kolmogorov-Smirnov uniformity checks give 0.00463 with p = 1.000. Rank pseudo-observations are uniform by construction. Those checks establish nothing.

Stress scenarios can use USD/ZAR; JSE-Brent tail dependence remains unproven

A monitoring rule survives. For a long South African equity book, the same-month USD/ZAR loading of about -0.5 to -0.7 is stable enough for stress scenarios. Brent also belongs among those scenarios. Its tail-dependence interval of [0.0000, 0.2366], however, should determine no position size. Gold offers no evidence of equity protection, with Kendall's tau of 0.0383 and p = 0.4034. The authors reach that conclusion too.

We could not test this ourselves.

Our data contains no JSE Top 40 series, USD/ZAR spot or Brent. Our history also begins around 2010, compared with the paper's 2007 start. Substituting US equities and WTI would measure a different dependence structure and answer a different question. A CAViaR comparison on months unseen by the model would change our assessment.