A commodity book diversified on median co-movement is sized for the wrong regime.
At the median, the total connectedness index is thirteen point four. In the tails, it is roughly sixty.
The first figure comes from the static QFVAR(1, 0.5, 1) fit at horizon 4. It estimates how much forecast-error variance in each of eight commodity indices originates with the other seven, using data from June 2012 to November 2023. The tail figure uses a different specification. In Figure 3, the same index is about 20 at the median and about 60 at both τ=0.1 and τ=0.9. Measured on that curve's basis, connectedness roughly triples between the middle and the tails.
The paper's pre- and post-Covid comparison carries the headline claim that transition metals gain centrality. It is also the less convincing half of the study.
The network they estimate
Bastianin, Casoli, Kočenda and Li use commodity price series from Refinitiv covering June 2012 to November 2023. Following the IMF Primary Commodity Price Index weighting scheme, they combine those series into eight weighted indices. Their starting universe is the IMF's 16 Energy Transition Metals (ETMs), defined by the IMF as metals used in clean energy technology.
The ETMs are grouped as base metals (copper, aluminium, nickel, zinc, molybdenum, lead, cobalt), precious metals (silver, platinum, palladium) and minor metals (silicon, manganese, chromium, rare earths, lithium, vanadium). The remaining indices cover gold, industrial metals (iron ore and tin), Newcastle coal, natural gas (Dutch TTF and Henry Hub) and crude oil (WTI, Brent, Dubai). All prices are converted into USD, deflated with interpolated US CPI and normalized to the 2016 average. The original 2,994 daily observations for each commodity are condensed into 599 weekly averages.
The model is a Quantile Factor VAR, using the estimator of Ando, Greenwood-Nimmo and Shin. Its autoregressive coefficients vary across quantiles of the conditional return distribution, while latent common factors absorb residual cross-correlation. SIC selects lag order 1. The Ando and Bai criterion chooses the factor count, subject to a cap of three. A generalized forecast error variance decomposition at horizon 4 then produces TO, FROM, NET and total connectedness measures for every quantile.
There is no portfolio exercise. The paper reports no returns, Sharpe or turnover. Connectedness describes risk transmission here, which is how the authors present it.
We ran no test of our own, including a substitute test. Standard statistical software can handle the quantile factor VAR and connectedness calculations without special hardware, but our available inputs are insufficient. Our universe has no price series for cobalt, molybdenum, silicon, manganese, chromium, rare earths, vanadium or Dubai crude, while three of the minor metals are denominated in CNY. Commodity ETFs would create a different network rather than reproduce this one.
The tails carry the result
The system appears almost dormant at the median. Total connectedness is 13.38. Gold sends the most spillovers, with TO of 34.76, and receives the most, with FROM of 27.17, leaving NET at +7.59. Crude oil records TO of 12.68 and FROM of 10.90, for NET of +1.78. Base and precious ETMs are the two largest net receivers, at -2.75 and -4.97.
Minor ETMs barely interact with the rest. Their own-variance share is 97.14%, while FROM is 2.86, the lowest figure in the table. Taken alone, the median table casts transition metals as passengers.
Figure 3 changes the picture. Static TCI rises from around 20 at the median to around 60 in both tails. The rolling-window estimates separate even more sharply, with roughly 70 to 90 in the tails and roughly 5 to 40 at the median. Those figures come from the bootstrap panels in Figure 7, where the factor count is fixed at 1 and the plotted line shows the bootstrap mean. Diversification suggested by the median table largely disappears in the states where investors would expect it to pay.
The model also needs more common structure at the extremes. Three latent factors are optimal in both tails, compared with one at the median. For τ=0.1 and τ=0.9, the Ando and Bai criterion is -1.98 at three factors. For τ=0.5, it is -1.11 at one factor. Adding factors reduces the sum of absolute residual correlations from 4.06 in the plain quantile VAR to 3.08. Because the QFVAR assumes a diagonal idiosyncratic covariance, its GFEVD rows sum to one without the ad hoc normalization commonly required in Diebold-Yilmaz work.
Gold changes direction as well. Its median NET of +7.59 makes it a transmitter, whereas in the upper tail (τ=0.9) it becomes a receiver. The authors interpret that reversal as classical safe-haven behaviour: during extreme commodity moves, whether positive or negative, gold absorbs shocks from elsewhere instead of setting them off. Anyone relying on gold as the stress leg of a commodity book should look closely at that switch.
