Modelling and forecasting crude oil price volatility with climate policy uncertainty

Crude oil plays an essential role in the world economy and oil price volatility has important implications for macroeconomic and financial markets. The accurate prediction of oil volatility is crucial and valuable to crude oil futures investors, oil-related companies, and energy organizations in many fields. Extensive academic research has focused on improving crude oil volatility predictability, yet accurate forecasts remain an open question. Recently, scholars have examined links between crude oil volatility and uncertainty indicators such as economic policy uncertainty (EPU), geopolitical risk (GPR), and equity market uncertainty (EMV). This paper aims to improve predictability of crude oil volatility using the climate policy uncertainty (CPU) index (Gavriilidis, 2021) and its change indicators, and to relate CPU-based predictors to oil fundamentals to identify the source of predictive power. The paper proposes two transmission channels from climate uncertainty to crude oil markets: (1) an economic activity channel, whereby climate uncertainty affects activity, alters crude oil supply and demand, drives price changes, and affects market stability; and (2) a financial markets channel, as climate risks impact financial markets and, given the financialization of oil, transmit to crude oil volatility. To better capture climate uncertainty information, the study considers the first difference of CPU and an asymmetric increase measure, motivated by limited attention and asymmetric responses to positive innovations. Thus, beyond the original CPU, two change indicators are constructed: changes in CPU (CCPU) and increases in CPU (ICPU). The empirical analysis using these three indicators yields several findings: (i) statistically, CPU change indicators forecast crude oil volatility in-sample and out-of-sample, while the level CPU does not; (ii) economically, CPU change indicators deliver sizeable gains for crude oil futures investors; (iii) predictability concentrates in economic expansions and CPU contains climate risk not captured by other predictors; and (iv) CPU change indicators explain changes in crude oil consumption, the economic source of their predictive power. The paper contributes to uncertainty-based prediction of oil volatility by employing CPU and showing change indicators outperform other uncertainty measures, and to the climate risk–oil literature by documenting CPU’s predictive content for volatility. The paper proceeds with methodology, data, empirical results and discussions, extended analyses and robustness tests, and concludes.

The paper relates to two strands of literature. First, on uncertainty indicators for forecasting crude oil volatility, prior work studies GPR and EMV, finding they provide useful predictive information (e.g., Liu et al., 2019; Dutta et al., 2021; Wang et al., 2021; Song et al., 2022). Zhang et al. (2023b) show an aligned GEPU index is informative. Departing from these, the study employs the CPU index (Gavriilidis, 2021) and shows CPU change indicators, rather than the level CPU, predict crude oil volatility and outperform other uncertainty indexes. Second, on climate risks and crude oil markets, prior research documents that climate risks (physical and transition) affect energy market dynamics and investor behavior (e.g., Zhou et al., 2023; Griffin and Jaffe, 2022; Zhang et al., 2023a; Iin et al., 2024). Complementing this literature, the paper tests the in-sample and out-of-sample predictive power of CPU for crude oil volatility, offering a new perspective on climate risk–oil market linkages.

