{"id":"e770ea1c-ff5e-45ac-b661-228ef7711ab6","arxiv_id":"2501.03945","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"The paper adapts Sequential Monte Carlo sampling to mixed causal-noncausal vector autoregressions and uses MDD and BIC to select the polynomial orders and the error distribution.","lead":"This paper shows how to estimate mixed causal-noncausal time series models with a parallelizable Sequential Monte Carlo sampler instead of Markov chain Monte Carlo, and uses model selection to choose both the lag structure and the error distribution. The approach is tested on simulated data and on a bivariate series of European ESG stocks and Brent oil prices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MDD leg of the proposed identification is not reliable: under a Skewed-t DGP, Eq. (5) with P=10000, M=100, lambda=2 selects the true VMAR(1,1) only 62% of the time, and the printed MDD formula omits the 1/P normalization.","rationale":"The reader's weakest assumption identifies the same load-bearing point: the SMC estimate of the marginal data density must be accurate enough for model selection under the chosen tuning. Table 1 provides direct evidence that it is not, particularly for the Skewed-t DGP, where the wrong error distribution is selected 21.5% of the time by MDD. This matters because the paper's stated novelty is the joint identification of polynomial orders and error distribution using MDD and BIC. BIC performs better, but the paper presents MDD as a central by-product and does not define the BIC formula in the text. I do not see a need to move the verdict from CONDITIONAL to a harsher category: the point-estimation results in Table 2 are mostly plausible, the BIC leg works in the simulations, and the MDD issue could potentially be fixed by more stages, a corrected formula, or a more robust marginal-likelihood estimator. However, the printed Eq. (5) is also missing the 1/P normalization, which is an internal inconsistency that should be corrected regardless. Other concerns, such as the sum-versus-product typo in Eqs. (9)-(11), the lack of uncertainty measures in the empirical application, and the absence of code, reinforce the need for conditional acceptance but do not replace the MDD accuracy issue as the primary load-bearing concern.","tokens_in":14653,"tokens_out":8560,"duration_ms":92601,"concrete_test":"Take the Skewed-t DGP from Section 4.3 with B=200 replications, P=10000, M=100, lambda=2, and focus on the replications where MDD selects Student-t VMAR(1,1) over the true Skewed-t VMAR(1,1). For those replications, compute an independent marginal likelihood for the true model using a Chib-Jeliazkov or Geweke-Keane estimator on the same posterior draws or on a long MCMC run. If the independent estimate ranks Skewed-t above Student-t in these cases, then Eq. (5) is biased at the proposed tuning and the MDD-based identification claim fails; if the independent estimate also prefers Student-t, then the issue is genuine model uncertainty rather than estimator error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two load-bearing parts: accurate posterior estimation and reliable identification. The first is supported reasonably by Table 2, though with a reported standard-error anomaly for the skewness parameters. The second, which is the paper's main novelty, rests on the MDD from Eq. (5) being accurate. Table 1 shows it is not robust: for the Skewed-t DGP, MDD identifies the true VMAR(1,1) with Skewed-t errors only 62.0% of the time (vs. 94.0% for BIC) and selects the wrong error distribution (Student-t, same orders) 21.5% of the time. The authors attribute this to needing more stages (Section 4.3), but M=100 with P=10000 and lambda=2 is the fixed configuration used throughout, so the reported 'strong ability' does not hold for one of the three error distributions considered. Moreover, Eq. (5) as printed is internally inconsistent: it is a product of sums of incremental weights, while the correction step in Appendix B normalizes by (1/P) times the sum. The missing 1/P changes the numerical MDD by a factor of P^M and, if implemented literally, invalidates any reported MDD level, although it would cancel in model comparisons that use the same P. A direct validation of the MDD estimator against an independent marginal-likelihood estimator is therefore needed before the identification claim can be accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Sequential Monte Carlo (SMC) algorithm for Bayesian estimation