The lower-tail result looks more like a flat allocation than market structure. At tau = 0.05, off-diagonal cells range from about 5.7 to 8.4 percent. Twelve of the thirteen diagonals are below 10%, with Casablanca highest at 10.71%. Each series accounts for roughly the same small share of every other series' forecast error variance, while explaining barely more of its own.
The paper's headline rests on those tables.
The system Korsah and Danso estimate
The data cover Seven African exchanges from Bloomberg (Ghana, Dar es Salaam, Egypt, Nairobi, Johannesburg, Casablanca), three precious metals (gold, silver, palladium) and three cryptocurrencies (Bitcoin, Ethereum, BNB) from investing.com. The stated daily sample runs from 10 July 2007 to 31 July 2023.
Log returns feed two models. One is a quantile VAR estimated at tau = 0.05, 0.50 and 0.95; the other is a TVP-VAR. Both produce Diebold-Yilmaz tables using the generalized decomposition of Koop et al. and Pesaran and Shin.
The reported total connectedness reaches 59.76% at the median, 90.48% in the lower tail and 89.97% in the upper tail. Across the full sample, the TVP-VAR averages 75%. Gold moves from a large net receiver in the median regime (-22.97) to a net transmitter in the lower tail (+3.80). Silver ranks as the largest net transmitter in both tails (+15.49 and +14.13).
Under normal conditions, Ghana leads the net transmitters at +26.52, followed by Dar es Salaam at +19.14. Johannesburg receives at -15.70, then transmits in the tails (+1.97 bearish, +6.86 bullish). The authors conclude that precious metals offer conditional safe-haven effectiveness because they frequently transmit volatility during extreme market conditions. Their title makes the stronger claim: When Safe Havens Fail.
The median table earns attention. Two small frontier exchanges transmit more than the JSE, reversing the usual intuition about the hub. Its own-variance shares also look economically plausible: 59.19% for Ghana, 21.76% for gold and 33.09% for Bitcoin.
Flat tails
The other two tables contain little comparable structure. Across all 156 off-diagonal cells in the bearish table, values range from about 5.68 to 8.42. Silver becomes the paper's dominant tail transmitter through column entries of roughly 7.2 to 8.4, compared with a field around 6.5 to 7.
The bullish table looks much the same. Its off-diagonals span about 5.85 to 8.57, while the diagonals range from 8.35 to 10.10. Those allocations produce total connectedness of 90.48% and 89.97%. The paper treats the near-equality as evidence that euphoria spreads as strongly as panic. A 13-variable VAR at tau = 0.05 may instead have very few effective observations in the conditioning region, leaving a flat decomposition.
The printed tables also fail to reconcile internally. The FROM column and TO row should yield the same total. In the median table, however, the FROM entries, each exactly 100 minus its diagonal, add to 797.20 against a printed TO total of 896.38. The FROM column averages 61.32, versus the reported TCI of 59.76. The bearish table repeats the discrepancy: 1181.84 against 1357.22. A table assembly error could explain this without implying an estimation error. Either way, the net rows should wait for a correction.
We did not find a reported lag length, forecast horizon H, or any confidence interval around the quantile TCIs. Those diagnostics would show whether the gap comes from markets or method.
Can a panel starting in July 2007 contain BNB?
BNB launched in 2017, and Ethereum in 2015. Yet the paper gives 10 July 2007 to 31 July 2023 as the sample for all thirteen series, citing data availability. We did not find the common estimation window, the number of observations retained after aligning thirteen exchange calendars, or the treatment of non-trading days in illiquid African markets.
Later figures carry the label 2016-2023, and the sub-period networks start in 2015. That suggests a working sample considerably shorter than the stated span. The summary table adds to the concern. It reports means of 7.737 for Ghana and 11.510 for Egypt, negative excess kurtosis for every series, and the authors' own statement that all series are non-stationary at levels. Those values resemble log price levels. Equation (1) defines a log return and describes it as monthly.
Safe havens require a different test
The paper cites the Baur and Lucey taxonomy. Under that definition, a safe haven preserves a weak or negative correlation with another asset during market turmoil. A positive net GFEVD share measures something else. It has no sign and does not isolate downside in African equities. It only shows that gold explains more of others' forecast error variance than they explain of gold's.
The authors also run the test that bears directly on the question. Their quantile-on-quantile surfaces place gold-equity slopes within roughly -4 to +4 across the joint distribution. They interpret that range as a relatively neutral, low-intensity influence on African equities, with gold acting as neither a strong destabiliser nor a strong stabiliser.
Their account is that gold's transmitting role remains real in aggregate GFEVD terms, though diffuse rather than concentrated in one corner of the joint distribution. In their view, this qualifies the headline without contradicting it. Gold-crypto surfaces behave differently, with excursions beyond plus or minus 10 concentrated at joint extremes.
"Diffuse but real" fits their surface. It cannot support a title about safe havens failing. A slope band from -4 to +4, with no concentration at equity lows, describes an asset that is mostly irrelevant to African equities in the tails.
The diagnostics tell a quieter story
The rolling 200-day TCI rises from about 28-33% before 2020 to a sustained 40-47% from 2021. That pattern is a level shift rather than a crisis spike. The semivariance difference is small and temporary. Bad-volatility connectedness exceeds good in 68% of windows, averaging 46.9% against 45.0%, and the gap of about 1.9 points reverses in early 2023.
Sub-period networks put the Russia-Ukraine window marginally below the pre-war window at every quantile, including 63.2 versus 66.4 at tau = 0.05. The authors report this result and adjust their account around it. The 1.9-point gap should concern them most because it directly measures the 90.48 versus 89.97 bearish/bullish contrast that they say it corroborates.
The authors make a related observation about the TVP-VAR. Its 75% average falls between the 59.76% median and the roughly 90% tails, which they say hides the regime-specific dynamics revealed by the QVAR. The abstract nevertheless leads with the averaged framing.
Why we left it alone
We could not test these results. Four of the seven equity series (Casablanca, Nairobi, Dar es Salaam, Ghana) lack a tradable index instrument in our coverage. Palladium is also absent from our futures roots, preventing reconstruction of the three-metal system as specified. US or global proxies would replace the frontier-integration mechanism at the center of the paper.
There is no portfolio to reproduce either. We found no hedge ratio, no weight, no Sharpe and no cost assumption anywhere in the paper, despite the diversification implications in its conclusion.
A re-estimation of the bearish and bullish tables could change our view. It would need a stated post-2017 common sample, reported lag and horizon, and bootstrap intervals for the net rows. Off-diagonals returning at 6 to 8 everywhere would leave the estimator as the likely source. If structure appears, the safe-haven claim becomes worth debating, using the quantile-on-quantile surfaces rather than the variance shares.