An EUA future linked to 18 of 19 neighbours in a crash graph still leaves a trader guessing which hedge to buy. Maciejowski and Leonelli measure a real difference between average and extreme-day networks. Their trading advice asks more of that measurement than it can give.
Firms in the EU Emissions Trading System must hold allowances (EUAs) against their emissions. The ICE front-month EUA future therefore prices carbon in that market. The paper compares EUA's links to energy, equity, FX, volatility and bond markets on ordinary and extreme days. Its graphical models retain an edge when dependence survives conditioning on every other variable. A Gaussian model handles average days; a Hüsler-Reiss model handles joint extremes. Spike graphs contain 119 to 128 of 190 possible edges, and crash graphs contain 110 to 125, against 59 to 81 in the average graphs. EUA's degree rises from 3 in the phase average graphs to between 13 and 18 in the tail graphs. The authors find clustered crash contagion in Phase 3 (2013 to 2020) and its absence in Phase 4 (2021 to January 2025), when the Fit for 55 reforms took place.
What goes into the graphs?
The authors use 20 daily Bloomberg series spanning January 2013 to January 2025. They include front-month ICE EUA futures (MO1), Brent, API2 coal, gas, four equity indices, MSCI Europe Energy, two clean-energy indices, VIX, gold, five euro crosses and two bond indices. Each series receives an AR-GARCH filter before its standardized residuals enter the models.
For the Bayesian Gaussian graphical model, an edge survives when the 95% credible interval for its partial correlation excludes zero. The Hüsler-Reiss graphical model applies the same conditional-dependence idea to extremes: a missing edge means the remaining variables fully explain the pair's joint extremes. The authors fit it to the top 20% of rank-transformed residuals separately for spikes and crashes. An L1 penalty is selected through 10-fold cross-validation. They repeat the analysis for the full sample, Phase 3 (2013 to 2020) and Phase 4 (2021 to January 2025), producing nine graphs.
A desk using ordinary-day correlations to choose stress hedges could pick the wrong neighbours if those links change in the tails. The authors examine that possibility through centrality and exponential random graph models (ERGMs). The ERGMs relate the presence of an edge to sector dummies, same-sector ties and triangle closure (gwesp).
Carbon moves toward the centre
MO1 has degree 3 in both the Phase 3 and Phase 4 standard graphs, with eigenvector centrality of 0.12 and 0.24. Its degree reaches 16 in three phase tail networks and 13 in the Phase 4 crash network. Across the full sample, MO1 reaches 17 in the spike graph and 18 in the crash graph; its eigenvector centrality is 1.00 in both. The STOXX 600 goes the other way: degree 12 and eigenvector 0.95 in the Phase 3 standard graph, then degree 6 or 7 in every tail graph. EURUSD has eigenvector 1.00 in all three standard graphs, falling to 0.47 to 0.57 in the spike networks.
A hedge from a crowded graph?
The full-sample spike graph has 128 of 190 possible edges. The authors expect such density. Extreme financial co-movements tend to be more widespread than average ones, and cross-validation rewards held-out likelihood rather than sparsity. They argue that node centrality and sector peripherality carry information beyond the edge count. They have a point, though once two-thirds of all pairs are linked, high degree depends heavily on where the 62 absent edges fall.
The pairwise evidence warrants a different emphasis. Chi-hat measures the frequency with which one series reaches its extreme tail given that another does. In both tail directions across all three sample periods, MO1's chi-hat with each other variable is about 0.14 to 0.35. CAC and DAX reach 0.73 to 0.78. The authors themselves describe MO1's "uniformly weak pairwise extremal dependence" with every other variable. Conditional edges can coexist with modest unconditional co-exceedance, as they explain.
The hedging recommendation strains that distinction.
The paper advises compliance desks to prioritise coal and natural gas, citing EUA's strongest conditional tail dependence with them. We could not verify that ranking: we did not find numeric strengths for MO1's conditional tail edges in the tables, only edge widths in Figure 1. The reported pairwise figures give little support to gas. Coal leads MO1's chi-hat row at 0.28 to 0.35 across tail directions and sample periods. Gas futures (NG1) register 0.20 to 0.24. In all six configurations, the STOXX 600 exceeds gas at 0.21 to 0.27, despite being an equity index of the kind the paper advises hedgers to deprioritise. The paper reports no hedge ratio, cost or out-of-sample hedge.
The estimator description merits scrutiny before the code is reused. Read literally, the paper gives its penalty the opposite effect from the one its graph interpretation needs. Small Γ entries signify strong tail co-movement in its account. Its L1 penalty then shrinks small entries of Γ to zero and treats those zeros as missing edges. Eq. (4), as printed, would remove the pairs described as most strongly dependent. Engelke and Hitz (2020), cited by the paper, locate graph zeros in a precision-type matrix instead.
Phase 4 thins the crash graph too
The abstract says the phase transition "restructures the tail network without thinning it". It says average dependence contracts sharply while tail dependence persists. Elsewhere the paper reports losses in both extreme networks: 13 edges disappear from the crash graph and 8 from the spike graph.
From Phase 3 to Phase 4, the paper gives relative declines of 10% for the standard graph, 6% for spikes and 11% for crashes. Crash and average graphs shrink by nearly the same proportion: 13 of 123 edges versus 7 of 66. The paper's 11% and 10% differ only by rounding. The discussion's 27% decline uses the full sample (81 edges) and Phase 4 (59 edges), a different comparison. For persistence, the authors point to the spike-to-standard density ratio rising from 1.93 to 2.01. The corresponding crash ratio stays put. Crash density over standard density is 0.647/0.347 in Phase 3 and 0.579/0.311 in Phase 4, about 1.86 both times.
Wide errors around contagion
The crash-graph gwesp estimates carry the contagion argument: 36.35 (SE 15.90) for the full sample, 23.22 (SE 12.63, p=0.066) for Phase 3 and 0.47 (SE 1.49) for Phase 4. The paper calls the Phase 3 result marginally significant. If the fits are treated as independent, (23.22 − 0.47) divided by the combined SE of 12.63 and 1.49 is about 1.8. Edges terms of −59.74 and −37.17 in the full-sample and Phase 3 crash fits suggest near-degeneracy.
There are further tensions in the reported fits. The table puts Phase 3 crash homophily at 18.11 with SE 1908, while the text calls nodematch significant in every network without exception. The main text says AIC favours M2 in all nine networks; the appendix says six and describes the full-sample spike network as degenerate from M2 onward. Table 11 nonetheless gives M2 the lowest AIC in all nine rows, including 214.05 for the full-sample spike network against 237.88 for M1. Appendix prose also discusses M0 to M3 results absent from its tables.
The paper calls the loss of triadic closure "not an artefact of the smaller Phase 4 sample". Its stated support is that the change "aligns with the qualitative change in Phase 4 market dynamics documented by the financialization literature". We read that as an interpretation and did not find a test behind it. Later, the limitations say a hierarchical model would be especially helpful for Phase 4 because its shorter sample limits precision. The authors also acknowledge that Phase 4 combines the regime change with the COVID recovery, the energy crisis and the war.
Why we stopped at reading
EUA futures are outside the futures roots we can trade. We also lack coverage of the euro crosses and European series. US proxies would describe a different tail system, and the paper reports no trading results.
Numeric conditional edge strengths for MO1 would let us assess the proposed ranking. A hedge backtest on EUA crash days, comparing a coal overlay with an equity-index overlay, would tell us whether that ranking helps a desk.