{"id":"88948b7c-d1b3-4a6c-a0e3-fa155ed5edce","arxiv_id":"2506.01094","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A semiparametric stochastic volatility model with dependent nonparametric error terms is proposed, but its claimed estimation advantages are not convincingly established because the comparison and algorithm are flawed.","lead":"This paper lets the random shocks in a stochastic volatility model follow any distribution and be correlated, using a kernel density estimate of the error shape. The authors report better estimation accuracy than a Gaussian benchmark, but the benchmark is unfair and the algorithm's details are internally inconsistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 5.3 claims NSVM-3 'generally produces lower' MAPE, but Tables 3 and 4 show NSVM-3 has the highest MAPE in every column of Table 3 and is never the lowest in Table 4, directly contradicting the paper's central volatility-accuracy claim.","rationale":"The reader's verdict of REJECT is well-founded. Among the concerns, I identify the internal contradiction between Section 5.3's claim and Tables 3 and 4 as the single most load-bearing because it directly undermines the paper's central quantitative claim without requiring any assumptions about MCMC coherence. Even if the plug-in kernel target in Eq. (12) were a perfectly valid posterior, the reported numbers show NSVM-3 does not 'generally produce lower' volatility estimation errors: it has the worst MAPE in the Gaussian simulation and is not better than stochvol in the t simulation. The parameter tables (Tables 1 and 2) mostly support the claim for parameters, but the volatility tables do not, and the differences are tiny and unaccompanied by uncertainty quantification. This was noted in the reader's rationale ('reported MAPE numbers contradict the text') but not listed as the weakest assumption; I agree with the reader's overall rejection but place the emphasis on the empirical contradiction, which is directly checkable from the manuscript. The plug-in posterior issue remains a serious secondary concern that would also need addressing, but the arithmetic failure is sufficient on its own.","tokens_in":13151,"tokens_out":7180,"duration_ms":70173,"concrete_test":"Recompute or re-read the MAPE rows of Tables 3 and 4 from the simulation outputs: order the three models by MAPE for each of Mean/Median/Mode. If NSVM-3 is highest in all three columns of Table 3 and not lowest in Table 4, then the Section 5.3 statement 'generally produces lower MAPE' is false. Optionally, run a paired Wilcoxon or t-test on the 100 per-run MAPE differences NSVM-3 minus stochvol; if the mean/median difference is not significantly negative, the volatility-accuracy claim lacks statistical support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline conclusion of superior volatility estimation rests on the error comparisons in Section 5.3. The text states that 'Tables 3, 4 indicate that NSVM-3 generally produces lower ... mean absolute percentage error (MAPE).' This is contradicted by the tables themselves. In Table 3 (Gaussian errors), NSVM-3 MAPE Mean/Median/Mode are 0.09959/0.09939/0.09944, the largest of all three models (stochvol: 0.09945/0.09927/0.09912; Gaussian: 0.09932/0.09908/0.09922). In Table 4 (Student-t errors), NSVM-3 MAPE values (0.09949/0.09959/0.09966) are slightly worse than stochvol (0.09948/0.09957/0.09965), though better than Gaussian. Thus the claimed 'generally lower' MAPE is false. The only apparent srMSE advantages (e.g., 0.008250 vs 0.008252 in Table 3) are in the fifth decimal, with no standard errors or significance tests across the 100 runs, so no meaningful superiority is demonstrated. The central claim fails even before considering the plug-in posterior target in Eq. (12); this arithmetic inconsistency alone invalidates Section 5.3's conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NSVM-3, a semiparametric stochastic volatility model in which the joint density of the return innovation and the volatility innovation is estimated nonparametrically by a bivariate kernel density estimate, thereby allowing for non-Gaussianity and dependence between the two error processes. The method is implemented by replacing the Gaussian components in the full conditional posterior for the latent volatility with the kernel density estimate, while the model parameters are sampled from conditional posteriors derived under the Gaussian model. The paper compares NSVM-3 against a Gaussian SV model and the R package stochvol in simulations with dependent Gaussian and Student-t errors, reporting lower square-root mean squared errors for parameters and, allegedly, for volatility estimates. An empirical application to S&P 500 daily returns is presented as an illustration. The central methodological and empirical