{"id":"e340c4ec-aef2-46c8-b5c8-d3b71f99b507","arxiv_id":"2506.11551","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"FABART, a FAVAR model with BART-based nonlinear factor loadings, is applied to U.S. data, claiming modest forecast gains and sign asymmetries in oil shock transmission.","lead":"The paper introduces FABART, a factor model that uses Bayesian Additive Regression Trees to link observed macro-financial data to latent factors, and applies it to oil price shock analysis. A reader might care because the method claims better forecasts and asymmetric responses to oil shocks than standard linear models, though the evidence is weak.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sign asymmetry cannot arise from the stated model: the measurement equation is linearized via the Crawford et al. projection (Eq. 12-13) and the transition is linear Gaussian, so GIRFs are symmetric by construction.","rationale":"The reader's weakest assumption correctly identifies that the Crawford et al. linear projection in Section 2.2.2 could destroy the nonlinear information needed for asymmetric GIRFs. My stress-test confirms that this is indeed the decisive, load-bearing flaw: as written, the model's inference and impulse responses are built entirely on a linear measurement equation and a linear Gaussian transition. The GIRF definition in Section 2.3.1 is applied to this linear system, so the reported sign asymmetry cannot emerge. The paper never specifies any alternative mechanism — such as simulating GIRFs through the original BART function or conditioning on histories in a way that would break symmetry — and provides no code to clarify. The simulation evidence is also weak: the claimed nonlinear DGP (26) is linear in the factor, and the RMSE differences in Table 2 (e.g., 0.648 vs 0.655) are tiny and unreported with uncertainty. The state-level asymmetry and all narrative conclusions rest on the same irreproducible asymmetry, so the central contribution is unsupported. This fully supports the reader's REJECT verdict; no adjustment to the verdict is needed beyond confirming it.","tokens_in":23671,"tokens_out":1778,"duration_ms":18576,"concrete_test":"Re-implement the estimation and GIRF computation exactly as described: set the loadings via A~_i = F^dagger X_i (Eq. 12), estimate the linear Gaussian state space with transition Eq. (2), and compute GIRFs by Monte Carlo simulation around the long-run mean (Eq. 15) for shocks of +10% and -10% in the oil supply news shock. If the responses are mirror images (or differ only by Monte Carlo noise), the reported sign asymmetry in Figure 7 is not generated by the stated model, and the headline empirical claim is unsupported. A secondary check: run the same replication using the true BART measurement equation (3) and test whether the asymmetry reappears; this would confirm the absence of an explicit nonlinear mechanism in the paper.","verdict_should_be":"REJECT","load_bearing_attack":"The central empirical claim is sign asymmetry in GIRFs (Figures 7-8). But the estimation algorithm replaces the BART measurement equation with the linear approximation X_i,t = A~_i' Y_t + e_i,t (equations 12-13), where A~_i = F^dagger X_i is the Moore-Penrose projection from Crawford et al. (2019). The paper then draws latent factors conditional on this linear Gaussian state space (Carter-Kohn, Section 2.2.2 step 5) and computes GIRFs from the VAR transition equation (2), which is linear with Gaussian shocks. With a linear transition and linear measurement, the conditional expectation in Eq. (14) is linear in the shock, so responses to positive and negative shocks of equal magnitude are exact mirror images and no sign asymmetry can emerge. The BART nonlinearity F(Y_t) from Eq. (3) is estimated only to obtain A~_i; it never enters the state-space model used for smoothing or the GIRF simulation. The asymmetry reported in Section 4.3.1 therefore cannot arise from the model as described. The simulation also provides no check: Table 2 reports only small RMSE improvements, and the 'nonlinear' DGP (26) X_t = B^2 F_t + V_t is actually linear in F_t; only DGP (27) tanh(BF_t) is genuinely nonlinear. No code or data accompany the paper, so the reported asymmetry cannot be independently reproduced.