The paper's most useful number is 13.38.
Across eight commodity price indices, that is the total connectedness index at the median quantile from June 2012 to November 2023. Table 4 reports the complementary own-variance shares on its diagonal, ranging from 72.83 for gold to 97.14 for minor ETMs. The same connectedness measure reaches 54.83 at the 10% quantile and 52.03 at the 90% quantile. Commodity diversification holds in ordinary markets and weakens sharply in the states where investors need it most.
Inside the eight indices
The weekly set comprises base energy transition metals (ETM-B), precious ETMs (ETM-P), minor ETMs (ETM-M), gold, industrial metals, Newcastle coal, natural gas and crude oil. Following the IMF's commodity index documentation, constituent weights reflect shares in global imports. Each commodity begins with 2,994 daily observations, averaged into 599 weekly points. The authors deflate those observations using interpolated US CPI, normalise them to the 2016 average, calculate weekly log returns and standardise the results.
Estimation uses the Quantile Factor VAR of Ando, Greenwood-Nimmo and Shin. Latent common factors absorb residual cross-correlation, allowing quantile regression equation by equation. SIC selects lag order 1. The Ando and Bai criterion chooses the factor count subject to a cap of three, giving one factor at intermediate quantiles and three in both tails. The paper's comparison of QVAR(1,0.5) with QFVAR(1,0.5,1) shows the sum of absolute residual correlations falling from 4.06 to 3.08 after factors are added.
Connectedness is calculated from a generalised forecast error variance decomposition at a four-week horizon within the Diebold and Yilmaz framework. The static estimates are joined by 451 rolling windows of 150 weeks, bootstrap-after-bootstrap confidence bands and an event study covering 342 IEA policy announcements.
No portfolio appears, and no returns are reported.
We could not rebuild the network using our own data, and we ran no backtest against this paper. A direct comparison is unavailable because the paper gives no returns, no Sharpe and no transaction costs. Reconstructing its eight indices calls for Brent, Dubai, Newcastle coal and Dutch TTF, together with price assessments for molybdenum, silicon, manganese, chromium, rare earth carbonate and vanadium. US-listed commodity ETFs cannot stand in for those series. Fund structure and roll mechanics alter the return dynamics being modelled.
The disclosure is unusually complete
The paper discloses more than most connectedness studies. Index weights appear to two decimals, including copper at 47.17% of ETM-B, silver at 48.28% of ETM-P and iron ore at 94% of the industrial metals index, iron ore and tin. Its outlier rule is precise. Following McCracken and Ng, values outside median plus or minus ten interquartile ranges are replaced with zero. The series counts are seven for ETM-M, four for Coal, one for ETM-P and zero everywhere else. The authors also specify the lag, horizon, window length, window count, quantile grid, factor cap and rule for matching event dates to the nearest Friday.
Choices left to implementation
Data acquisition is the first. Table 1 quotes chromium, rare earths and lithium in CNY, with Dutch TTF and Brent Forties in EUR, before converting everything to USD. The chosen FX conversion path therefore enters the returns. Appendix A's Table A.1 identifies silicon, manganese, chromium and vanadium as CIF North West Europe price assessments instead of exchange-traded contracts. Rare earths are represented by carbonate REO 42-45% purity.
Then comes zeroing. Before the tail-quantile estimation, seven weekly returns for ETM-M are changed to zero. The authors explain that extreme observations can make estimated factors chase idiosyncratic variation instead of common movement, and they report the counts. The choice is defensible. It also excludes ETM-M's seven most extreme weeks from the fits at tau=0.1 and tau=0.9.
Across the full sample, ETM-M is the largest net receiver in both tails, with NET of -14.31 and -8.95. Before Covid, it records -1.95 at tau=0.1. After Covid, at tau=0.1, it changes to +8.82. Receiver status therefore belongs to the full sample and the pre-Covid period. I would run both versions of the zeroing rule.