claims are, however, not supported by the evidence in the manuscript.","tokens_in":13481,"tokens_out":4016,"duration_ms":42305,"significance":"The topic is relevant: relaxing both the distributional and the independence assumptions on the shocks of a stochastic volatility model is a natural extension, and a well-posed semiparametric Bayesian procedure could be of value to financial econometrics. The paper also has a constructive element: it provides a full MCMC algorithm and presents simulation and empirical comparisons. But the manuscript as it stands does not establish a coherent inferential target: the plug-in kernel density in Eq. (12) is treated as the true error density without correction for estimation from the same data, the parameter samplers do not use the semiparametric likelihood, and the reported volatility-accuracy claim is directly contradicted by the paper's own tables. With a properly derived posterior and a fairer comparison design the idea might become publishable, but the current version does not provide a valid basis for the claimed advantages.","major_comments":[{"comment":"The posterior target in Eq. (12) uses a bivariate kernel density estimate \\hat{k} that is constructed from standardized residuals \\hat{u}_t and \\hat{w}_t obtained from a preliminary Gaussian SV fit on the same dataset. This estimate is plugged into the full conditional of h_t as if it were the exact joint density of (u_t, \\nu_t). No correction is made for the fact that \\hat{k} is estimated from the data, and no consistency, coverage, or Bernstein-von Mises type result is provided. Consequently, the MCMC samples produced by Algorithms 2 and 4 are not draws from a well-defined posterior distribution of the semiparametric model; they are draws from a procedure that uses the data twice. This is a load-bearing issue because the entire inference, including the parameter and volatility estimates in Tables 1-4, depends on this target.","section":"Section 3, Eq. (12)"},{"comment":"The samplers for \\delta, \\alpha, and \\sigma_\\nu use the conditional posterior distributions derived from the Gaussian parametric model in Eqs. (5)-(7), and Algorithm 3 explicitly states that 'Let p be the posterior distribution in a parametric setting.' Thus the parameter estimates reported in Tables 1 and 2 are not estimates under the semiparametric model with dependent, nonparametrically distributed errors; they are Gaussian-model estimates. The paper therefore provides no evidence that the proposed semiparametric framework improves parameter estimation, because the parameters are not actually being estimated from the semiparametric likelihood.","section":"Section 4.1, Algorithms 3, 7, and 8"},{"comment":"The text claims that 'Tables 3, 4 indicate that NSVM-3 generally produces lower ... mean absolute percentage error (MAPE),' but this is contradicted by the numbers in the tables. In Table 3, NSVM-3 has the largest MAPE for all three posterior summaries (0.09959, 0.09939, 0.09944) compared to stochvol (0.09945, 0.09927, 0.09912) and Gaussian (0.09932, 0.09908, 0.09922). In Table 4, NSVM-3's MAPE values (0.09949, 0.09959, 0.09966) are slightly worse than stochvol's (0.09948, 0.09957, 0.09965). The srMSE and MAE differences are in the fifth decimal place (e.g., 0.008250 vs 0.008252) and are reported without standard errors or significance tests across the 100 runs. The claim of superior volatility estimation, which is central to the paper's conclusion, is therefore not supported by the presented results.","section":"Section 5.3, Tables 3 and 4"},{"comment":"The simulation comparison omits a parametric baseline that allows for correlation between the error terms. The data are generated with \\rho(u_t, \\nu_t) = -0.5, yet the two comparators (the Gaussian SV model and the stochvol implementation as used here) assume independent errors. A natural baseline would be a parametric SV model with bivariate Gaussian or bivariate Student-t errors and an estimated correlation parameter. Without such a baseline, the reported gains of NSVM-3 in parameter srMSE and in volatility error metrics could be attributed entirely to the inclusion of dependence rather than to the nonparametric flexibility of the error density. The paper never reports an estimate of the correlation parameter, so this alternative explanation is not addressed.","section":"Sections 5.1 and 5.2"},{"comment":"The empirical application to the S&P 500 does not provide any quantitative measure of volatility estimation accuracy. Since the true volatility is unobserved in real data, the visual comparison in Figure 10 cannot establish that NSVM-3 'responds more accurately' to large fluctuations. No model comparison metrics, such as predictive likelihoods, coverage, or out-of-sample losses, are reported. The empirical claim is therefore