The indices show where commodity risk clusters, but adding the Market RII never lowers HAR's out-of-sample RMSE. The descriptive map is the useful result. The forecasting evidence is thinner, as the authors' own table shows.
The paper separates risk into three layers: contract-specific micro risk, sector-level market risk and a latent economy-wide component. It estimates each contract's sensitivity to those layers, then puts them on a common scale through Risk Intensity Indices. The authors impose no asset-pricing restriction and estimate no price of risk. The indices serve as a monitoring tool, with no money mechanism attached.
Fixed sensitivities, changing conditions
Aslanidis, Bariviera, Kapetanios, Sarafidis and Ventouri examine 52 futures at weekly frequency from August 2014 to April 2026, covering 611 weeks. Their universe contains 24 agricultural contracts, 14 energy and petrochemical contracts, and 14 metal contracts. WTI, gold and copper sit alongside local contracts such as Indian cardamom and Dalian eggs. All prices are converted to dollars.
Each contract's weekly return is regressed on the three layers. For the micro layer, the inputs are realised volatility, skewness and kurtosis calculated from daily returns within the week, plus log open interest. The market layer uses leave-one-out sector averages of those four characteristics and the leave-one-out sector return. The macro layer is latent.
Omitted macro drivers may also move the characteristics, so the authors use IV in the first stage. They strip PCA factors from the characteristics and their lags, then use what remains as instruments. Identification depends on the latent macro component lying within the span of those factors, the paper's spanning condition. A second-stage PCA of the residuals finds one component explaining 91.4% of their variation. Boosting with Multiple Testing (BMT), a forward-selection routine governed by a multiple-testing threshold, retains 6 of 36 candidate indicators: the broad dollar index, MOVE, Michigan sentiment, five-year expected inflation, VIX and oil-related geopolitical risk.
To construct each Risk Intensity Index (RII), the authors multiply sensitivities by standardized variables, sum the products within a layer and take the absolute value. They then average the result over time or across contracts. Full-sample shares are 42.1% market, 29.7% macro and 28.2% micro. Energy has a 52.4% market share. Agriculture divides into 36.1% macro, 35.4% market and 28.5% micro; metals into 38.1% market, 33.0% micro and 28.9% macro.
Those shares describe realised risk intensity, as the authors say, rather than a structural variance decomposition or a causal attribution. Sensitivities stay fixed over time. A market-share spike in April 2020 or March 2022 therefore reflects extreme sector characteristics multiplied by full-sample betas. Energy's largest market sensitivity in the return equation is its sector-return beta, at 1.038 by Mean Group (the cross-sectional average of commodity-level estimates) and 1.253 by Median Group (the cross-sectional median), against 0.213 for agriculture. Yet the Market RII excludes that beta by construction. It sums betas on sector volatility, skewness, kurtosis and open interest. Energy's 52.4% market share comes from those terms, including sector volatility at +0.198 for energy versus -0.067 for metals. The paper connects the share to pronounced market sensitivities without specifying which ones.
The macro layer leaves a consequential ambiguity. The summary-statistics table presents selected macro series as index levels, while the text reads the coefficients as changes. We could not tell which enters the regression. If levels enter, persistence alone may account for part of the Macro RII's smooth profile.
Does intensity help forecast volatility?
Both Micro and Market RII coefficients are significant at all three horizons. For one standard deviation of RII (coefficients ×100), the Market RII predicts volatility at 0.212 one week ahead, 0.149 at four weeks and 0.047 at twelve. The corresponding Micro RII figures are 0.083, 0.083 and 0.102. Macro RII remains insignificant throughout, with -0.016 at one week. Across all six volatility and absolute-return specifications, p-values for equality of the micro and market coefficients range from 0.121 to 0.845. These are in-sample regressions, as the authors label them, using betas estimated over the whole sample.
For the out-of-sample exercise, the authors add Market RII to an expanding-window heterogeneous autoregressive (HAR) volatility model. Clark-West p-values at one, four and twelve weeks are 0.029, 0.075 and 0.539. RMSE ratios are 1.0000, 1.0008 and 1.0017; OOS R² values are -0.007%, -0.157% and -0.332%.
The authors acknowledge the point-forecast result: adding Market RII "produces little change in conventional point-forecast accuracy relative to the HAR benchmark." They nevertheless read the Clark-West test as evidence of incremental information at the shorter horizons. Their footnote gives a sound statistical reason. Under a nested null, estimation of the extra coefficient adds noise to the larger model's error, which Clark-West corrects. The one-week test thus suggests a probably nonzero population coefficient. A desk must use estimated coefficients, and the resulting index never beats HAR on RMSE.
We found an out-of-sample test only for Market RII against volatility. The abstract's claim about micro RII and absolute returns relies on the in-sample table. We also did not find a statement saying whether the Market RII betas are re-estimated within each window.
The printed tables disagree
Table 2 has columns headed Green, Stable and DeFi. It gives a full-sample micro share of 16.8% and a market share of 51.2%, at odds with the prose's 28.2/42.1/29.7 split. Table 3 assigns the top 10% of contracts 29.1% of micro RII and the lowest 20% 98.4%. The text gives 30.3% and 4.8%.
Concentration could tell a risk manager which contracts carry the book's risk, making it the most practical output here. Until those tables are fixed, it deserves only a directional reading.
There is a separate econometric qualification. After projection, the leftover sector-return term shrinks at rate 1/N_s, with 14 to 24 contracts in each sector. The panel over-identification test (J = 28.610, p = 0.329) checks the spanning condition only indirectly; it is the sole check reported.
Limits of our own run
The micro layer requires open interest, which our futures data lacks. Volume measures something else. Several of the 36 macro candidates, including the newspaper-based uncertainty indices, also fall outside our FRED coverage. We therefore could not reproduce the selection stage with our data, or rebuild the indices ourselves.
Our view would change with a version of Table 5 that re-estimates betas inside each window and brings the one-week RMSE ratio below 1.0000.