China's -0.149 UCT coefficient is hard to square with a uniformly positive volatility story. Finland's is +0.254. Both are significant at 1%, both come from the same specification, yet the abstract describes "a positive relationship between UCT and stock market volatility." Anyone considering the series as a volatility input has to reckon with that spread.
The index and the test
Rogers, Sun and Sun built the US-China Tension index, a monthly newspaper count published on policyuncertainty.com. Following the Baker, Bloom and Davis construction, it tracks the share of articles in leading US papers that mention both countries, refer to a contentious bilateral issue, and contain tension-related language.
Salisu and Nsonwu test whether the series forecasts equity volatility. Their sample contains 347 monthly observations from January 1995 to February 2024. It covers 16 developed and 18 emerging index markets from investing.com, classified by MSCI. Realized volatility is the dependent variable, calculated as the annualized standard deviation of monthly returns across a three-month rolling window. Each observation therefore measures the dispersion of three monthly returns, while consecutive observations share two of those three by construction. Lagged volatility enters the regressions and absorbs some of that persistence.
The authors estimate each country separately with the Westerlund-Narayan predictive regression. Lagged UCT is the coefficient of interest. The contemporaneous change in UCT addresses endogeneity from predictor persistence, and geopolitical risk enters as a control. Panel A uses Caldara and Iacoviello's global GPR. Panel B substitutes their country-specific indices.
The sample is then divided 80:20 for Clark-West tests of nested models at horizons of 3, 6 and 12 months. A Bayesian dynamic multivariate panel model from Helske and Tikka comes last. It imposes an acyclic ordering from UCT into GPR into volatility, then simulates 20% and 40% UCT shocks in 2010.
There is no portfolio anywhere in the paper. It reports no returns, costs or Sharpe. Any economic use would have to pass through vol targeting or option positioning, which the paper does not attempt.
Can one sign represent the estimates?
With global GPR as the control, 15 of 16 developed markets have positive UCT coefficients significant at 1%. France is at 0.048 and Finland reaches 0.254. Australia is the exception at -0.034. Under country-specific GPR, Australia's coefficient moves to -0.010 and becomes insignificant. The UK falls to 0.023 at the 5% level, while Hong Kong drops to 0.006. The sign survives even as the size shrinks sharply.
Emerging markets do something else. Brazil is 0.064, Chile 0.136, Malaysia 0.072, Mexico 0.079, Poland 0.111 and Taiwan 0.041, all positive and significant. Yet under the same global GPR control, eight of 18 are negative and significant: China -0.149, Colombia, Egypt -0.140, Hungary, India -0.078, the Philippines, Saudi Arabia -0.084 and Thailand -0.070. Four more, Korea, Peru, South Africa and Turkey, are insignificant.
The body acknowledges the split. The authors say the mixed pattern "suggests that UCT shock transmission is less uniform in emerging markets". Their conclusion nevertheless states, "The results show that UCT significantly increases short-term stock market volatility", without mentioning the negative coefficients. Markets with negative and insignificant coefficients are grouped together as those that "appear relatively insulated".
China resists that description. A coefficient of -0.149 at 1% associates rising bilateral tension with lower Shanghai Composite realized volatility, even though Shanghai is the market one would least expect to be insulated from US-China tension. The paper offers only general possibilities, including financial integration, trade exposure, institutional structures and policy buffers. We did not find a China-specific explanation.
Forecast gains move with the control
The out-of-sample analysis is the strongest part of the paper. The authors state directly that "significant in-sample predictability does not necessarily translate into improved out-of-sample forecasts", and the forecast tables support the warning.
For developed markets under global GPR, 15 of 16 Clark-West statistics exceed the 10% one-sided value of 1.282 at all three horizons. Hong Kong barely clears it, with 1.470, 1.461 and 1.418. Finland leads at 7.252, 7.440 and 7.698, followed by Israel at 5.780, 6.087 and 6.479. Belgium fails with 0.373, 0.415 and 0.318.
Country-specific GPR changes the membership. Belgium enters at 1.612 at h=3, while Spain at 0.748 and Portugal at 0.642 fall away. Among emerging markets, the switch removes forecast gains for the Philippines, at -1.198 at every horizon, as well as Brazil, from 0.508 to 0.641, Korea, from 0.484 to 0.688, and South Africa, from 0.224 to 0.559. Mexico remains insignificant with either control, posting 0.958, 1.168 and 1.220 under global GPR. UCT retains a forecasting signal, but the set of markets carrying it changes.
We did not find an MSFE ratio, a utility calculation or a comparison with a GARCH-family benchmark. Clark-West pits the full model against the same model without UCT. The test establishes that the predictor contributes something, while leaving the size of that contribution unanswered. This is the same gap discussed in our note on a volatility ranking with no defined target, carried one step further.
One result deserves to survive the criticism. UCT's correlation is 0.0056 with global GPR and 0.0245 with country-specific GPR, leaving it close to orthogonal to the standard geopolitical risk measure. UCT also predicts commodity volatility where GPR does not. The in-sample coefficients are 0.091 for gold and 0.126 for oil, both significant at 1%, versus insignificant GPR coefficients of 0.012 and 0.009.
Gold keeps the out-of-sample result, with statistics of 5.312, 4.943 and 4.172. Oil fails at h=3 with 1.207 and at h=6 with 1.144. It clears the 5% value only at h=12, reaching 1.733.
A simulated fade
The policy conclusion comes entirely from the dynamic panel exercise. The authors call it "a form of counterfactual analysis" and select 2010 as the shock year because of its "symbolic global significance" after the financial crisis. They provide no event study of the 2018 tariff escalation or another actual episode. The authors choose the 20% and 40% shock sizes themselves, while the ordering from UCT to GPR to volatility is imposed rather than tested.
The conclusion retains the appropriate qualification: "the counterfactual dynamic simulations suggest that the volatility effects may be transitory under the model assumptions". The abstract removes that qualification and says "the volatility effects associated with US-China tensions are short-lived". Policy advice to "caution against rapid policy changes" relies on the stronger version.
The simulated fade also clashes with forecast statistics that generally increase with horizon. Finland rises from 7.252 at h=3 to 7.698 at h=12. A rapidly dissipating effect should produce the weakest improvement at the twelve-month-ahead horizon, rather than the strongest of the three.
The authors recognize the timing problem created by newspaper indices. Such measures "may not perfectly coincide with the exact timing at which financial markets incorporate new information", they write. Their defence is that the indices "provide a systematic measure of the perceived information environment facing investors and have been shown to closely track policy actions, business decisions, and financial market outcomes". That answer is adequate for in-sample inference. For a monthly series intended for trading, we did not find any discussion of publication lag or whether the month's value is available at that month's close.
We could not test these claims. The UCT index is absent from our data catalogue, and reconstructing it would require the newspaper corpus and topic-model keyword logic used by Rogers, Sun and Sun. The sample begins in January 1995, long before our price coverage, and consists almost entirely of non-US equity indices.
A single-market economic-value test would change my mind. Take the US or Hong Kong series, convert the UCT forecast into a variance position or a vol-targeted equity book, and report the net result. Until such a test exists, this remains a credible measurement paper about an index close to orthogonal to GPR. Under the global GPR control, however, nine of 34 markets are negative and significant, with four more insignificant.