AQAI QuantAI research lab for systematic strategies

Automated analysis

This analysis was drafted by our research engine and has not been checked by a human editor. It may contain errors. It separates the paper’s own results from our tests, and any figures called ours come from our own backtest.

Our automated analysisOur backtest

Contemporaneous Heat Maps Do Not Size a Clean-Energy Book

Nawaz and co-authors find useful conditional signs, but levels regressions and a two-sided filter block implementation.

2026-09-08 · 7 min read · US-listed clean-energy and renewable-energy ETFs

Reviewing: Do fossil energy prices and climate policy uncertainty affect clean energy markets? Insights for financial sustainable development · Ali Nawaz, Chi‐Wei Su, Ran Tao et al. · Read it on openalex

Our backtest of this idea

Our automated quick test, not the paper's

Regime-Aware Multi-Scale Clean-Energy ETF Tactical Allocation

Backtest period 2020-01-01 to 2024-07-01 · hypothetical, net of modelled costs

Why these figures are not the paper's (3)

Run on a different market than the paper

The paper studies global clean-energy equity markets/indices, whereas the implementable strategy would trade liquid US-listed clean-energy ETFs. This is a limited market substitution: the proposed mechanism is a price-and-policy-risk exposure affecting renewable-equity valuations and can plausibly apply to US-listed clean-energy funds, but the paper's reported relationships and coefficients do not transfer directly to the ETF universe.

The paper's own figures describe its universe and do not carry over to ours.

This is not a replication of the paper (3)

  • The paper's likely external climate-policy-uncertainty index is not available as a precomputed series. The implementation must use a custom news-derived climate-policy-uncertainty proxy from available text/embeddings, limiting the test period to approximately 2020 onward and testing a proxy rather than the paper's exact explanatory variable.
  • The paper analyzes monthly data from January 2008 through November 2024, while news-based policy-uncertainty inputs are only available from approximately 2020; the full reported sample cannot be reproduced.
  • Wavelet-based quantile-on-quantile regression can be implemented in code, but it is primarily an inferential dependence method rather than an immediately specified out-of-sample trading rule. Signal thresholds, lag structure, and portfolio mapping require independent design choices.

The figures below measure what we could run, not the paper's own method, so they are not evidence for or against its claim.

Our own audit found this run does not follow the paper faithfully (13)

  • Clean-energy dependent series and tradable ETF universe: ICLN adjusted month-end prices proxy for the paper's S&P Global Clean Energy Transition Index, while holdings are selected from an explicit set of US-listed clean-energy and energy-transition ETFs. (invalidates: The paper's reported CE mean, median, standard deviation, skewness and kurtosis; exact CE quantile locations; and the magnitudes of the paper's raw and horizon-specific QQR coefficients.)
  • Climate policy uncertainty series: The implementation uses a point-in-time multi-source news-frequency proxy rather than Gavriilidis's published climate policy uncertainty index. (invalidates: The paper's CPU descriptive statistics, CPU quantile assignments including [0.3,0.45], [0.5,0.7] and [0.55,0.80], and all reported CPU-to-CE QQR coefficient locations and magnitudes.)
  • Haar mapping to the paper's horizon intervals: The implementation maps Haar dyadic detail levels into the paper's non-dyadic 1-3, 6-9 and 12-18 month intervals using normalized period-band overlap weights. (invalidates: Exact replication of the paper's short-run, medium-run and long-run QQR coefficient surfaces and the horizon-specific regions reported in Figures 2 and 3.)
  • Authoritative CE proxy definition: ICLN adjusted month-end close, rather than an equal-weight aggregate of eligible ETFs, is the sole implementation proxy for the paper's S&P Global Clean Energy Transition Index. (invalidates: The paper's CE descriptive statistics, exact CE quantile assignments, and the raw and horizon-specific GPEI-to-CE and CPU-to-CE QQR coefficient surfaces.)

9 further finding(s) are described in the note.

These are our findings about our own implementation, not criticisms of the paper. Read the figures below as a description of what we ran.

Jan 2020Total 5.2%Jul 2024
Sharpe
0.07
Total Return
5.2%
Max Drawdown
-40.7%
CAGR
1.1%
Volatility
15.5%
Beta vs SPY
0.48
Trades
275

A trader cannot size a clean-energy position from these surfaces. Every reported relationship is contemporaneous, estimated on index levels and passed through a wavelet filter built from the full sample. The signs make economic sense, and the short-horizon reversal conditional on clean energy's own state deserves attention. Neither feature supplies an executable exposure.

Opposing forces

The economic case is compact. Expensive coal, gas and crude make renewables cheaper at the margin. Risk-averse investors with existing fossil cash-flow exposure may then rotate toward clean-energy equity for substitution and hedging. Nawaz, Su, Tao and Lobonț cast this as an extension of mean-variance optimisation. Expected clean-energy return equals a base return plus a loading on the change in fossil prices, less a penalty for fossil-price variance.