not substantiated beyond anecdotal visual inspection.","section":"Section 6"}],"minor_comments":[{"comment":"The text states that the simulation was repeated 50 times, and the box plots refer to that number, but the srMSE formulas and the table captions use 100 runs. Please clarify whether the results are based on 50 or 100 replications.","section":"Section 5.1"},{"comment":"The formula for s appears to contain typographical errors: the term 's1δs1' in the manuscript is not a valid expression, and the summation expressions following it are garbled. The definitions of s1, s2, s3 in Eq. (9) are clear, but Eq. (8) needs to be rewritten consistently.","section":"Equation (8)"},{"comment":"The symbol N is used both for the sample size and for the kernel function within the same equation (the kernel is written as N(x, y)), which is confusing. Use a different letter, e.g., K, for the kernel.","section":"Equation (12)"},{"comment":"The tuning constant c* is stated to be 1.2 in the text, but Algorithm 5 and the later algorithms use log(1.1) in the code-like pseudocode. The relationship between c* and the 1.1 factor should be explained or made consistent.","section":"Section 4.1"},{"comment":"The definition of c uses h_m as the mode of q, but in Algorithm 2 the acceptance steps use a constant c that is defined after its first use in the pseudocode. Reordering the algorithm so that c is computed before the proposal step would improve readability.","section":"Section 4.1, Eq. (14)"},{"comment":"The abstract claims lower 'bias and variance,' but the paper only reports srMSE, which combines bias and variance. Reporting bias and variance separately, or at least stating that srMSE is a combined measure, would be more precise.","section":"Abstract and Sections 5.1-5.2"},{"comment":"No convergence diagnostics are reported for the MCMC chains, such as trace plots, effective sample sizes, or Gelman-Rubin statistics. Given the nontrivial proposal mechanism, the reader cannot assess whether the MCMC results in Tables 1-4 are reliable.","section":"Section 4"},{"comment":"Algorithm 1 is only a sketch and references parameters 'T' and 'b' that are not defined in the main text (likely burn-in and total iterations). These should be defined explicitly in the algorithm caption.","section":"Section 4, Algorithm 1"},{"comment":"The reference to 'Feng and Gangopadhyay (2025)' is cited as an arXiv preprint without a full citation or version; the reader cannot easily locate or verify the claimed methodological relationship.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper's central claims are not supported by its own reported results. The posterior target is a plug-in approximation with no correction, the parameter samplers are not based on the semiparametric likelihood, and the volatility comparison claim is contradicted by Tables 3 and 4. These are not presentation issues; they affect the validity of the method and the conclusions. I would be willing to look at a substantially revised version that derives a coherent posterior (or an approximate posterior with stated assumptions), includes a proper parametric correlated-error baseline, and reports volatility metrics that actually support the claimed advantages."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about arXiv:2506.01094. First, the model idea is real: extending your earlier NSVM-1/2 to a bivariate kernel density for the errors (u_t, ν_t) is a natural step, and it is actually new relative to Feng and Gangopadhyay (2025). Second, the evidence presented for it does not hold up. The paper's central claim—that NSVM-3 gives lower volatility estimation error—is directly contradicted by its own Tables 3 and 4, and the MCMC target is a plug-in approximation with no correction for the estimated density.\n\nWhat it does well: the bivariate kernel formulation is clearly written, the algorithms are spelled out, and the simulation design covers both Gaussian and t errors. The parameter srMSE results in Tables 1 and 2 do look favorable for NSVM-3.\n\nWhere it falls apart. Section 5.3 states NSVM-3 'generally produces lower' MAPE. In Table 3, NSVM-3 has the highest MAPE in all three columns (mean, median, mode). In Table 4, it is never the lowest. The srMSE advantages are in the fifth decimal, with no standard errors or significance tests over the 100 runs. So the headline conclusion about volatility is not supported by the very numbers in the paper.\n\nThe deeper issue is the posterior target. Equation (12) puts \\hat{k}(y_t/√h_t, (ln h_t − μ_t)/σ_ν) into the posterior for h_t, where \\hat{k} is a kernel density estimate computed from residuals of a preliminary Gaussian fit on the same data. No account is taken of estimating \\hat{k}. The parameter samplers (Algorithms 6–8) still use the Gaussian posterior densities, so the δ, α, σ_ν estimates are actually from the parametric model, not the semiparametric one. That makes the favorable Tables 1 and 2 hard to interpret.