of mixed causal-noncausal vector autoregressive models (VMAR(r,s)) with Student-t, Cauchy, and Skewed-t error distributions. It introduces a model-selection procedure, based on the SMC-computed marginal data density (MDD) and the Bayesian information criterion (BIC), that chooses both the causal and noncausal polynomial orders and the innovation distribution. Monte Carlo experiments with T=150 and B=200 replications for a bivariate VMAR(1,1) report posterior estimates close to the true values under all three error distributions, with BIC selecting the true specification in 83--94% of replications. The MDD selection is strong for Cauchy and Student-t DGPs but drops to 62% for the Skewed-t DGP. An empirical application to the S&P Europe 350 ESG Index and Brent crude oil prices selects a VMAR(1,1) model with Student-t errors.","tokens_in":15069,"tokens_out":7286,"duration_ms":65639,"significance":"If the numerical issues identified below are corrected, the paper would offer a useful computational alternative for noncausal models: SMC avoids the mode search and MCMC tuning of existing methods, is parallelizable, computes the MDD as a byproduct, and accommodates heavy-tailed and skewed innovation distributions. The simultaneous identification of polynomial orders and error distribution is a genuine extension of previous Bayesian noncausal VAR work. The Monte Carlo point estimates in Table 2 are mostly close to their true values, and BIC identifies the correct specification in a large majority of replications. However, the current manuscript contains a likelihood formula error, an unnormalized MDD formula, and a strong MDD-selection failure for the Skewed-t DGP, so the abstract's claim of 'strong ability ... to correctly identify the process' is not yet supported.","major_comments":[{"comment":"The likelihood is written as a sum over t of the innovation densities, and Eq. (9) repeats the sum inside the density. For an i.i.d. innovation sequence the likelihood is the product over t of the densities, not the sum. As printed, the posterior is not the posterior of the stated model; if the implementation actually uses a product, the displayed formulas should be corrected. This is load-bearing because both MDD and BIC are computed from this likelihood.","section":"Section 4.2, Eqs. (9)-(11)"},{"comment":"The MDD estimator is printed as the product over m of the unnormalized sum of incremental weights, with no 1/P factor, while the correction step in Appendix B normalizes the incremental weights by (1/P) times the sum. The correct estimator is the product over m of (1/P) times the sum of the incremental weights. The missing factor changes the numerical MDD by P^M, so the MDD levels reported in Table 3 are not valid as printed. The factor cancels in comparisons that use the same P, but the formula must be corrected and independently validated.","section":"Section 3.2, Eq. (5), and Appendix B"},{"comment":"Under the Skewed-t DGP, the MDD criterion selects the true VMAR(1,1) with Skewed-t errors only 62.0% of the time and instead selects a Student-t error distribution for the same orders 21.5% of the time, while BIC achieves 94.0%. Since P=10000, M=100, and lambda=2 are the fixed tuning values used throughout the Monte Carlo study, the abstract's 'strong ability' claim is not supported for one of the three error distributions considered. The paper should either temper the claim, demonstrate a practically feasible configuration that fixes the MDD performance, or benchmark the MDD estimator against an independent marginal-likelihood estimator.","section":"Section 4.3, Table 1"},{"comment":"The reported average of the per-replication standard errors (STD) is 0.0117 and 0.0143 for alpha_1 and alpha_2, respectively, while the Monte Carlo variance of the point estimates is 0.2349 and 0.2931, corresponding to a Monte Carlo standard deviation of about 0.48, and the RMSE is 0.4950 in both cases. The per-replication standard errors are implausibly small relative to the dispersion of the point estimates, suggesting an error in the computation or reporting of posterior standard deviations. This matters because the accuracy claim for the skewness parameters rests on these numbers.","section":"Table 2, Skewed-t DGP rows for alpha_1 and alpha_2"}],"minor_comments":[{"comment":"The caption says 'MMD' where the text and table body use 'MDD'; this should be corrected.","section":"Table 1 caption"},{"comment":"The BIC value for the Student-t VMAR(0,1) model is