How strong is the post-Covid case?
The abstract says: "While crude oil remains influential, its dominance weakens post-Covid as ETMs, particularly base ETMs, gain centrality." Even with that qualification, the statement depends on a sub-sample comparison. The authors set the break at the onset of Covid. Their pre-Covid period runs from 8 June 2012 to 28 February 2020, while post-Covid spans 6 March 2020 to 22 November 2023, giving the latter under four years of weekly observations. That interval contains the pandemic collapse, reopening, the 2021-22 inflation, the Russia-Ukraine war and faster transition policy.
Evidence for the claim extends beyond the two static networks. Section 4.4 estimates a 150-week rolling window over 451 windows, finding that upper-tail TCI "remained persistently elevated from 2020 through early 2022". The conclusion draws on those results. Still, the pre/post split is shown through network diagrams, and I did not find a formal test in the main text for differences in TCI or NET across the sub-periods.
The authors acknowledge the limits of this window. They say it combines continuing pandemic effects with the recovery, making the two hard to separate. "The results for this timeframe should be interpreted with caution." They continue: "They mainly provide a general view of how the economy and connectedness patterns have changed before and after the onset of Covid-19."
That framing is candid. It supports a broad comparison before and after Covid-19 less strongly than it supports a transition-driven reordering of commodity centrality. Yet the conclusion adopts the larger interpretation, referring to "a gradual move away from a purely fossil-centric network configuration".
Two modelling choices also affect that reading. In every rolling window and quantile, the factor count is selected again. Consequently, part of the movement in TCI may come from changes in the fitted factor structure rather than shifts in spillovers. The bootstrap intervals use another setup, forcing the factor count to 1 throughout instead of applying the main specification.
Tail treatment matters too. Observations outside the median plus or minus ten times the IQR are replaced with zero. This affects seven observations in minor ETMs, four in coal and one in precious ETMs, with none among the other five indices. Those are Twelve weeks out of 599, and they are tail weeks in a study centered on tail behaviour.
The minor ETM index raises a related question. Chromium, rare earths and lithium are quoted in CNY. Its 97.14% own-variance share could reflect thin or administered prices rather than true segmentation. This interpretation is mine, rather than the paper's. The cross-section contains eight indices, several dominated by a single commodity: iron ore makes up 94% of industrial metals, copper 47% of base ETMs and silver 48% of precious metals. Gold and coal each consist of one commodity.
Policy effects appear late
The event study provides the paper's clearest link to the energy transition. Using the bootstrap-after-bootstrap method of Greenwood-Nimmo, Kočenda and Nguyen, the authors test announcements recorded in the IEA Policies database. They identify 77 significant increases in spillovers. Of these, 44 occur at τ=0.1, 32 at τ=0.9 and exactly one at the median.
Resource governance and mining policy is the dominant category, accounting for 18 lower-tail events and 15 upper-tail events among 76 categorized cases. The table includes only events with probability above 0.94.
Of the 77 significant events, 76 occur in the tails and 1 at the median. Policy news changes connectedness during extreme states while leaving the median regime largely untouched.
As presented, the result is not directly tradeable. Connectedness is calculated through a 150-week rolling window, so each detected effect represents a change in the local spillover regime. The paper makes this interpretation explicit. Timing reinforces it. In the lower tail, only 7 events (16%) appear within one week. Another 17 (39%) emerge after two weeks, and 20 (45%) after four. The corresponding upper-tail counts are 9, 14 and 9. No multiple-testing correction is reported across the three quantiles and three horizons, a material issue when the main result is an event count.
The confidence intervals contain another odd feature. They are narrower in the tails than at the median, which the authors interpret as greater precision in estimates of tail connectedness. Extreme quantiles ordinarily provide fewer effective observations, suggesting the reverse. These intervals also come from the specification in which the factor count is fixed at 1.
The Covid break is imposed rather than estimated, and observations end in November 2023. A break test selecting its own date would make the centrality claim more persuasive to me. So would fitting the same network through 2025 while excluding the war-inflation window.
The tail estimate already clears that bar. Figure 3 places connectedness at about 20 near the median and about 60 in both tails, a threefold move. For a commodity book sized from mean co-movement, that multiple captures the error in the states that cost money.