{"id":"c97c344d-17ba-4535-8f04-407feee956ae","arxiv_id":"2412.06193","paper_version":2,"verdict":"REJECT","confidence":"LOW","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Using a fitted CAViaR quantile model, the paper reports that US high-yield credit market tail risk spills over to stocks, FX and interbank markets more strongly than the reverse, making credit the central warning market.","lead":"This preprint applies a machine-learning version of a standard tail-risk model to US stock, credit, foreign exchange and interbank data, and concludes that the credit bond market is the main source of extreme risk spillovers. A general reader might care because the claim, if true, suggests that monitoring high-yield credit markets could give early warnings of US stock market stress.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'early-warning' claim is not established: all coefficients and Figure 5 pseudo-IRFs are in-sample full-sample fits, with no out-of-sample forecast evaluation or stability test, so 'historical extreme risks predict other markets' VaR' remains unsupported.","rationale":"The reader's weakest_assumption identifies the same core issue: no out-of-sample period, no structural-break test, and no forecasting evaluation, with HIBOR as an additional proxy concern. I agree and would put the emphasis on the unsupported predictive claim, because the abstract's 'can serve as a predictor' is the strongest version of the central claim. This is a correctness risk rather than a stylistic or reporting issue: even if the coefficient estimates in Table 3 are exactly right in-sample, they do not establish that RCB tail information improves VaR forecasts for other markets. The paper provides no benchmark comparison, no forecast evaluation, and no stability check. A forecasting exercise with rolling windows would directly settle whether the credit market's historical tail-risk information has predictive content beyond the autoregressive dynamics already in each market's own VaR. Since this concern supports the reader's REJECT verdict rather than redirecting it, the verdict should remain unchanged.","tokens_in":8208,"tokens_out":4052,"duration_ms":43012,"concrete_test":"Run a rolling-window out-of-sample exercise: for each t in, say, 2018-2024, estimate Eqs (1)-(2) using only data up to t, generate 1-day-ahead 1% VaR forecasts for RS, RE, and RM with the unrestricted MVMQ-CAViaR model (including lagged RCB tail terms) and with a restricted model that excludes the RCB spillover terms. Compare quantile losses, expected shortfall, and Kupiec/Christoffersen conditional coverage. The central claim survives only if the model with RCB information has superior out-of-sample predictive accuracy; otherwise the credit market's 'warning' status is not demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and conclusions claim RCB's tail-risk history can predict VaR of other US markets and that RCB has a central warning position. The only evidence offered is Section III.C: significant cross-coefficients in Table 3 and the pseudo-impulse responses in Figure 5. Both are computed from one full-sample estimation over 2014-2024 (Table 3, Eqs 1-2), with no holdout, no rolling window, and no structural-break test. Pseudo-impulse responses are not defined in the text; whatever they trace, they are in-sample summaries, not predictive validation. In nonlinear quantile autoregressions, in-sample fit and significant coefficients need not translate into out-of-sample predictive gains; the autoregressive terms (b11, b22) may carry all forecastable information, making RCB's marginal predictive content negligible. The paper also reports no comparison against a univariate CAViaR benchmark, no forecast-encompassing test, and no evaluation of VaR exceedances. Consequently, the directional/asymmetric spillover pattern may be real as a historical description, but the 'early warning' and 'predictor' components of the central claim do not follow from the evidence. Using HIBOR as a US interbank proxy is an additional validity threat, but the missing out-of-sample evidence is sufficient to block the claim as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript estimates bivariate multivariate-quantile CAViaR systems at the 1% quantile for three pairs of markets—credit bonds with stocks, with foreign exchange, and with interbank lending—using daily data from May 2014 to June 2024. Parameters are estimated by gradient descent and genetic algorithm, and the paper uses joint significance tests and a figure called \"pseudo-impulse response\" to claim that the credit bond market is the dominant source of tail-risk spillovers and that its historical tail-risk information can predict other markets' VaR. The central claims are that the credit market occupies a \"central warning position\" and that its extreme-risk history serves as a predictor for stocks, FX, and interbank markets.","tokens_in":8462,"tokens_out":7715,"duration_ms":73382,"significance":"The question of which US market leads tail-risk transmission is practically important for systemic-risk monitoring, and the MVMQ-CAViaR framework is a reasonable starting point for studying quantile spillovers. The paper's merits