reported as 1521.6, identical to the Cauchy VMAR(0,1) value; this is likely a typo and should be checked.","section":"Table 3, BIC row"},{"comment":"The expression uses the symbol n for the stage index in the incremental weights, but the stage index elsewhere is m; use a consistent notation.","section":"Section 3.2, Eq. (5)"},{"comment":"The definition of Theta_1 lists vec(Psi_1),...,vec(Psi_s) and vec(Phi_1),...,vec(Phi_r), but Psi is the causal polynomial of order r and Phi is the noncausal polynomial of order s; the indices appear to be swapped and should be corrected.","section":"Section 3.1"},{"comment":"The notation p(y|Theta) = sum over t of p(u_t|Theta) appears before Eq. (9); if the intended likelihood is a product, the summation notation should be removed throughout Sections 4.2 and Appendix A.","section":"Section 4.2, likelihood definition"},{"comment":"The data are described as monthly from July 2017 to February 2024, but the horizontal axis of Figure 1 starts in 2015; please clarify the discrepancy.","section":"Section 5 and Figure 1"},{"comment":"The paper does not report random seeds or provide replication code; adding code or a detailed pseudocode with tuning choices would strengthen reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first SMC treatment of noncausal vector autoregressions, and the BIC-based selection results are fairly encouraging. But the MDD leg of the identification fails one of the three DGPs in their own Table 1 (62% hit rate for Skewed-t vs 94% for BIC), and Eq. (5) as printed omits the 1/P normalization in the MDD. These are fixable, but the abstract's 'strong ability' is over-stated.\n\nWhat is new: adapting the Herbst-Schorfheide/Bognanni-Herbst sampler to VMAR(r,s) models, and proposing MDD/BIC to select among Cauchy, Student-t, and Skewed-t errors jointly with orders. That is a genuine step beyond the MCMC-based, Student-t-only work of Lanne et al. (2012) and Lanne and Luoto (2016). The Monte Carlo estimates in Table 2 are mostly close to truth, with small bias and RMSE, which supports the core algorithm.\n\nSoft spots: (1) The MDD formula in Eq. (5) and Appendix B is missing the 1/P factor. As printed, the reported MDD levels would be off by P^M; the actual numbers in Table 3 suggest the implementation used the correct formula, but the text is wrong and needs to be fixed. (2) The MDD model selection under a Skewed-t DGP selects the true VMAR(1,1) with Skewed-t errors only 62% of the time and picks Student-t 21.5% of the time. The authors note this and blame the need for more stages, but it remains a weakness in the proposed identification procedure. BIC does better, so the paper's contribution is not entirely lost, but the abstract should not claim 'strong ability' uniformly. (3) Table 2 has an internal inconsistency: the reported average standard errors for the skewness parameters α1, α2 are 0.0117 and 0.0143, while the Monte Carlo variances are 0.2349 and 0.2931. That's a factor of 20 mismatch and needs a correction. (4) The empirical application gives point estimates but no standard errors or credible intervals, and no code or data are provided.\n\nCirculation of references is fine; they cite the relevant MCMC and SMC literature. The paper is honest about its limitations in Section 4.3, which earns goodwill.\n\nBottom line: the central estimator is not obviously broken; the identification claim is under-supported at the edges. A serious referee could get this into publishable shape with revisions. I would send it to review.","headline":"A useful but uneven SMC adaption for noncausal VARs: the estimator works, but the MDD identification leg is weaker than advertised and the printed MDD formula needs fixing.","tokens_in":15619,"tokens_out":5199,"would_cite":true,"duration_ms":43981,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62M10","65C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A sequential Monte Carlo sampler estimates noncausal vector autoregressions and jointly selects orders and error distribution.","keywords":["mixed causal-noncausal autoregressive models","sequential Monte Carlo","Bayesian estimation","marginal data density","model selection","heavy-tailed error distributions","vector autoregression","noncausal identification"],"falsifier":"Rerun the Skewed-t VMAR(1,1) Monte Carlo with larger settings such as $M=300$ and $P=30000$: if the MDD identification rate stays near 62% while BIC stays near 94%, the fixed tuning is the bottleneck and the MDD-based