Confidence bands require a third choice because their model differs from the one producing the point estimates. The note to Figure 7 says the factor number is "consistently set to 1 across all rolling windows and quantiles". Figure 6 takes another route for the rolling TCI, selecting factors dynamically through the criterion displayed in Figure 5. At the tails, the selected count exceeds one and can reach the cap of three.
The authors interpret the narrower tail bands as evidence that tail dynamics "are not statistical artifacts". Their interpretation rests on the one-factor bootstrap. They separately report forcing the factor count to one to avoid singularity in 2 windows at tau=0.1 and 22 at tau=0.9.
Has crude oil lost the centre?
The abstract says "crude oil remains influential" while "its dominance weakens post-Covid as ETMs, particularly base ETMs, gain centrality." The conclusion uses more restraint. Oil "dominates transmission mainly in the tails and can itself become a net receiver in extreme scenarios", which agrees with the full-sample tail tables. Its NET is -4.58 at tau=0.1 and -4.73 at tau=0.9.
The abstract foregrounds influence and structural reconfiguration. The conclusion settles on "suggesting a gradual move away from a purely fossil-centric network configuration". Those framings amount to different papers. The hedged version is the credible one.
At tau=0.9, the pre-Covid and post-Covid split supports the stated direction. Crude oil moves from +11.51 to -9.09, while ETM-B rises from +12.06 to +27.13.
Other entries in the table make the "particularly base ETMs" emphasis harder to sustain across both tails. At tau=0.1, ETM-B's net transmission declines from +13.34 before Covid to +1.37 afterward. ETM-P increases from +4.11 to +16.51, and gold moves from -14.15 to +12.61. Gold changes sign in both tails, going from -14.15 to +12.61 at tau=0.1 and from -14.31 to +7.63 at tau=0.9. It remains a net transmitter at the median, moving from +3.11 to +6.33. The conclusion itself says "ETMs and gold", so the broader result belongs to the authors. The base-metal claim is concentrated in the upper tail.
The post-Covid subsample contains about 195 weeks, compared with roughly 404 beforehand. The authors acknowledge the limitation: "since it is difficult to clearly separate the shock-impact period from the recovery period, the results for this timeframe should be interpreted with caution. They mainly provide a general view of how the economy and connectedness patterns have changed before and after the onset of Covid-19." A general view is reasonable from 195 weeks. The abstract's structural reconfiguration claim asks more of the sample.
Index aggregation also shapes the centrality result. Copper and aluminium account for 69.04% of ETM-B, leaving the transmitting node dominated by two LME contracts. The paper describes minor ETMs as predominantly shock receivers. That index is 92.30% silicon, manganese and chromium. Lithium contributes 2.31%, while rare earths contribute 3.85%.
What the 77 event-study hits represent
The event study records 77 significant hits: 44 at tau=0.1, 32 at tau=0.9 and 1 at the median. According to the table note, every event-quantile-horizon pair counts as a separate observation. The 44 therefore consists of 7 plus 17 plus 20 across the one, two and four week horizons. The 16/39/45 percentages are shares of those pairs. Among the 342 candidate events listed in Table D.1, 51 are significant at tau=0.1 under the looser 0.90 threshold, and several occur on the same date.
The screen applies three horizons and three quantiles to 342 candidates, producing roughly 3,000 combinations. We did not find a multiple-testing adjustment described in that section. Event clustering adds another problem. The tau=0.1 table contains four entries dated 15 September 2023: three Indonesian mining regulations and Malaysia's rare earth export ban plan, each flagged at the four-week horizon. Their individual effects cannot be separated.
State dependence is the result I would trade. The same eight nodes produce connectedness of 13.38 at the median and above 52 in both tails, with one factor selected at the median and three in the tails. Commodity diversification emerges as a normal-state property. A longer post sample, with copper and aluminium separated from ETM-B and the same pre/post split rerun, would change my view of the centrality reordering.