Policy uncertainty pushes the other way. Drawing on Pástor and Veronesi's (2013) treatment of political uncertainty, the authors argue that investors initially struggle to estimate the political and economic costs of new government policies. Their second hypothesis follows directly: greater climate policy uncertainty reduces clean-energy equity performance. Few listed assets are as exposed to policy as clean-energy equities.

The estimator has two stages and uses monthly data from January 2008 to November 2024, roughly 203 observations. Three series enter. The IMF Global Price of Energy Index spans coal, gas and crude, with mean 177.82 and standard deviation 60.03. The Gavriilidis (2021) climate policy uncertainty index records mean 145.33 and standard deviation 68.55. For the S&P Global Clean Energy Transition Index, the mean is 988.42, standard deviation 605.24, skewness 2.45 and kurtosis 9.94. Jarque-Bera rejects normality for all three, at 7.78, 32.09 and 603.59.

MODWT separates each series into short (1 to 3 month), medium (6 to 9 month) and long (12 to 18 month) bands. Within each band, the authors regress clean energy separately on the fossil index and the uncertainty index. They use Sim and Zhou's quantile-on-quantile estimator, a Gaussian kernel and bandwidth h = 0.05. The resulting surfaces cover a two-dimensional quantile grid, with the explanatory variable's quantile on one axis and clean energy's quantile on the other.

The useful part of the surfaces

The broad result offers little surprise: fossil prices are generally positive for clean energy, while uncertainty is generally negative. The conditioning carries more information.

At the 1 to 3 month band, clean energy's own position determines the sign of the fossil effect. Moderately rising energy prices, at fossil quantiles 0.3 to 0.4, hurt weak clean energy at quantiles 0 to 0.2. High energy prices, at 0.7 to 0.9, help strong clean energy at 0.65 to 0.8. One oil beta cannot capture that reversal. In the 6 to 9 month band, the positive region gathers around joint upper quantiles, 0.7 to 0.95 on both. At 12 to 18 months, it appears at fossil quantiles 0.3 to 0.5 and clean-energy quantiles 0.6 to 0.95.

The uncertainty result is patchier than the abstract suggests. The abstract describes climate policy uncertainty as consistently negatively correlated with clean energy equity performance. Yet the undecomposed data contain a strong positive association at joint upper quantiles, 0.7 to 1 on both. The medium band also has a positive patch, with uncertainty quantiles 0.6 to 0.8 against clean-energy quantiles 0.75 to 0.95. The authors then say that "the negative relation is more pronounced than the positive over medium term." Bouri, Iqbal and Klein document exactly that crisis rotation into green assets, and the paper cites it in the literature review. Across most of the grid, uncertainty hurts clean energy. Those two upper-quantile corners run in the other direction.

Levels, timing and inference

Equation (12) regresses the clean-energy level at time t on the explanatory level at time t. The specification includes no lag, control variable, unit-root test or cointegration test. With three trending and strongly non-normal index levels, the exercise describes contemporaneous co-movement in levels. The paper makes no forecasting claim. Its clean-energy series has kurtosis 9.94 and ranges from 405.85 to 3880.37.

The method also falls short of the framing. The abstract promises "an integrated examination of global fossil energy price dynamics and climate policy uncertainty", while the contributions paragraph says the authors "move beyond prior studies that addressed market or policy drivers separately". Estimation consists of two separate single-regressor models: clean energy on the fossil index, then clean energy on the uncertainty index. A joint specification never appears. Rates, market beta and macro conditions remain uncontrolled.

We also did not find standard errors, t-statistics or confidence intervals for the QQ surfaces. The findings come as quantile ranges and colour bands, including "the dark blue area indicates that the relationship's general short-term direction is unfavorable". Readers therefore cannot tell whether a region is genuine or a grid corner supported by only a handful of months. Roughly 203 monthly observations are spread across a two-dimensional quantile grid, with bandwidth fixed at 0.05. Tail cells are thin by construction. The paper states that choice clearly and supports it by citing the authors' earlier work rather than applying a data-driven rule.

Its check averages QQR coefficients across quantiles and compares them with plain quantile regression. The estimates are close. Averaging QQR coefficients back to the quantile-regression estimate is close to an arithmetic identity between the two.

There is another scope mismatch. Gavriilidis's index comes from US newspaper coverage, although the paper never mentions that limitation, while the equity index is global. The authors acknowledge other constraints. They write, "We rely on global-level data, which may obscure important regional or country-specific heterogeneities," and say the method "may still fall short in accommodating abrupt structural breaks or rapidly evolving market dynamics." They leave the contemporaneous levels specification, missing standard errors, two-sided filter and separate single-regressor models unaddressed.

The filter uses future observations

MODWT is a two-sided filter. The paper describes one decomposition over the complete sample, without a rolling or expanding scheme. Under that reading, the 12 to 18 month component at a given date partly incorporates observations arriving later.

That treatment is standard for describing historical scale-dependence. A trader, however, could not have known the plotted long-horizon quantile position at the time. Turning the medium- and long-band findings into signals requires rebuilding the decomposition on an expanding window. Doing so yields different, noisier band values. Even so, the paper's investor recommendations rest on these surfaces: use clean-energy equities as a partial hedge against fossil price risk, and overweight insulated names during high uncertainty.

