{"id":"f2adf072-1348-45ea-a4c2-4507870a484f","arxiv_id":"1908.06325","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Bayesian global VAR with drifting coefficients and factor volatility in mean produces an endogenous international uncertainty measure and finds time-varying negative real and financial effects of uncertainty shocks.","lead":"This paper builds a statistical model of six large economies in which international uncertainty is a hidden factor that moves all economies together and also affects their average behavior. It estimates that uncertainty shocks raise unemployment and lower industrial production, inflation, and equity prices, with effects that differ across countries and over time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's headline impulse responses to an uncertainty shock are never defined: no shock size, no normalization, and no propagation algorithm for the latent ht are given, so the central quantitative claims in Section 4.2 cannot be checked as stated.","rationale":"The reader's weakest assumption is that ht is separately identified from the factor loadings and fixed σh, and that volatility innovations are exogenous. My concern is closely related but more primitive: even before discussing exogeneity, the paper never defines the impulse response to an uncertainty shock. This is stated in the reader's rationale — no shock size, no normalization, no algorithm — but the formal 'weakest_assumption' field emphasizes identification and exogeneity rather than the missing IRF definition. The full text supports my concern: Section 2.1 describes the model and says the setup can be used for impulse responses, but it does not derive them; Appendix B lists the MCMC steps and contains no IRF step; Section 4.2 presents the figures without a computational definition; and footnote 7 explicitly withholds numerical tables. These are limitations flagged in the manuscript itself, and they bear directly on the central claim that uncertainty shocks have 'pronounced real and financial effects' with a particular time profile. I therefore do not propose changing the reader's conditional verdict, but I would make the condition explicit: the authors must supply the IRF definition and a replication to verify the reported magnitudes. My agreement is 'partial' because the reader's formal weakest assumption centers on structural identification, whereas my load-bearing concern is the absence of a well-defined IRF object; the two are complementary, and the missing IRF definition must be resolved before the exogeneity question can even be assessed.","tokens_in":20128,"tokens_out":6765,"duration_ms":80167,"concrete_test":"Obtain a complete written definition of the impulse response and an independent implementation: from the posterior draws of βit, L, ht, and ωij,t, simulate an innovation to ξt at date t0 of size one unit and, separately, of size one standard deviation (σh), propagate it through ht+1 = ht + ξt and through the variance term exp(ht)LL′ and the mean term βit ht, and recompute the median and credible bands of the responses in Figures 4 and 5. In addition, re-estimate with σh sampled instead of fixed at 0.2 under the same prior, and report whether the response magnitudes shift by more than the plotted credible intervals. If the responses change materially, the headline effects are not robust as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 and the abstract claim that uncertainty shocks have 'pronounced real and financial effects' with time-varying magnitudes, but the impulse response object behind that claim is not specified anywhere in the paper. Equations (1) and (2) define a nonlinear model in which ht enters the mean equation through βit ht and scales the factor covariance through exp(ht)LL′, while ht follows a random walk with innovation variance σh. No equation, algorithm, or appendix entry defines the shock: is it a one-unit change in ht, a one-standard-deviation innovation ξt, or some other size? How is the shock propagated through both the mean and variance equations? What baseline path and what normalization of ht are used? The paper explicitly fixes σh = 0.2 in footnote 1 because the likelihood is flat, and footnote 7 says detailed tables are available only on request, so the reported response magnitudes are not reproducible from the manuscript. Because ht is a latent state with no explicit scale normalization, observationally equivalent reparameterizations may change the numerical responses unless the IRF definition is supplied. This gap is load-bearing because the abstract's quantitative conclusions are read directly from Figures 4 and 5; if the required definition is supplied and changes magnitudes or sign patterns, the central claims would need revision.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian global vector autoregression with drifting coefficients and factor stochastic volatility in mean for six advanced economies, using monthly data from 1991:04 to 2018:07. A scalar latent process h_t drives the common volatility of factor innovations and also enters the mean equations with country- and variable-specific coefficients. Estimation uses hierarchical normal-gamma priors that shrink the model toward constant coefficients and homoscedastic errors. The paper reports the estimated h_t as an international uncertainty measure, compares it with several text-based and VIX-based proxies, and presents impulse response functions to an uncertainty shock for unemployment, industrial production, exports, inflation, equity prices, and Nelson-Siegel yield-curve factors. The main empirical claim is that uncertainty shocks produce pronounced real and financial effects in all countries, with magnitudes and timing that differ across economies and over