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FABART, a factor-augmented VAR in which the measurement equation linking observables to latent factors is modeled with Bayesian additive regression trees. Estimation proceeds by constructing a linear approximation to the BART function and then drawing factors from a linear Gaussian state space via Carter-Kohn. The framework is applied to recursive forecasting of U.S. macro-financial variables and to generalized impulse responses to oil supply news shocks identified with the Känzig (2021) external instrument. The paper claims improved forecasts for industrial production, especially during the COVID-19 period, and pronounced sign asymmetries in the transmission of oil supply news shocks at both aggregate and state levels.","tokens_in":24034,"tokens_out":5754,"duration_ms":58720,"significance":"If the claims were valid, FABART would be a useful addition to the nonlinear factor-model literature and the empirical oil-shock literature, and the forecasting evidence for industrial production would be of practical interest. The paper engages with a relevant literature, spells out priors and the estimation algorithm in detail, and connects its results to existing theoretical explanations of oil-price asymmetries. However, the central empirical contribution—sign asymmetry in GIRFs—is not obtainable from the model as described, because the measurement equation used for inference and impulse responses is a linear projection of the BART fit, and the transition equation is linear Gaussian. The simulation evidence also labels a linear DGP as nonlinear. These issues affect the paper’s core claims rather than only its presentation.","major_comments":[{"comment":"The estimation algorithm replaces the BART function F(Y_t) with the linear approximation X_{i,t} = \\tilde A_i' Y_t + e_{i,t}, where \\tilde A_i = F^\\dagger X_i, and the transition equation (2) is linear with Gaussian shocks. The Carter-Kohn step then conditions on this linear Gaussian state space. Consequently, the conditional expectation in the GIRF definition (14) is linear in the shock; for a positive and a negative shock of equal magnitude, the responses must be exact mirror images. The sign asymmetries reported in Section 4.3.1 and Figure 7 therefore cannot arise from the model as specified. The BART nonlinearity is not used in the GIRF simulation, so the reported asymmetry is not a model output but an artifact of the implemented procedure.","section":"Sections 2.2.2 and 2.3.1, Eqs. (12)–(14)"},{"comment":"The “quadratic nonlinearity” DGP is X_t = B^2 F_t + V_t. Since B is a fixed 20×1 vector, B^2 is a fixed set of loadings, and this equation is linear in the factor F_t; it does not introduce any nonlinearity or sign dependence. Only the tanh specification in Eq. (27) is genuinely nonlinear. The simulation and Monte Carlo sections therefore do not establish recovery under two distinct nonlinear DGPs; they compare one linear and one nonlinear DGP. The near-identical RMSEs for “Nonlinear I” and “Nonlinear II” in Table 2 (0.648 vs. 0.650) are unsurprising given this structure.","section":"Section 3, Eq. (26)"},{"comment":"The forecast evaluation reports RMSEs and log scores without any measure of uncertainty, such as standard errors, Diebold-Mariano tests, or Bayesian posterior intervals. For example, the industrial production 12-month RMSE improvement (1.725 vs. 2.773) is presented as a substantial gain, but without uncertainty quantification it is impossible to assess whether the difference is statistically meaningful. This is a load-bearing limitation for the forecasting claim, which is a prominent part of the abstract and conclusions.","section":"Section 4.2, Table 4 and Figures 2–3"}],"minor_comments":[{"comment":"The two subsections both provide the identification of oil price shocks and contain nearly identical equations; one appears to be a leftover draft version and should be removed.","section":"Sections 2.3.2 and 2.3.3"},{"comment":"The note describes the figure as illustrating “size asymmetry” while the surrounding text discusses sign asymmetry; the terminology should be made consistent.","section":"Figure 7 note"},{"comment":"The text states that the structural analysis uses the pre-pandemic sample 1974:01–2016:12, while Table 3 and the dataset description report 1974M1–2024M08; the sample used for the GIRFs should be stated unambiguously.","section":"Section 4.3.1 vs. Table 3"},{"comment":"Step 5 says “Conditional on the draws obtained in Steps 1.-5.”, which references itself; it should presumably refer to Steps 1–4.","section":"Section 2.2.2, algorithm Step 5"},{"comment":"The manuscript contains several typos, including “hightened”, “postive”, and “F ABART” in the title, and no replication code or data are provided, which is particularly unfortunate given that the GIRF results cannot be verified from the text alone.","section":"General"}],"recommendation":"reject","confidential_remarks":"The central empirical claim is not supported by the model as written; the linearized measurement equation and linear Gaussian transition imply symmetric impulse responses. The simulation DGP labeling issue compounds the problem. These are not local fixes: restoring a genuinely nonlinear link between the BART function and the GIRFs would require a fundamental re-specification of the estimation and impulse-response procedure. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's headline result—pronounced sign asymmetries in the transmission of oil supply news shocks—does not survive contact with the model as written. The measurement equation is replaced by the linear projection X_i,t = A~_i' Y_t + e_i,t (equations 12–13), and the transition is a linear Gaussian VAR. With both blocks linear, conditional expectations in equation (14) are linear in the shock, so positive and negative shocks of equal size produce mirror-image responses. The BART nonlinearity enters only to construct A~_i; it never appears in the state space used for smoothing or in the GIRF simulation. The asymmetry in Figures 7–8 and the state-level results is therefore not a consequence of the model presented. This is a load-bearing flaw, not a minor gap.\n\nWhat is genuinely new: putting BART in the measurement equation of a FAVAR, rather than in the VAR coefficients as BAVART does, is a legitimate combination. The state-level application is a useful extension, and the literature review is fair and reasonably complete. The writing is mostly clear, though the duplicated sections 2.3.2/2.3.3 and the Monte Carlo appendix text suggest the manuscript was not fully polished.\n\nThe simulation evidence is weaker than the text implies. The first 'nonlinear' DGP, X_t = B^2 F_t + V_t, is linear in the factor—squaring the loading matrix does not make the mapping nonlinear. Only the tanh DGP is genuinely nonlinear. And the reported RMSE gains are small: 0.648 versus 0.655 for the linear FAVAR. There is no uncertainty quantification around these numbers. The evaluation sample is also described inconsistently (Table 4 says 2014:01–2024:06; the text at one point says 2023:07–2024:06). No code or data are provided, so none of the empirical results can be independently checked.\n\nWho is this for? An econometrician working on nonlinear factor models or asymmetric oil-shock transmission might find the setup worth discussing, but in its current form the central empirical conclusion is not credible. That said, the methodological idea is worth engaging with: a referee could help the author see where the linearization kills the nonlinearity and how to fix it, for example by using the BART function directly in the GIRF simulation rather than only in the loading projection. I would tell an editor to send it to a referee, but with a clear instruction to verify the GIRF construction before anything else.","headline":"The sign-asymmetry claim cannot arise from the model as written: the BART nonlinearity is linearized away before the GIRFs are computed, even though the BART-in-FAVAR idea itself is worth a careful look.","tokens_in":24515,"tokens_out":3217,"would_cite":false,"duration_ms":31739,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes FABART, a factor-augmented VAR whose factor loadings are estimated by Bayesian additive regression trees, and claims that this nonparametric mapping improves industrial-production forecasts and reveals sign asymmetries…","keywords":["Bayesian additive regression trees","factor-augmented VAR","nonparametric factor model","oil supply news shocks","sign asymmetry","generalized impulse response functions","forecast evaluation","large macroeconomic panels"],"falsifier":"Fit the FABART model to a dataset generated by a known nonlinear factor model, then compute GIRFs two ways: through the linear projection $\\tilde A_i = F^\\dagger X_i$ and through direct simulation of the fitted BART function. If the linearized responses are statistically indistinguishable from a linear FAVAR's symmetric responses, or if the direct and projected responses diverge, the claimed sign asymmetry is not being carried by the model as described.","tokens_in":23460,"feed_emoji":"🛢️","tokens_out":7599,"duration_ms":73157,"temperature":0.7,"pith_summary":"This paper builds a factor-augmented vector autoregression in which the link between observable macro-financial series and latent factors is estimated by Bayesian additive regression trees rather than by a linear mapping. It argues that this nonparametric FAVAR, called FABART, recovers latent factors under nonlinear data-generating processes without overfitting when the truth is linear, and that it forecasts U.S. industrial production more accurately than linear benchmarks during turbulent periods. The paper also claims that oil supply news shocks transmit asymmetrically: a positive shock that raises real oil prices contracts real activity and raises inflation more strongly and persistently than a negative shock expands activity, with the same sign asymmetry visible in state-level employment. If true, the work matters because linear