Volatility measure: Monthly realized volatility (RV) of the crude oil market is computed by summing squared daily returns within each month: RV_t = Σ_{i=1}^{N_t} r_{it}^2. Given non-Gaussianity, the logarithm of RV is used in regressions: LV_t = ln(RV_t). Predictive models: The benchmark is a monthly autoregressive (AR) model for future log RV: LV_{t+1} = β_0 + Σ_{i=1}^{k} β_i LV_{t+1-i} + ε_{t+1}, with lag order k chosen by AIC. To test a predictor X_t, the model is extended: LV_{t+1} = β_0 + Σ_{i=1}^{k} β_i LV_{t+1-i} + θ X_t + μ_{t+1}. Newey–West t-statistics test θ = 0. CPU predictors: The CPU index exhibits an upward trend and high persistence; innovations are proxied by first differences. Two change indicators are constructed: (i) CCPU_t = CPU_t − CPU_{t−1}; and (ii) ICPU_t = CCPU_t if CCPU_t > 0, and 0 otherwise, capturing asymmetric attention to increases. Thus, three predictors are considered: CPU, CCPU, and ICPU. Forecast evaluation: Out-of-sample performance is assessed via R^2_os (Campbell and Thompson, 2008): R^2_os = 1 − Σ(LV_{t+1} − LV^{CPU}{t+1})^2 / Σ(LV{t+1} − LV^{AR}{t+1})^2. Statistical significance is tested with the Clark–West (2007) test using f_t defined as the difference in adjusted forecast errors; the t-stat from regressing f_t on a constant is the CW statistic. Economic value: Two portfolio evaluation methods are employed. (1) Mean-variance investor allocating between oil futures and risk-free bills with margin trading, weight w_t constrained in [−1.5, 1.5], leverage δ ∈ {5, 8, 10}, risk aversion γ = 3, expected excess return via historical average, and volatility forecast from models. Economic value is measured by realized certainty equivalent return (CER) = μ_p − 0.5 γ σ_p^2; CER gain is the annualized difference versus the AR benchmark, with and without 50 bps per-trade transaction costs. (2) Bollerslev et al. (2018) method with constant Sharpe ratio SR (annualized 0.6): w_t = (SR/γ)/√E(RV{t+1}); average utility gain is the difference versus AR. Mechanism analysis: Regressions relate monthly changes in oil fundamentals and emissions to CPU predictors: Δy_t = α + φ X_{CPU,t} + η_t, where Δy_t is change in oil production, oil consumption, or crude-oil-related CO2 emissions. Predictability over business cycles is analyzed via separate R^2_os computed for NBER-defined expansions and recessions. Robustness checks include direction-of-change accuracy, using futures volatility, fixing AR lag to 6, and alternative evaluation periods (starting January 2012 and January 2016).

Statistical predictability:

• In-sample (Table 2): CPU level is not significant (θ = −0.083, t = −1.277, ΔR^2 = 0.584%). CPU change indicators are significant: CCPU (θ = −0.191, t = −2.970, ΔR^2 = 2.097%); ICPU (θ = −0.304, t = −4.791, ΔR^2 = 2.224%). ICPU outperforms CCPU, indicating asymmetric increases contain extra information.
• Out-of-sample (Table 3): CPU has negative R^2_os (expanding −0.219%; rolling −0.492%). CCPU and ICPU yield sizable positive R^2_os with statistical support: expanding window CCPU 5.329% (CW 1.858, 5% sig), ICPU 6.371% (CW 1.409, 10% sig); rolling window CCPU 5.317% (CW 1.846, 5% sig), ICPU 6.584% (CW 1.450, 10% sig). Cumulative squared forecast error plots show models with CCPU/ICPU consistently outperform AR over time. Economic value (Table 4 and Bollerslev et al. metric):
• Mean-variance CER gains (annualized, %) without/with 50 bps costs are negative for CPU and positive for CCPU/ICPU across leverage ratios and windows. Examples: Expanding window, leverage 5, no costs: CPU −0.542, CCPU 0.542, ICPU 0.165. Rolling window, leverage 8, 50 bps: CPU −0.452, CCPU 0.478, ICPU 0.130. Results are robust across δ = 5, 8, 10 and cost settings.
• Bollerslev et al. (2018) average utility gains (%, rolling): CPU −0.252; CCPU 0.113; ICPU 0.072. Expanding window results are qualitatively similar. Thus, CPU changes deliver economically meaningful gains; CPU level does not. Mechanism via fundamentals and emissions (Table 5):
• Production: No significant relation for CPU, CCPU, or ICPU.
• Consumption: Significant negative relations for change indicators, not for CPU. CCPU: φ = −2.171, t = −2.522, R^2 = 2.161%; ICPU: φ = −2.514, t = −2.490, R^2 = 1.216%. CPU level insignificant.
• CO2 emissions: Significant negative relations for change indicators. CCPU: φ = −2.216, t = −2.376, R^2 = 1.887%; ICPU: φ = −3.280, t = −2.737, R^2 = 1.733%. CPU level insignificant. These results align with a consumption channel: higher CPU changes reduce oil consumption and emissions, underpinning volatility predictability. Business-cycle heterogeneity (Table 6):
• Expansions: Strong predictability. Expanding window R^2_os (%): CPU 7.978*; CCPU 11.627**; ICPU 13.798*. Rolling: CPU 4.593; CCPU 11.495**; ICPU 13.442*.
• Recessions: Predictability disappears; R^2_os negative or insignificant across predictors. Controls for other predictors:
• With economic indicators (Table 7): Adding CPU-related predictors to models with 13 economic variables generally improves R^2_os. CPU remains weak; CCPU/ICPU deliver larger, significant gains (e.g., with DY: CCPU 5.427% (CW 1.877, 5%), ICPU 6.541% (CW 1.472, 10%)).
• With other uncertainty indexes (Table 8): Against benchmarks including EPU, GPR, EMV, MPU, VIX, MU, FU, RU, CPU level remains insignificant/negative; CCPU/ICPU remain significantly positive (e.g., with GPR: CCPU 5.297% (CW 1.818, 5%), ICPU 6.796% (CW 1.416, 10%)). CPU change indicators provide independent information not contained in other uncertainty measures. Robustness: Findings hold under direction-of-change accuracy evaluation, using futures volatility, fixing AR lag to 6, and alternative evaluation periods (Jan 2012 and Jan 2016 starts).