time.","tokens_in":20349,"tokens_out":3871,"duration_ms":42825,"significance":"If the results hold, the paper makes a useful methodological contribution by combining a global VAR with drifting parameters, factor stochastic volatility in mean, and hierarchical shrinkage priors, and it extends the international uncertainty literature to a joint treatment of time variation and cross-country heterogeneity. The appendices provide a fairly detailed MCMC algorithm, posterior credible intervals are reported, and the estimated uncertainty measure is compared with established proxies. The central quantitative claim, however, rests on an impulse-response object that is never formally defined in the manuscript. Because the reported magnitudes in Figures 4 and 5 and the abstract's 'pronounced real and financial effects' are read directly from that object, the significance is conditional on the authors supplying a precise definition of the uncertainty shock and its propagation, as well as a scale normalization for the latent h_t.","major_comments":[{"comment":"The impulse response to an international uncertainty shock is never defined. The paper does not state whether the shock is a one-unit change in the latent h_t, a one-standard-deviation innovation to the random walk h_t, or some other size; it does not specify how the shock is propagated through the mean equation (beta_it h_t) and the factor covariance exp(h_t)LL'; and it gives no baseline path or normalization of h_t. Footnote 7 states that detailed numerical tables are available only upon request, so Figures 4 and 5 cannot be reproduced from the manuscript as it stands. This is load-bearing because the abstract's quantitative conclusions are read directly from these figures, and footnote 1 reports that the likelihood is flat in sigma_h, so the response magnitudes are sensitive to the adopted normalization and shock definition.","section":"Section 4.2, Eqs. (1)-(2), Appendix B"},{"comment":"The scale and location of the latent process h_t are not identified from the likelihood: Var(epsilon_t) = exp(h_t)LL' + Omega_t is invariant to multiplying exp(h_t) by a constant and rescaling L and Omega_t accordingly, while the random walk for h_t has no location anchor. The paper fixes sigma_h = 0.2 but does not state a location normalization, such as a prior mean for h_0 or a sum-to-zero constraint. The posterior median values of h_t in the range of about -10 to -2 in Figure 1 are therefore arbitrary up to the prior. If the impulse response is a one-unit shock to h_t, the magnitudes in Section 4.2 depend directly on this normalization; if it is a sigma_h-standardized innovation, the magnitudes depend on the fixed value of sigma_h. The authors should state the normalization explicitly and show that the qualitative conclusions are robust to alternative normalizations.","section":"Section 2.1, Eq. (1), Fig. 1"},{"comment":"The recursive identification of an exogenous uncertainty shock is asserted rather than formalized. The text relates the approach to ordering uncertainty indices first, but h_t is an estimated latent process from the same system whose responses are then attributed to uncertainty shocks. Because the priors on beta_it are centered at zero and do not impose sign restrictions, the negative responses in Section 4.2 are data outcomes rather than imposed restrictions. As a correctness check, the authors should state explicitly what structural assumption identifies the innovation to h_t as an exogenous uncertainty shock, and report a robustness experiment, for example including lags of h_t in the mean or using an alternative ordering, that supports the causal interpretation.","section":"Section 2.1, identification of the uncertainty shock"}],"minor_comments":[{"comment":"The statement that detailed tables are available upon request is not consistent with current reproducibility standards; please provide numerical impulse-response tables in the paper or in a supplementary file.","section":"Footnote 7"},{"comment":"There is a typo: \"bancruptcy\" should be \"bankruptcy\".","section":"Section 4.1"},{"comment":"There is a typo: \"simliar\" should be \"similar\".","section":"Appendix C"},{"comment":"The note \"1992:01 to 2017:07 on biannual frequency\" does not explain how the color gradient maps to time; please clarify which colors represent early and late periods.","section":"Figure 4 note"},{"comment":"Some statements about \"significant\" responses appear inconsistent with credible sets that cover zero, for example for industrial production and exports; please define the significance criterion used (e.g., 68% or 90% posterior interval excluding zero) and apply it consistently.","section":"Section 4.2"},{"comment":"The Nelson-Siegel decay parameter lambda is fixed at 0.0609 rather than estimated; since the yield-curve factors are constructed from this parameter, a brief sensitivity discussion would strengthen the results.","section":"Section 3, Eq. (8)"}],"recommendation":"major_revision","confidential_remarks":"The paper overlaps in motivation with Carriero et al. (2018a) and Mumtaz and Musso (2019), but its empirical framework with drifting coefficients, factor stochastic volatility in mean, and hierarchical shrinkage is sufficiently distinct. The main obstacle is the missing formal definition of the impulse response and the identification/normalization of h_t; these are fixable within the manuscript's scope. I do not see grounds for rejection on novelty or technical soundness grounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The reader's conditional verdict is about right, and the stress-test note lands. The paper is a competent, incremental but real extension of the Bayesian TVP-VAR toolkit, and its uncertainty measure looks sensible; however, the central IRF results are not defined anywhere in the paper, so the quantitative claims in Section 4.2 cannot be verified as written.