Clean energy indices take opposite signs on the same shocks in Erben Yavuz's paper, a distinction a trader could use. The paper's reading of extreme geopolitical risk offers less. What it measures is a short-horizon change in a newspaper article count.
What goes into the estimates?
The economic argument is familiar. Geopolitical risk threatens semiconductor supply chains and raises risk premia, both reasons to expect weaker artificial intelligence (AI) equity returns. Clean energy faces competing pressures: energy-security fears draw money toward renewables, while financing costs and exposure to critical minerals work against them. Grid software and data-centre power demand connect AI with clean energy.
Erben Yavuz examines daily observations from 18 December 2017 to 1 May 2026. The geopolitical risk measure is the Caldara-Iacoviello GPR index, a count of newspaper articles covering wars, terrorism and international tensions. Three indices represent AI: STOXX Global AI, Indxx Global Robotics & AI, and Nasdaq CTA AI & Robotics. The clean energy side uses FTSE Environmental Opportunities Renewable & Alternative Energy, NASDAQ Clean Edge Green Energy, and S&P Global Clean Energy. The resulting 15 pairs comprise 6 GPR-market relationships and 9 relationships between AI and clean energy indices.
The analysis uses three tools. Wavelet quantile-on-quantile regression (WQQR) uses a MODWT wavelet to separate each return series into frequency bands. It estimates the response of a target quantile to a driver quantile within each band, using a 19 by 19 grid. Quantile-on-quantile connectedness (QQCON) instead fits a quantile VAR(1) over rolling 200-day windows. Its NET measure subtracts the share of 20-step forecast variance received by a series from the share it sends, across an 11 by 11 grid.
FQCON separates those variance shares by frequency, into a 2-10 day band and a band beyond 10 days. The estimates come from two four-variable systems: GPR alongside the three AI indices, and GPR alongside the three clean energy indices. The headline findings are small GPR effects concentrated in the top and bottom GPR quantiles, strongest over 2-4 days. AI effects on clean energy are larger and more systematic.
A 0.95 GPR quantile measures a change
The paper log-differences every series, including GPR. Its description of that transformation is explicit: it "represents the daily rate of change in geopolitical risk". Thus, the 0.95 GPR quantile in QQCON selects the top 5% of daily log changes in the article index. WQQR selects the top 5% of the band-filtered change at each scale, from 2-4 days through 16-32 days. A noisy day following a quiet week can enter that tail. A week during a war, with article counts steady at a high level, cannot.
Yet the interpretation treats the quantiles as levels. The paper describes the WQQR effects on AI indices as "concentrated at extreme levels of geopolitical risk". In QQCON, AI indices become net transmitters around GPR's 0.3-0.5 quantiles. Those quantiles receive the description "normal geopolitical conditions".
Both statements concern changes.
There is another awkward finding in FQCON. GPR predominantly receives net shocks from the three AI indices in the short band, particularly around 2020-2021 and 2023. A newspaper count does not read equity prices. The likeliest interpretation when an exogenous news series "receives" shocks is statistical co-movement captured by the variance decomposition. The other NET maps warrant the same discount. We also found no description of the alignment between a seven-day news index and trading-day returns across time zones.
Opposite signs for FTSE and S&P clean energy
The strongest result is the split between indices. At the 2-4 day scale, the 0.95 GPR quantile has a negative effect on FTSE and a strongly positive effect at SP's lower and middle quantiles. AI returns have positive slopes for FTSE across all 361 quantile combinations and all four wavelet scales. NASDAQ clean energy has predominantly negative AI effects over 2-4 days, switching to predominantly positive effects over 8-16 and 16-32 days. SP's effects are small and negative over 2-4 days, then positive everywhere from 4-8 days onward.
QQCON gives a different view of that last relationship. The AI index transmits net shocks to SP in all 121 quantile cells, a pattern absent from every other clean energy index. WQQR nevertheless gives the same pair a negative short-horizon slope. These findings can coexist. NET measures which series' shocks explain the other's forecast errors; the direction of the response is a separate quantity. Reading "dominant net shock transmitters to the SP index" as a claim that SP follows AI upward would leave a desk with the wrong sign over 2-4 days.
The author supplies a direct caveat: "the heatmaps report point estimates of the slope coefficients". Only sign patterns are therefore interpreted. With 15 pairs and four scales, the exercise produces 21,660 coefficients, without standard errors and using a single bandwidth (h = 0.05).
The horizon ends at 32 trading days
D4, the longest wavelet band, spans 16-32 days. In trading terms, D1 describes a two-to-four-day swing; D4 is roughly a monthly rebalance. The paper describes D1 to D4 as "corresponding to short-term, medium-term, long-term, and very long-term dynamics". Its "long-term" category therefore covers 8-16 trading days.
FQCON puts everything beyond 10 days in its long band, with estimates drawn from 200-day windows. The author acknowledges that these bands differ from the wavelet scales: the short band includes D1, D2 and part of D3. None reaches the long-term strategic investment horizon discussed in the policy section.
The filtering also limits practical use. MODWT multiresolution components use two-sided filters with periodic boundaries, allowing later observations to enter each day's component. This is suitable for description. It rules out using the components as live signals.
Diversification remains untested
The abstract promises implications for "risk management, portfolio diversification". The body places a clear limit on the latter: "since net connectedness measures do not directly capture causal shock transmission or the gains obtainable from portfolio diversification, these implications should be evaluated within the relevant methodological limitations." Implementation, the policy section adds, "requires additional stress tests, portfolio performance analyses".
That qualification is honest and settles the diversification question. We found no portfolio, hedge, cost or out-of-sample test in the paper. Diversification remains a hypothesis presented in the abstract.
Index selection survives as the useful result. On both the 0.95 GPR jump and AI returns over 2-4 days, an SP clean energy sleeve takes the opposite short-horizon sign from FTSE. Check those signs before assuming the thematic label provides a hedge.
We could not replicate the study. Our data do not include the daily Caldara-Iacoviello GPR series, and we could not confirm complete histories for the six thematic indices. The sample extends to May 2026, beyond our backtest window. A walk-forward test showing that a trailing GPR-jump flag moves the realised FTSE-minus-SP spread after costs would change my view.