factor models would miss both the forecast gains and the asymmetric propagation.","feed_headline":"Oil supply shocks hit harder when they raise prices, model finds","feed_subtitle":"Regression-tree factor model also forecasts industrial production better than linear benchmarks.","key_machinery":"The load-bearing object is the sum-of-trees function $f_i(Y_t)=\\sum_{s=1}^{S} g_{is}(Y_t\\mid \\tau_{is},\\mu_{iS})$ with $S=250$ regression trees, each a recursive binary partition of the factor space with constant terminal-node values, estimated under the BART regularizing prior by Metropolis-Hastings tree moves and Bayesian backfitting. What carries the argument is a two-step bridge: the fitted tree function is projected onto the observables through $\\tilde A_i = F^\\dagger X_i$, the Moore-Penrose projection of $X_i$ on the matrix of fitted nonlinear values $F$, producing a linear measurement equation $X_{i,t}=\\tilde A_i' Y_t+\\epsilon_{i,t}$ that permits Gaussian state-space draws of the latent factors. Impulse responses are then computed as generalized impulse response functions conditional on the long-run mean, which is how the paper lets shock sign and history enter the responses.","core_discovery":"The central claim is that replacing the linear factor-loading equation with a sum-of-trees approximation, where each observable is a function of the latent factors estimated by 250 regularized regression trees, lets latent factors inherit nonlinear structure from the data and preserves that structure when computing generalized impulse responses. In the simulations, FABART delivers lower forecast errors for the latent factor than a linear FAVAR under a linear DGP and two nonlinear DGPs, and Monte Carlo replications show posterior factor estimates with average correlations above 0.96. In the empirical application, FABART improves forecasts of industrial production and keeps predictive densities stable during the COVID-19 period, while the linear BVAR and FAVAR benchmarks deteriorate. The paper's headline empirical finding is sign asymmetry in oil supply news shock transmission: positive shocks generate stronger and more persistent contractions in real activity and inflation than negative shocks generate expansions, and state-level employment contracts more after positive shocks than it improves after negative shocks.","pith_inferences":["The reported sign asymmetry may partly come from the linear projection used to sample factors; a direct comparison of GIRFs from the full tree function against GIRFs from the projected linear measurement equation would show how much of the asymmetry is genuinely produced by the BART model.","The same nonparametric factor-loading device could be applied to other observed factors, such as the federal funds rate or an uncertainty index, where the paper's identification strategy would transfer without change.","The state-level cross-sectional regressions suggest a sectoral mechanism, but a panel GIRF design with time-varying state exposure could test whether the manufacturing-share channel is causal rather than a slow-moving proxy.","Because the external instrument covers only part of the sample, the asymmetry estimates use a shorter identification window; extending the instrument or using narrative sign restrictions would reveal whether the result is sample-dependent."],"forward_implications":["Under nonlinear data-generating processes, FABART recovers the latent factor with lower RMSE than a linear FAVAR and with posterior correlations above 0.96 across 100 Monte Carlo replications.","In the empirical application, FABART forecasts U.S. industrial production with lower RMSE and better log scores than a linear BVAR and a linear FAVAR at one-, three-, and twelve-month horizons, and its density forecasts stay stable when the COVID-19 shock reaches the data.","The estimated GIRFs imply that a positive oil supply news shock of 10 percent raises inflation and contracts U.S. and global industrial production more strongly and persistently than an equal-sized negative shock expands them.","State-level employment responses show the same sign asymmetry: most states fall below the 45-degree symmetry line, and the average difference is significantly negative; manufacturing-heavy states contract more while mining-intensive states are more insulated.","For real oil prices the forecast gains are small, and for financial variables such as the excess bond premium and the S&P 500 the linear BVAR remains competitive, so the nonparametric factor stage mainly helps real-activity forecasts and density stability."],"supporting_citations":[{"why":"Introduces the FAVAR framework whose measurement equation and factor structure FABART extends.","marker":"Bernanke et al. (2005)"},{"why":"Supplies the sum-of-trees