\n\nWhat's new: the hierarchical global-local prior setup that centers country-specific coefficients and state innovation variances on a common cross-country mean, with an additional shrinkage layer on the common means. That is a genuine, useful innovation over existing NG and double-gamma priors, and the MCMC details in Appendices A and B are careful. The uncertainty measure ht is constructed endogenously, matches event chronology (GFC peak, Euro crisis, Brexit/Trump), and compares reasonably to GPR/GEPU/WUI/VIX. That part is solid and worth publishing.\n\nThe soft spots are real. No equation or algorithm defines the impulse response to an uncertainty shock. The model has ht in both the mean (through βit ht) and the covariance (through exp(ht) LL'), so a shock to ht propagates through both, but the paper never says whether the shock is one unit, one standard deviation of ξt, or something else; nor how future ξt are handled; nor what baseline ht path is used. Without that, the response magnitudes and even the sign patterns in Figures 4 and 5 are not checkable. This is not a cosmetic omission, because the whole empirical contribution is those figures.\n\nThe identification story also leans on a recursive ordering putting uncertainty first, which is the standard but contestable assumption; the paper cites Carriero et al. (2019) for support but does not resolve the scale identification of ht. Fixing σh = 0.2 with a footnote saying the likelihood is flat is acceptable only if accompanied by sensitivity analysis; the grid search is described, but no results are shown. Footnote 7 says detailed tables are available on request, which is no substitute for a replication package.\n\nBottom line: I'd send this to a serious referee, with the clear request that the IRF definition, shock normalization, sensitivity to σh, and replication materials be supplied before publication. The framework is worth engaging, but the central quantitative claims are currently not reproducible.","headline":"Competent, incremental extension of Bayesian TVP-GVAR with a genuinely new hierarchical prior, but the headline uncertainty-shock IRFs are never defined, so the central quantitative claims cannot be checked as written.","tokens_in":20950,"tokens_out":2699,"would_cite":false,"duration_ms":29562,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that one latent process, $h_t$, extracted from six economies' data, measures international uncertainty, and that shocks to it raise unemployment and depress output, exports, inflation and equity prices, with effects that…","keywords":["international uncertainty","global vector autoregression","factor stochastic volatility","stochastic volatility in mean","hierarchical shrinkage priors","time-varying parameters","uncertainty shocks","Bayesian state-space model"],"falsifier":"Regress the first differences of the posterior median of $h_t$ on lagged industrial production, unemployment, inflation, equity prices, and bond yields from all six countries; if those lagged variables jointly predict the next innovation in $h_t$, the 'uncertainty first' ordering fails.","tokens_in":19858,"feed_emoji":"📉","tokens_out":10110,"duration_ms":86967,"temperature":0.7,"pith_summary":"This paper tries to establish that a single latent quantity, the common factor volatility $h_t$ inside a six-economy global vector autoregression, is a valid measure of international macroeconomic uncertainty. The model estimates $h_t$ endogenously from the comovement of national data, and also lets $h_t$ enter the mean of each country's VAR, so a shock to $h_t$ is an uncertainty shock that moves both the level and the volatility of the system. The paper argues that such shocks raise unemployment and depress industrial production, exports, inflation and equity prices in all six economies, with timing and magnitude differing across countries and over time. A sympathetic reader should care because this is a single coherent way to measure international uncertainty and trace its real and financial consequences without relying on external news-based indices.","feed_headline":"One latent gauge tracks world uncertainty and its shocks bite","feed_subtitle":"A six-economy Bayesian VAR with drifting parameters links one latent uncertainty factor to real and financial effects.","key_machinery":"The central object is the scalar log-volatility $h_t$ of the common factors in the error decomposition $\\epsilon_t = L f_t + \\eta_t$, with $f_t \\sim N(0, \\exp(h_t) I_d)$ and $h_t$ following a random walk. This $h_t$ is also included in the mean of every country's VAR equation, which is what turns a change in $h_t$ into an uncertainty shock rather than just a volatility adjustment. The estimation machinery is a non-centered state-space parameterization combined with hierarchical Normal-Gamma (double Gamma) priors that shrink the model toward constant coefficients, homoscedastic errors, and cross-country homogeneity while still allowing time variation and country idiosyncrasies if the data demand them. That shrinkage is what makes the high-dimensional multi-country system tractable and keeps the uncertainty measure identified.","core_discovery":"On its own terms, the central discovery is that the estimated log-volatility process $h_t$ behaves like a meaningful international uncertainty measure: it rises around the Asian and Russian crises, 9/11, the Iraq War, the global financial crisis, the European sovereign debt crisis, and the Brexit/Trump period, and it tracks established proxies such as GPR, GEPU, WUI and the VIX. Treating an innovation to $h_t$ as an exogenous uncertainty shock, the model finds significant increases in unemployment (up to eight basis points in the United States), declines in