The study addresses whether climate policy uncertainty can forecast crude oil price volatility and through which channels. Results show that innovations in CPU (changes and especially increases) predict oil volatility both in-sample and out-of-sample, while the level CPU does not. The stronger performance of ICPU supports an asymmetric attention mechanism where positive surprises in climate policy uncertainty receive more attention and information content. The predictive power is economically meaningful, translating into improved portfolio CER and utility outcomes for oil futures investors, robust to transaction costs and leverage choices. Mechanism analysis ties predictability to real economy fundamentals: CPU changes are linked to declines in oil consumption and related CO2 emissions, but not to production. This supports a demand-side channel: heightened climate policy uncertainty induces firms and consumers to reduce fossil fuel usage and accelerate energy transition investment, which affects crude oil demand and, in turn, market volatility. The concentration of predictability in economic expansions is consistent with stronger demand-side sensitivity when activity is high; during recessions, the link between CPU changes and oil consumption weakens, reducing forecastability. Moreover, CPU change indicators retain predictive content after controlling for a wide array of economic variables and other uncertainty measures, indicating they capture unique climate-risk information beyond traditional predictors. These insights underscore the relevance of transition-risk news for volatility management in energy markets.

The paper shows that changes in climate policy uncertainty—measured by first differences (CCPU) and increases (ICPU) of the CPU index—reliably predict crude oil price volatility, whereas the CPU level does not. The predictive power is statistically robust in-sample and out-of-sample and economically valuable in portfolio applications. The economic source of this predictability is traced to a demand-side mechanism: CPU changes are associated with reductions in crude oil consumption and related CO2 emissions, not with production changes. Predictability is notably stronger during economic expansions and remains after controlling for standard economic predictors and other uncertainty indexes, indicating that CPU changes convey distinct climate-risk information relevant to oil volatility. Policy and practice implications include incorporating climate policy uncertainty innovations into risk management and asset allocation for oil futures, and for policymakers to consider oil market conditions when designing climate policies to avoid exacerbating volatility. Future research directions proposed by the authors include constructing sector-targeted CPU measures (e.g., oil-specific CPU), distinguishing impacts of physical versus transition climate risks on oil volatility, and exploring nonlinear forecasting methods (e.g., random forests, neural networks) to potentially enhance predictability.