include the use of a recognized conditional autoregressive quantile model, a long daily sample, and a transparent comparison of two optimization approaches. However, the significance of the contribution depends on the early-warning claim, and that claim is not established by the evidence as presented. The missing out-of-sample validation, the undefined pseudo-impulse-response procedure, and the apparent misreporting of significance tests are all load-bearing problems.","major_comments":[{"comment":"The pseudo-impulse response procedure used for Figure 5 is never defined. The paper does not specify how a one-unit shock to one market's VaR is imposed in the nonlinear MVMQ-CAViaR system, how many horizons are traced, whether coefficients are held fixed, or how the duration comparison is measured. Because the conclusion that the credit market's spillover effect and duration are largest is read directly from this figure, the missing definition is a load-bearing gap.","section":"§III.C, Figure 5"},{"comment":"The claim that historical extreme-risk information from the credit market can 'serve as a predictor' of other markets' VaR is not supported by any out-of-sample exercise. All coefficients in Equations (1)-(2) and Table 3 are estimated on the full 2014-2024 sample, and the joint tests and Figure 5 are in-sample summaries. No holdout period, rolling-window forecast, forecast-encompassing test, or VaR exceedance backtest is reported. Given the autoregressive terms b11 and b22, significant cross-coefficients do not by themselves imply predictive gains over a univariate CAViaR benchmark.","section":"Abstract and §IV"},{"comment":"The column heading 'Accept/Refuse' and the entry 'refuse' are incorrect for p-values of 0.0619, 0.4906, and 0.4954. At the conventional 5% level these null hypotheses cannot be rejected, and the latter two cannot be rejected at the 10% level either. This is not a labeling detail: the directional conclusion that FX and interbank markets are spillover receivers in Models 1 and 3 depends on rejecting a12=b12=0, while the corresponding a21=b21=0 rows actually show no significant reverse spillover.","section":"Table 5"},{"comment":"The text states that in Model 2 (stock/credit) b12 and b21 are both significant at the 5% level. The displayed estimate for b21 is -0.0103 with a standard error of 0.0117, which gives a t-ratio of about -0.88 and is not significant. The asterisk alignment in Table 3 is also unclear, with parentheses and stars seemingly attached to the wrong coefficients in several rows, so the significance claims in this table cannot be verified as printed.","section":"Table 3 and §III.B.2"},{"comment":"HIBOR, a Hong Kong interbank rate, is used as the proxy for the US interbank market. Since the paper's stated object is US financial markets, the interbank-market results and Figure 5's claim about 'the other two markets' are only valid if HIBOR tracks US interbank conditions; no justification or sensitivity analysis with a US-based rate (for example, SOFR or the TED spread) is provided.","section":"Table 1 and §III.B.1"}],"minor_comments":[{"comment":"The subsection labels (a)-(c) are inconsistent with Table 3: the text says 'Model 3 reports risk spillovers in the foreign exchange market and credit bond market,' but Table 3 lists Model 3 as interbank/credit and Model 1 as FX/credit, which makes the estimation results hard to follow.","section":"§III.B.2"},{"comment":"The quantile regression objective is written without the usual check function and the notation is garbled; please restate it as min over beta of (1/n) times the sum of rho_k(r_t - f_t(beta)).","section":"Equation (7)"},{"comment":"The comparison between genetic algorithm and gradient descent is informal; no hyperparameters (mutation rate, crossover rate, learning-rate schedule) or convergence criteria are reported, and it is unclear whether the loss curves shown are the same objective function used in the CAViaR estimation.","section":"§III.B.3, Figures 2-4"},{"comment":"Several citations do not correspond to the claimed content (for example, [12] is cited for credit risk in China but is a deep-network pattern recognition paper), and the reference list includes items never discussed in the text, such as [9] and [16].","section":"References"},{"comment":"The table title reads 'MCMQ-CAViaR' but the model is elsewhere called MVMQ-CAViaR; the typo should be corrected.","section":"Table 3 title"}],"recommendation":"reject","confidential_remarks":"This is an early preprint with the core empirical claim not yet supported. The missing out-of-sample validation, the undefined pseudo-impulse-response procedure, and the incorrect significance labels are substantial, and they affect the central early-warning conclusion rather than being presentation issues. I do not see a path to acceptance without essentially redoing the empirical design. If the authors add a genuine out-of-sample predictive exercise, define the pseudo-IRF, correct the statistical interpretation, and address the HIBOR proxy, a future version could be worth re-reviewing. The paper would also benefit from a data/code availability statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper applies a known MVMQ-CAViaR framework to US data and claims the high-yield credit market is the dominant tail-risk transmitter to stocks, FX and interbank markets. That empirical finding is new and, if true, worth knowing. But the evidence as presented does not establish the \"early warning\" or \"predictor\" part of the claim.