industrial production, exports, inflation and equity prices, and a flattening of the yield curve across all six countries. The responses are not constant: cumulative effects on equity prices and some other variables diminish over the sample, and the strength of the real effects varies over time, with muted effects in the aftermath of the Great Recession for several series.","pith_inferences":["Editorial extension: the identification of $h_t$ could be stress-tested by re-estimating the model with zero restrictions that keep slow-moving real variables from responding on impact; if the impulse responses change materially, the uncertainty-first ordering is doing the work.","Editorial extension: because $h_t$ is extracted from six advanced economies, applying the same machinery to a panel that includes emerging markets could reveal whether the scalar factor remains international or becomes a proxy for US or European stress.","Editorial extension: the estimated $h_t$ could be exported as an observable regressor in single-country studies of uncertainty transmission, giving an external check on whether it captures information beyond text-based uncertainty indices."],"forward_implications":["The estimated $h_t$ series can serve as an international uncertainty indicator that peaks at the global financial crisis, the European sovereign debt crisis, and the Brexit/Trump episodes, and moves with GPR, GEPU, WUI and VIX.","Uncertainty shocks raise unemployment in every country in the sample, with the largest cumulative effects in Germany and the United States.","Industrial production, exports, inflation and equity prices fall in response to an uncertainty shock, and the yield curve flattens as term spreads narrow.","The strength of these effects changes over time: equity-price responses shrink across the sample, and post-Great-Recession responses are muted for several real variables before reverting later."],"supporting_citations":[{"why":"Supplies the factor stochastic volatility decomposition of the error covariance that defines the common volatility $h_t$.","marker":"Aguilar and West (2000)"},{"why":"Introduces the global VAR cross-sectional weighting scheme used to build the foreign variables $y_{it}^*$.","marker":"Pesaran et al. (2004)"},{"why":"Precedent for reading the common factor volatility as uncertainty and for comparing the resulting impulse responses.","marker":"Crespo Cuaresma et al. (2017)"},{"why":"Provides the endogenous-uncertainty benchmark whose measurement and real effects this model extends to multiple economies with time variation.","marker":"Carriero et al. (2018b)"},{"why":"Supplies evidence that macroeconomic uncertainty does not respond endogenously, supporting the recursive identification of uncertainty shocks.","marker":"Carriero et al. (2019)"},{"why":"Establishes the recursive ordering that puts the uncertainty measure first, the identification assumption behind the impulse responses.","marker":"Bloom (2009)"},{"why":"Contributes the double-Gamma prior on state innovation variances used to shrink the drifting coefficients.","marker":"Bitto and Frühwirth-Schnatter (2019)"},{"why":"Provides the non-centered parameterization that lets the model place shrinkage priors on state innovation variances and volatility variances.","marker":"Frühwirth-Schnatter and Wagner (2010)"}],"fun_headline_variants":["Latent factor reveals global uncertainty's real bite","One gauge tracks world shocks that hit jobs and markets","Drifting-parameter VAR finds uncertainty's time-varying toll","Bayesian latent factor captures crisis-driven economic pain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach depends on treating the single estimated volatility component as an exogenous uncertainty shock that arrives before the economy moves, rather than as a symptom of movements already under way.","fun_headline_variants_meta":{"raw":{"variants":["Latent factor reveals global uncertainty's real bite","One gauge tracks world shocks that hit jobs and markets","Drifting-parameter VAR finds uncertainty's time-varying toll","Bayesian latent factor captures crisis-driven economic pain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000134,"raw_usage":{"total_tokens":1102,"prompt_tokens":870,"completion_tokens":232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":170}},"tokens_in":486,"tokens_out":232,"duration_ms":2996,"temperature":1.0,"reasoning_tokens":170,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:49:58.502670+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Regress the first differences of the posterior median of $h_t$ on lagged industrial production, unemployment, inflation, equity prices, and bond yields from all six countries; if those lagged variables jointly predict the next innovation in $h_t$, the 'uncertainty first' ordering fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the factor stochastic volatility decomposition of the error covariance that defines the common volatility $h_t$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the global VAR cross-sectional weighting scheme used to build the foreign variables $y_{it}^*$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Precedent for reading the common factor volatility as uncertainty and for comparing the resulting impulse responses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the recursive ordering that puts the uncertainty measure first, the identification assumption behind the impulse responses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the double-Gamma prior on state innovation variances used to shrink the drifting coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the non-centered parameterization that lets the model place shrinkage priors on state innovation variances and volatility variances."}],"review_version":1}