\n\nThe comparison is also tilted: the baseline Gaussian model is misspecified in exactly the dimension NSVM-3 relaxes (dependence). A parametric leverage SV model, or a correlated-error t model, would be the right control. Without it, the gains could just reflect dependence rather than nonparametric flexibility. Minor issues: run counts say 50 in Section 5.1 but 100 in the tables, and no code or data are given.\n\nWho should read it: anyone thinking about semiparametric SV might take the bivariate kernel idea as a starting point, but the paper needs major work before its claims are credible. I would not send this to a serious referee in its current form. It should go back for major revision with an honest reporting of Table 3/4, a coherent posterior derivation, and a proper baseline.","headline":"A genuine but incremental extension of the authors' NSVM work, undermined by a self-contradictory volatility comparison and an incoherent plug-in posterior.","tokens_in":13951,"tokens_out":4769,"would_cite":false,"duration_ms":44093,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Semiparametric SV model with dependent errors lowers estimation bias","keywords":["Stochastic Volatility Model","Bayesian Inference","Nonparametric Method","Markov Chain Monte Carlo","Volatility Modeling","Financial time series","kernel density estimation","dependent errors"],"falsifier":"Generate data from a bivariate normal error process with independent errors and compare the empirical coverage of NSVM-3's posterior credible intervals for the parameters and volatility against the nominal level; if the plug-in kernel step destroys posterior coherence, the coverage will drift systematically while the Gaussian model's intervals remain calibrated.","tokens_in":12932,"feed_emoji":"📈","tokens_out":4687,"duration_ms":44673,"temperature":0.7,"pith_summary":"This paper tries to establish that a stochastic volatility model in which both the return innovation and the volatility innovation are allowed to have a joint, data-estimated nonparametric distribution—rather than a prespecified Gaussian—yields lower bias and variance in estimated parameters and volatility. The claim matters because financial returns show heavy tails, skewness, and correlation between return and volatility shocks that Gaussian SV models cannot represent. The authors build NSVM-3, an extension of their earlier independent-error semiparametric models, and test it on simulated correlated Gaussian and Student-t data and on S&P 500 returns. If correct, the semiparametric SV framework offers a more flexible and adaptable default for financial econometrics.","feed_headline":"Semiparametric SV model with dependent errors lowers estimation bias","feed_subtitle":"Simulations show the new NSVM-3 model beats Gaussian and popular Bayesian baselines on parameter and volatility error.","key_machinery":"The load-bearing object is the bivariate kernel density estimate \\(\\hat{k}\\) of the joint error density of \\((u_t, \\nu_t)\\), computed with the kde2d routine on standardized residuals from an initial Gaussian SV fit. This estimate is plugged directly into the volatility posterior in place of the two Gaussian factors, giving \\(p(h_t \\mid \\ldots) \\propto $h_t^{{-3/2}}$ \\hat{k}(y_t / \\sqrt{h_t}, (\\ln h_t - \\mu_t)/\\sigma_\\nu)\\). The inverse-gamma proposal distribution is retained to drive an efficient Metropolis-Hastings sampler for the latent volatilities.","core_discovery":"The paper's central discovery is that replacing the Gaussian densities for the return error and the volatility error in the volatility posterior with a bivariate kernel density estimate of their joint distribution—estimated from residuals of a preliminary Gaussian SV fit—produces more accurate parameter and volatility estimates. In simulations with correlated Gaussian errors and with bivariate Student-t errors, NSVM-3 attains the lowest square root mean squared error for the persistence parameter, the level parameter, and the volatility-of-volatility parameter, and it generally produces lower volatility estimation error than the Gaussian SV model and a popular Bayesian implementation. The empirical application to S&P 500 daily returns shows volatility estimates that react more markedly to large price movements, suggesting the model captures real market behavior that Gaussian assumptions miss.","pith_inferences":["If the plug-in kernel step is not corrected, the reported gains may partly reflect MCMC sampling from a pseudo-posterior rather than a coherent Bayesian model; a two-stage or Dirichlet-process mixture prior would be a direct testable fix.","A natural stress test would compare out-of-sample