\n\nWhat it does well: the joint tests in Tables 4 and 5 are the right idea, the descriptive stats and ADF tests are standard, and the comparison between genetic algorithm and gradient descent is at least a real exercise. The direction of the finding—credit spreads leading other markets in the tail—is plausible and consistent with some earlier work on China.\n\nSoft spots, in order of severity. First, the predictive claim is in-sample only. All coefficients and the pseudo-IRFs come from a single full-sample fit over 2014–2024. There is no holdout, no rolling window, no structural-break test, and no comparison to a univariate CAViaR benchmark. In nonlinear quantile autoregressions, significant cross-coefficients can disappear out-of-sample, so the abstract's statement that \"historical information can serve as a predictor\" is unsupported. Second, the paper never defines the pseudo-impulse response procedure; Figure 5 is the main evidence for the credit-market centrality and its duration, and a reader cannot tell what is being traced. Third, Table 5 labels p-values above 0.49 as \"refuse\", which is backwards: with p = 0.49 you cannot reject the null of no spillover. That directly undercuts the claim that the credit market is less affected by FX and interbank spillovers—those tests actually fail to reject, possibly due to low power. Fourth, HIBOR is a Hong Kong rate, not a US interbank rate; using it as a proxy for the US interbank market is a validity threat. Fifth, the text is heavily corrupted (OCR-like), coefficient stars do not match the text's claims, and the reference list contains several irrelevant entries, making verification difficult.\n\nThis is not a paper to desk-reject on the idea. The empirical claim is important enough that a cleaned version with out-of-sample validation and a properly defined IRF could make a contribution. But as it stands, the evidence does not support the headline. I would send it back for major revision if I were the editor, or reject and invite resubmission.","headline":"Applies a known MVMQ-CAViaR framework to US data and finds credit-market tail-risk centrality, but the predictive claim rests entirely on in-sample fits and an undefined pseudo-IRF.","tokens_in":9020,"tokens_out":2565,"would_cite":false,"duration_ms":24206,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The US credit bond market is the main transmitter of tail risk to other US markets, with larger and longer-lasting spillovers than any reverse effect.","keywords":["tail risk spillover","CAViaR","credit bond market","early warning","genetic algorithm","gradient descent","Value-at-Risk","quantile regression"],"falsifier":"Re-estimate the model on data through 2019 and generate one-step-ahead $1\\%$ VaR forecasts for the S&P 500 for 2020–2024 from the credit-market equations; if these forecasts do not beat a univariate CAViaR benchmark on quantile scores or hit rates, the credit-market-centrality claim loses its predictive content.","tokens_in":7939,"feed_emoji":"📉","tokens_out":8539,"duration_ms":66782,"temperature":0.7,"pith_summary":"This paper aims to establish that the US credit bond market, not the stock market, is the main origin of extreme (tail) risk for the other major US financial markets. Using a multivariate CAViaR model that models the $1\\%$ Value-at-Risk of each market, the paper estimates how a tail event in one market shifts the tail risk of another. It reports that credit-market tail shocks spill into the stock market with greater size and persistence than any spillover into the credit market, while foreign-exchange and interbank shocks barely affect credit. If this asymmetry is correct, monitoring high-yield credit spreads could provide an early warning of stock-market crashes and improve systemic risk management.","feed_headline":"Credit-market tail risk leads stocks, not the reverse","feed_subtitle":"A CAViaR model with genetic-algorithm optimization makes the high-yield bond market the center of systemic risk warnings","key_machinery":"The main object is the MVMQ-CAViaR model (multivariate multi-quantile conditional autoregressive Value-at-Risk): a two-equation system in which the $1\\%$ VaR of each market depends on its own lagged VaR, the other market's lagged VaR, and lagged shocks. The off-diagonal coefficients $b_{12}/b_{21}$ and $a_{12}/a_{21}$ measure directional tail-risk spillovers, and joint tests on those coefficients decide whether a market transmits, receives, or both. Model parameters are estimated by regression quantiles with gradient descent and genetic algorithm optimization, with the genetic algorithm reported as the better optimizer.","core_discovery":"The central claim is that the credit bond market is a systemic risk hub: its extreme risk events transmit to the stock, foreign exchange, and interbank markets, while only stock-market tail risk feeds back into credit. This is established