prior, the regularizing hyperparameters, and the Bayesian backfitting sampler used for the factor-loading functions.","marker":"Chipman et al. (2010)"},{"why":"Provides the linear-projection device $\\tilde A_i = F^\\dagger X_i$ that lets the MCMC draw latent factors conditional on the BART function.","marker":"Crawford et al. (2019)"},{"why":"Defines the Generalized Impulse Response Function used to compute history-dependent, sign-sensitive responses in the nonlinear model.","marker":"Koop et al. (1996)"},{"why":"Provides the external instrument from oil futures around OPEC announcements and the sign and magnitude benchmark for the oil supply news shock responses.","marker":"Känzig (2021)"},{"why":"Motivates the nonparametric VAR approach and the forecast-stability argument during the COVID-19 period.","marker":"Huber et al. (2020)"},{"why":"Documents forecasting gains from nonlinear dimension reduction in turbulent periods, the comparison class for FABART's empirical results.","marker":"Hauzenberger et al. (2023)"},{"why":"Supplies the sign-asymmetry phenomenon in oil price-output relations that the paper's GIRFs are compared against.","marker":"Hamilton (2003)"},{"why":"The FRED-MD database defines the large panel of observables used for factor extraction.","marker":"McCracken and Ng (2016)"}],"fun_headline_variants":["Tree-based factor model forecasts better, shows oil shock asymmetry","Non-parametric factor model beats linear benchmarks, finds oil shock asymmetry","Positive oil supply shocks cause stronger U.S. contractions than negative ones","FABART: tree-based factor model improves forecasts, reveals asymmetric oil shocks","Oil price rise shocks hurt U.S. economy more than price fall shocks, says tree model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The sign-asymmetry result rests on the assumption that the linear projection $\\tilde A_i = F^\\dagger X_i$ used to draw the latent factors keeps the nonlinear information in the BART function; if the projection acts as an ordinary linear loading matrix, the measurement equation is linear, the impulse responses are symmetric, and the reported asymmetry cannot arise.","fun_headline_variants_meta":{"raw":{"variants":["Tree-based factor model forecasts better, shows oil shock asymmetry","Non-parametric factor model beats linear benchmarks, finds oil shock asymmetry","Positive oil supply shocks cause stronger U.S. contractions than negative ones","FABART: tree-based factor model improves forecasts, reveals asymmetric oil shocks","Oil price rise shocks hurt U.S. economy more than price fall shocks, says tree model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3221,"prompt_tokens":936,"completion_tokens":2285,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2203}},"tokens_in":552,"tokens_out":2285,"duration_ms":15154,"temperature":1.0,"reasoning_tokens":2203,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:03:57.642439+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the FABART model to a dataset generated by a known nonlinear factor model, then compute GIRFs two ways: through the linear projection $\\tilde A_i = F^\\dagger X_i$ and through direct simulation of the fitted BART function. If the linearized responses are statistically indistinguishable from a linear FAVAR's symmetric responses, or if the direct and projected responses diverge, the claimed sign asymmetry is not being carried by the model as described.","supporting_citations":[{"cited_title":"S., Boivin, J., and Eliasz, P","cited_arxiv_id":null,"evidence_quote":"Introduces the FAVAR framework whose measurement equation and factor structure FABART extends."},{"cited_title":"A., George, E","cited_arxiv_id":null,"evidence_quote":"Supplies the sum-of-trees prior, the regularizing hyperparameters, and the Bayesian backfitting sampler used for the factor-loading functions."},{"cited_title":"R., Runcie, D","cited_arxiv_id":null,"evidence_quote":"Provides the linear-projection device $\\tilde A_i = F^\\dagger X_i$ that lets the MCMC draw latent factors conditional on the BART function."},{"cited_title":"H., and Potter, S","cited_arxiv_id":null,"evidence_quote":"Defines the Generalized Impulse Response Function used to compute history-dependent, sign-sensitive responses in the nonlinear model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the nonparametric VAR approach and the forecast-stability argument during the COVID-19 period."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents forecasting gains from nonlinear dimension reduction in turbulent periods, the comparison class for FABART's empirical results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sign-asymmetry phenomenon in oil price-output relations that the paper's GIRFs are compared against."}],"review_version":1}