identification claim is conditional on it. A direct check is to compare the SMC marginal data density against a high-accuracy estimate from a much longer MCMC run; systematic disagreement for Skewed-t data would show the byproduct MDD is not reliable enough for model choice.","tokens_in":14417,"feed_emoji":"📈","tokens_out":7304,"duration_ms":65142,"temperature":0.7,"pith_summary":"The paper sets out to show that Sequential Monte Carlo, an algorithm that evolves a swarm of parameter draws through gradually more data-informed distributions, is a practical and fast Bayesian estimator for mixed causal-noncausal autoregressive models. The payoff is that the same sampler works for Cauchy, Student-t, and Skewed-t errors and returns the marginal data density as a byproduct, so model selection can choose the causal and noncausal order and the error distribution in one pass. Simulation evidence shows accurate posterior means and mostly correct identification; the authors report that BIC identifies the true model more often than MDD when the error is Skewed-t. This matters because noncausal models are used to model bubbles and expectation-driven dynamics, and existing Bayesian tools are tied to Student-t errors and are slower to run.","feed_headline":"SMC estimates noncausal VARs and chooses the error distribution","feed_subtitle":"Also picks between Cauchy, Student-t and skewed-t errors; applied to ESG stocks and Brent oil, it selects VMAR(1,1).","key_machinery":"The load-bearing object is the tempered Sequential Monte Carlo sampler: a swarm of $P$ particles is drawn from the prior, then moved through $M$ bridge distributions $\\pi_m(\\Theta)\\propto [p(y|\\Theta)]^{\\rho_m}p(\\Theta)$ with $\\rho_m=((m-1)/(M-1))^\\lambda$, using correction (reweighting), selection (resampling when the effective sample size drops below $P/2$), and mutation (short MCMC steps with invariant distribution $\\pi_m$). This schedule is what lets the sampler find multiple modes instead of getting stuck in local minima, and it makes the marginal data density $\\hat p(y)=\\prod_{m=1}^M \\sum_i \\tilde w^i_m$ a free byproduct rather than a separate computation. The supporting identification machinery is the grid of 21 candidate models (seven order pairs across three error distributions), compared by MDD and BIC.","core_discovery":"The paper claims that the Sequential Monte Carlo sampler of Herbst and Schorfheide (2014) and Bognanni and Herbst (2018) can serve as a general Bayesian engine for vector mixed causal-noncausal autoregressions $\\Psi(L)\\Phi(L^{-1})y_t=u_t$, whatever the error distribution, as long as the posterior kernel can be evaluated pointwise. It supplies posterior mean estimates and, without extra computation, the marginal data density, which together with BIC selects the causal order $r$, noncausal order $s$, and the error law. In Monte Carlo exercises with Cauchy, Student-t, and Skewed-t errors, the estimates have small bias and RMSE, and both criteria usually recover the true VMAR(1,1); the paper reports that BIC outperforms MDD when the errors are Skewed-t. Applied to detrended monthly S&P Europe 350 ESG Index and Brent oil prices, the procedure chooses a bivariate VMAR(1,1) with Student-t errors and finds positive cross-effects in both directions, which the authors read as expectation-driven dynamics.","pith_inferences":["The underperformance of MDD under Skewed-t at the chosen tuning suggests that users should either prefer BIC for heavy-tailed asymmetric DGPs or increase $M$ and $P$; the paper itself notes more stages would help but become computationally costly.","The same sampler could be extended to other non-Gaussian error families, Markov-switching parameters, or stochastic volatility, since the only requirement is pointwise kernel evaluation.","A practical test of the identification rule would be to compare SMC MDD values against a long-run MCMC marginal-likelihood estimate; if they disagree mainly for Skewed-t data, a correction or adaptive tempering would be needed."],"forward_implications":["For any error distribution with a pointwise-evaluable posterior kernel, the same SMC code replaces bespoke MCMC samplers, removing the Student-t-only restriction in Bayesian noncausal modelling.","Marginal-data-density computation becomes free, so Bayes factors for comparing $r,s$ and error laws no longer require a separate marginal-likelihood step.","Because particles are independent, estimation can be parallelized across cores, cutting wall-clock time