Our adaptation used five ETFs

Three limitations come before the results. We cannot trade a global clean-energy index, so we substituted five US-listed ETFs (ICLN, TAN, QCLN, LIT, GRID), using ICLN as the clean-energy proxy. We could not obtain the Gavriilidis index as a live series. Our replacement classifies news text directly, limiting the sample to roughly 2020 onward and testing a proxy instead of the paper's variable. The paper also gives no trading rule, leaving us to choose the lag and portfolio mapping. This is an adaptation rather than a test of the authors' surfaces.

We rebalanced monthly from 2020-01-01 to 2024-07-01. The wavelet bands were rebuilt on an expanding window and aligned with the paper's 1-3, 6-9 and 12-18 month scales. QQR was estimated on an expanding window over the 0.05 to 0.95 grid with h = 0.05. Each estimate used data through the previous month end, with execution at the close of the first trading day afterward.

A regime layer translated the local quantile estimates into gross exposures of 0%, 50%, 75% or 100%. It imposed an explicit zero-exposure veto when the short-run fossil quantile fell in [0.3, 0.4] and the clean-energy quantile fell in [0, 0.2]. That rule directly encodes the paper's sign reversal. Positions across the five ETFs were sized by inverse volatility.

The result was poor.

Across those 54 months, total return was 5.18%, Sharpe 0.07, Sortino 0.11, Calmar 0.03, maximum drawdown -40.71% and annualised volatility 15.51%. A 0.07 Sharpe paired with a 40.71% drawdown reflects badly on our implementation. These figures should be read first as evidence about that implementation. The live window is short and one-sided, covering a clean-energy boom followed by an unwind that erased 40.71% from the peak. We chose the mapping from quantile estimates to gross exposure, and another mapping would alter the figures.

The paper supplies no strategy returns, so no published Sharpe exists for comparison. Our outcome does not contradict the authors' estimates. A four-state gross-exposure gate, run monthly for 54 months, cannot settle a conditional relationship estimated across roughly 203 months.

Evidence that would change the verdict

I would want a walk-forward test that rebuilds the wavelet bands every month and lags the uncertainty index. Clean-energy returns should sit on the left side instead of levels. The resulting exposure schedule should then face buy-and-hold ICLN and a plain volatility target.

If the quantile conditioning survives those tests and beats a trend filter, the short-horizon sign reversal becomes tradeable. Until then, the paper describes how three indices moved together, while its medium and long bands use information unavailable to the trader at the time.

Our backtest stops at 2024-07-01, and everything after that date is deliberately left untouched so the same strategy can be checked out of sample later.

How our backtest worked

The steps the code we ran actually executed, from its strategy card. Ours, not the paper's — it is one automated implementation of the idea, not the authors' own.

At each completed calendar month end:
  1. Use only observations available through that cutoff; require at least 18 monthly observations.
  2. Resample adjusted prices to month end. Use ICLN as the sole clean-energy regime proxy.
  3. Form the fossil signal as 0.5 × expanding z-score(CL monthly log return)
     + 0.5 × expanding z-score(NG monthly log return).
  4. Form monthly climate-policy-uncertainty frequencies from documents whose classifier
     probability is at least 0.5, then average available source-level frequencies.
  5. Recompute expanding Haar MODWT components and map them to short (1–3 month),
     medium (6–9 month), and long (12–18 month) bands.
  6. On the 0.05-to-0.95 quantile grid, estimate Gaussian-kernel local-linear QQR models
     of clean energy separately against fossil energy and CPU, using bandwidth h = 0.05.
     Also estimate matched conventional QR models and average QQR coefficients across
     explanatory-variable quantiles before the robustness comparison.
  7. Set fossil exposure:
       - 0.00 if short-run fossil quantile is [0.3, 0.4] and short-run CE quantile is [0, 0.2].
       - Otherwise 0.50 when the applicable local fossil estimate is negative.
       - Otherwise 1.00 when it is positive and both CL and NG have positive cumulative
         returns over at least two of the 3-, 6-, and 12-month horizons.
       - Otherwise 0.75.
       - Cap at 0.50 if CL or NG volatility is at or above its trailing 12-month 90th percentile.
  8. Set CPU exposure:
       - 0.50 when CPU is elevated for two completed months and the applicable local estimate is negative.
       - 1.00 when the applicable estimate is positive, including specified positive upper-quantile regions.
       - Otherwise 0.75.
  9. Set the regime gross limit to min(fossil exposure, CPU exposure).
 10. Rank eligible ETFs using the intended 3-, 6-, and 12-month relative-strength inputs;
     the specification does not define their combining transformation. Select up to five.
 11. Normalize inverse-volatility weights using 20- and 60-day realized volatility, apply
     a 12% annualized portfolio-volatility target, enforce a 20% per-ETF cap, and apply the
     regime gross limit. The specification leaves the exact cap/scaler ordering contradictory.
 12. Hold residual capital in cash and execute changes at the close of the first trading day
     after the cutoff; skip a trade when no genuine execution price is available.