predictive log-likelihoods on held-out periods, since in-sample parameter error does not directly measure forecasting performance.","The approach could be extended to multivariate or multi-asset volatility models, where a joint nonparametric error density would capture cross-asset tail dependence in a purely data-driven way."],"forward_implications":["Practitioners can estimate stochastic volatility models without committing to a parametric family for the error distributions while still allowing dependence between return and volatility shocks.","The method extends the semiparametric approach to settings where the error dependence structure is unknown and must be learned from data.","The simulation results quantify concrete gains in parameter bias and variance for persistence, level, and volatility-of-volatility under both Gaussian and heavy-tailed dependent errors.","The S&P 500 application demonstrates that the semiparametric volatility estimates respond more strongly to large market moves, which could inform risk management and option pricing.","The framework opens the door to fully data-driven error distributions in other state-space and latent-variable models beyond stochastic volatility."],"supporting_citations":[{"why":"Defines the base stochastic volatility model that the paper modifies and extends.","marker":"Taylor (1994)"},{"why":"Supplies the Bayesian MCMC framework and inverse-gamma proposal used to sample latent volatilities.","marker":"Doan et al. (1994)"},{"why":"Introduces the earlier semiparametric SV models (NSVM-1 and NSVM-2) that this work extends to dependent errors.","marker":"Feng and Gangopadhyay (2025)"},{"why":"Provides the popular Bayesian SV implementation used as a comparison baseline in the simulations.","marker":"Kastner (2016)"},{"why":"Supplies the kde2d bivariate kernel density estimation routine used to construct the joint error density.","marker":"Venables and Ripley (2002)"},{"why":"Motivates the need for fat-tailed and correlated error specifications in stochastic volatility models.","marker":"Jacquier et al. (2004)"}],"fun_headline_variants":["Semiparametric SV with dependent errors cuts estimation bias","Non-Gaussian errors in SV model yield lower bias and variance","Bivariate kernel errors improve volatility estimation accuracy","Better volatility estimates via non-Gaussian dependent errors","Dependent-error SV model beats Gaussian baselines in accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The posterior treats the kernel density estimate, built from residuals of a preliminary Gaussian fit on the same data, as the exact joint density of the error terms; nothing adjusts for the uncertainty in that estimate, so the MCMC samples may not come from a coherent posterior of a well-defined semiparametric model.","fun_headline_variants_meta":{"raw":{"variants":["Semiparametric SV with dependent errors cuts estimation bias","Non-Gaussian errors in SV model yield lower bias and variance","Bivariate kernel errors improve volatility estimation accuracy","Better volatility estimates via non-Gaussian dependent errors","Dependent-error SV model beats Gaussian baselines in accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000572,"raw_usage":{"total_tokens":2660,"prompt_tokens":857,"completion_tokens":1803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":1726}},"tokens_in":473,"tokens_out":1803,"duration_ms":13863,"temperature":1.0,"reasoning_tokens":1726,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:51:02.777144+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate data from a bivariate normal error process with independent errors and compare the empirical coverage of NSVM-3's posterior credible intervals for the parameters and volatility against the nominal level; if the plug-in kernel step destroys posterior coherence, the coverage will drift systematically while the Gaussian model's intervals remain calibrated.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the base stochastic volatility model that the paper modifies and extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian MCMC framework and inverse-gamma proposal used to sample latent volatilities."},{"cited_title":"and Gangopadhyay, A","cited_arxiv_id":null,"evidence_quote":"Introduces the earlier semiparametric SV models (NSVM-1 and NSVM-2) that this work extends to dependent errors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the popular Bayesian SV implementation used as a comparison baseline in the simulations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the kde2d bivariate kernel density estimation routine used to construct the joint error density."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the need for fat-tailed and correlated error specifications in stochastic volatility models."}],"review_version":1}