that currently limits Monte Carlo studies and empirical applications.","In the ESG/oil application, the selected VMAR(1,1) with Student-t errors implies that expected ESG index movements feed into Brent prices and vice versa, giving a testable forward-looking channel."],"supporting_citations":[{"why":"Introduces mixed causal-noncausal autoregressions and the non-Gaussian identification condition that the SMC estimator builds on.","marker":"Breidt et al. (1991)"},{"why":"Defines the vector VMAR(r,s) representation and supplies the approximate likelihood functions used for estimation.","marker":"Lanne and Saikkonen (2013)"},{"why":"Establishes the earlier Bayesian MCMC estimation of univariate noncausal models that this paper extends.","marker":"Lanne et al. (2012)"},{"why":"Provides the multivariate Bayesian VMAR baseline, prior settings, and notes the non-elliptical posteriors SMC addresses.","marker":"Lanne and Luoto (2016)"},{"why":"Supplies the tempered sequential Monte Carlo algorithm that the paper adapts.","marker":"Herbst and Schorfheide (2014)"},{"why":"Provides the MDD formula, the transition-density details for the mutation step, and the multiple-equation SMC variant.","marker":"Bognanni and Herbst (2018)"},{"why":"Gives the importance-sampling convergence theory underlying the particle approximations.","marker":"Geweke (1989)"},{"why":"Supports the parallelization claim that motivates SMC over MCMC.","marker":"Durham and Geweke (2014)"}],"fun_headline_variants":["SMC estimates noncausal VARs and selects error law","Parallel SMC identifies orders and error distribution","SMC for Bayesian noncausal vector autoregressions","SMC picks error distribution in mixed causal-noncausal models","Noncausal VAR order and error selection via SMC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the SMC estimate of the marginal data density is accurate enough at the fixed settings $P=10000$, $M=100$, $\\lambda=2$ to drive model selection; if that estimate is biased for heavy-tailed or asymmetric posteriors, the proposed Bayes-factor identification fails even when posterior means are close to the truth.","fun_headline_variants_meta":{"raw":{"variants":["SMC estimates noncausal VARs and selects error law","Parallel SMC identifies orders and error distribution","SMC for Bayesian noncausal vector autoregressions","SMC picks error distribution in mixed causal-noncausal models","Noncausal VAR order and error selection via SMC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1532,"prompt_tokens":911,"completion_tokens":621,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":541}},"tokens_in":527,"tokens_out":621,"duration_ms":5873,"temperature":1.0,"reasoning_tokens":541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:43:48.262683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the Skewed-t VMAR(1,1) Monte Carlo with larger settings such as $M=300$ and $P=30000$: if the MDD identification rate stays near 62% while BIC stays near 94%, the fixed tuning is the bottleneck and the MDD-based identification claim is conditional on it. A direct check is to compare the SMC marginal data density against a high-accuracy estimate from a much longer MCMC run; systematic disagreement for Skewed-t data would show the byproduct MDD is not reliable enough for model choice.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces mixed causal-noncausal autoregressions and the non-Gaussian identification condition that the SMC estimator builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the vector VMAR(r,s) representation and supplies the approximate likelihood functions used for estimation."},{"cited_title":"Luoma, and J","cited_arxiv_id":null,"evidence_quote":"Establishes the earlier Bayesian MCMC estimation of univariate noncausal models that this paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multivariate Bayesian VMAR baseline, prior settings, and notes the non-elliptical posteriors SMC addresses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the tempered sequential Monte Carlo algorithm that the paper adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MDD formula, the transition-density details for the mutation step, and the multiple-equation SMC variant."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the importance-sampling convergence theory underlying the particle approximations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the parallelization claim that motivates SMC over MCMC."}],"review_version":1}