through the MVMQ-CAViaR estimates, joint hypothesis tests on the spillover coefficients, and pseudo-impulse response functions, which together show the credit-to-stock spillover is larger and longer-lived than the stock-to-credit spillover. The paper states that historical extreme-risk information from the credit bond market can therefore serve as a predictor of the Value-at-Risk of other markets and places credit in a central warning position.","pith_inferences":["The paper's pseudo-impulse responses are computed in-sample; a true out-of-sample forecasting exercise would be needed to turn the credit-market-centrality result into a usable early-warning system.","The reported asymmetry is consistent with structural credit-risk models in which default risk drives equity values, but the paper does not test that underlying mechanism.","Using HIBOR, a Hong Kong rate, as the US interbank proxy could weaken the interbank results; substituting a US-based rate such as SOFR might change the estimated spillover network.","If the asymmetry holds, it implies an asymmetric information flow in which credit markets aggregate distress information earlier than equity markets, a prediction that could be tested with Granger causality in tail quantiles."],"forward_implications":["Regulators and investors should treat the US high-yield credit market as a leading indicator for equity tail risk.","Stock portfolio risk models should include lagged credit-spread quantiles as conditioning variables.","Tail shocks from the foreign-exchange and interbank markets have comparatively little effect on credit-bond risk, so early-warning systems can focus on credit and equity.","The significant two-way spillover between stocks and credit implies that tail-risk hedging across these two markets is more valuable than hedging credit against FX or interbank risk."],"supporting_citations":[{"why":"Foundational CAViaR model: specifies the conditional quantile as an autoregressive process, which the paper's MVMQ-CAViaR extends to multiple markets.","marker":"[21]"},{"why":"Cited for the quasi-maximum likelihood estimation method used to estimate the MVMQ-CAViaR parameters.","marker":"[23-24]"},{"why":"Supplies the regression-quantile estimator of Koenker and Bassett that the paper compares with gradient descent and genetic algorithm optimization.","marker":"[25]"},{"why":"Justifies using the ICE BofA US High Yield Index Effective Yield as the credit-market proxy because high-yield bonds price most tail risk.","marker":"[29-31]"},{"why":"The paper invokes this marker as prior evidence that credit-bond tail risk can warn of stock-market volatility, the idea the paper extends to US markets.","marker":"[12]"},{"why":"Documents the fat-tailed, peaked distributions that motivate focusing on tail quantiles rather than means or volatility.","marker":"[32]"}],"fun_headline_variants":["Credit market drives stock tail risk, study finds","Credit, not stocks, is the systemic risk hub","Tail risk flows from credit to stocks, CAViaR shows","Bond market tail risk predicts stock market VaR","ML pinpoints credit as the risk epicenter"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central result assumes that the spillover coefficients estimated on the full 2014–2024 sample are stable and capture genuine one-way causation, so that the fitted model's impulse responses describe how future tail risks will actually propagate; no out-of-sample forecasting evaluation is reported.","fun_headline_variants_meta":{"raw":{"variants":["Credit market drives stock tail risk, study finds","Credit, not stocks, is the systemic risk hub","Tail risk flows from credit to stocks, CAViaR shows","Bond market tail risk predicts stock market VaR","ML pinpoints credit as the risk epicenter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1181,"prompt_tokens":855,"completion_tokens":326,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":250}},"tokens_in":471,"tokens_out":326,"duration_ms":3768,"temperature":1.0,"reasoning_tokens":250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:54:27.942858+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-estimate the model on data through 2019 and generate one-step-ahead $1\\%$ VaR forecasts for the S&P 500 for 2020–2024 from the credit-market equations; if these forecasts do not beat a univariate CAViaR benchmark on quantile scores or hit rates, the credit-market-centrality claim loses its predictive content.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Foundational CAViaR model: specifies the conditional quantile as an autoregressive process, which the paper's MVMQ-CAViaR extends to multiple markets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the regression-quantile estimator of Koenker and Bassett that the paper compares with gradient descent and genetic algorithm optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The paper invokes this marker as prior evidence that credit-bond tail risk can warn of stock-market volatility, the idea the paper extends to US markets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the fat-tailed, peaked distributions that motivate focusing on tail quantiles rather than means or volatility."}],"review_version":1}