{"id":"af5bac9f-65a2-4f66-b6ae-d1ce94f3de30","arxiv_id":"2608.12251","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In compact neural volatility forecasters, regime information improves accuracy and training stability when it is routed through a residual mixture-of-experts gate, and degrades both when concatenated to the model input.","lead":"This paper proposes RG-ResMoE, a neural volatility forecaster that uses market regime information only to route small residual corrections, not as a direct input. On U.S. and Japanese equity panels, it reports more accurate forecasts and more stable training than capacity-matched neural baselines, and finds that appending regime variables to model inputs hurts both.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing significance test on the gate/no-gate comparison: the claimed accuracy gain from routing regime variables is only +0.0004 IC, smaller than seed noise, so 'the integration pathway is decisive' is not established.","rationale":"Reading in good faith, the paper is a controlled empirical study with real strengths: a shared MLP block, fixed hyperparameter and seed budgets, a two-market replication, and a consistent stability ordering. The strongest claim is that the integration pathway is decisive. That claim requires a positive gate-pathway effect. The reported evidence is Table 4's ΔIC = +0.0004 with no significance test, which is the least secure load-bearing point. The reader's weakest_assumption was capacity matching between MLP-L and RG-ResMoE; that is a legitimate concern, but it is not the weakest point because the gate/no-gate comparison uses the same residual architecture, so capacity cannot explain its tiny effect. Moreover, the MLP-L versus RG-ResMoE advantage is at least partly attributable to the residual design shown in Table 6, independent of regime variables. The reader did note in their rationale that the gate-specific gain is 'tiny and untested for significance,' so there is partial agreement, but the reader did not elevate it to the primary concern. The missing p-value is easy to supply if daily predictions are retained, so the overall verdict should remain CONDITIONAL rather than being upgraded or downgraded; acceptance should be contingent on either demonstrating that the gate/no-gate IC gain is statistically significant or revising the decisive-pathway claim accordingly.","tokens_in":11713,"tokens_out":9432,"duration_ms":87071,"concrete_test":"Apply the paper's own pairwise test (Eq. 9) to the daily seed-averaged IC, RMSE, and QLIKE differences between RG-ResMoE and RG-ResMoE(-z) across the 30 walk-forward windows, with Newey–West standard errors, and report the p-values. If p >= 0.05 for IC (and for RMSE/QLIKE), the abstract and Section 6.1 statements that gating regime variables 'improves accuracy' must be weakened to 'does not degrade accuracy while reducing collapsed runs,' and the decisive-pathway conclusion should be revised. A useful secondary check is to report the parameter counts for MLP-L versus RG-ResMoE to rule out the capacity confound the reader flagged, but the missing significance test is the decisive check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two halves: appending regime variables to the input hurts, and using them only in the gate improves. The first half is supported by Table 4 (MLP-L(+z) IC 0.5378 vs MLP-L 0.5421, with 30/30 collapsed seeds). The second half rests on Table 4's comparison RG-ResMoE(-z) vs RG-ResMoE, where IC moves from 0.5466 to 0.5469, a delta of only +0.0004. No Diebold–Mariano p-value is reported for this comparison, unlike every other headline comparison in Tables 3, 5, 7 and 8. The difference is an order of magnitude smaller than the input-pathway degradation and is comparable to the reported across-seed standard deviations (±0.0011 and ±0.0012). The VaR and stress-window improvements in Tables 7–8 are measured against MLP-L, not against RG-ResMoE(-z), so they cannot be attributed to the gate pathway. The stability evidence is also thin: one collapsed seed (1/30 vs 0/30). Thus the decisive-pathway conclusion overstates the positive half of the claim; the only robust gate-specific effects are absence of harm and elimination of a single collapsed run. This concern is independent of capacity matching: even if MLP-L and RG-ResMoE had identical parameter counts, the 0.0004 IC delta is untested. The paper's own pairwise testing framework in Section 4.4 (Eq. 9) could settle this directly, but it was not applied to the comparison that most needs it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks where regime information should enter a compact neural volatility forecaster. It proposes RG-ResMoE, in which a frozen base MLP predicts volatility from stock features and a soft gate uses market and idiosyncratic volatility state variables to route zero-initialized residual correction experts. The evaluation is a rolling walk-forward study on 1,027 U.S. equities with 30 seeds, plus a Japanese replication panel, comparing against persistence, GARCH, HAR, ridge, small and large MLPs, input-integration variants, hard-routing variants, and standard MoE. The headline claims are that appending regime variables to the forecasting input degrades accuracy and stability, that routing them only through the gate improves both, that soft routing beats hard routing, and that the gains concentrate in high-volatility periods.","tokens_in":12012,"tokens_out":3110,"duration_ms":28979,"significance":"If the central claim is established, the paper offers a concrete and transferable design principle for regime-aware forecasting in compact models, and it cleanly separates capacity effects from integration-pathway effects. The experimental protocol is a genuine strength: rolling non-overlapping test windows, 30 random seeds, validation-based hyperparameter selection, Diebold-Mariano tests with Newey-West standard errors for the main comparisons, and an independent cross-market replication. The residual, zero-initialized MoE formulation is well motivated and the stability results are unusually clean. However, the headline 'integration pathway is decisive' claim currently rests on a statistically untested difference of +0.0004 in IC between RG-ResMoE and RG-ResMoE(-z), and the capacity-matching assertion is unverified because no parameter counts are reported. These gaps are load-bearing for the paper's central message, so the manuscript requires major revision rather than acceptance in its present form.","major_comments":[{"comment":"The gate-pathway comparison RG-ResMoE versus RG-ResMoE(-z) is the single most load-bearing comparison in the paper, yet no Diebold-Mariano p-value is reported for it. The IC difference is +0.0004, which is smaller than the reported across-seed standard deviations (±0.0011 and ±0.0012) and an order of magnitude smaller than the input-pathway degradation (-0.0043 for MLP-L vs MLP-L(+z)). The paper's own testing framework in Section 4.4, Eq. (9), should be applied to this comparison in exactly the same way it is applied to Tables 3, 5, 7, and 8. Without that test, the claim that 'restricting them to the routing gate improves accuracy' is not supported; at best the evidence shows absence of harm plus elimination of one collapsed seed (0/30 vs 1/30). The same absence of pairwise significance testing applies to the input-pathway deltas in Table 4: they are large and suggestive, but no p-values are reported.","section":"§6.1, Table 4"},{"comment":"The manuscript asserts that MLP-L with hidden width H=44 'matches the overall forecasting capacity of the full RG-ResMoE model,' but it reports no parameter counts for any model. This is load-bearing because the paper's conclusion that RG-ResMoE's gains come from the residual expert architecture 'rather than from additional parameters' depends on equality of parameter budgets. If MLP-L actually has fewer parameters, or if the optimization landscape differs due to width, the architecture-vs-capacity attribution is undermined. Please report parameter counts for every neural model in Table 2 (or at least MLP-S, MLP-L, RG-ResMoE, and the MoE variants), and confirm that the counts are equal or state the actual discrepancy.","section":"§4.2, 'capacity-matched' claim"},{"comment":"The VaR calibration improvements (Table 7) and the stress-window IC gains (Table 8) are computed relative to MLP-L or MLP-S, not relative to RG-ResMoE(-z). These comparisons therefore reflect the combined effect of the residual architecture, the gate, and any capacity differences; they cannot be used to attribute the gains to the regime-gating pathway. The abstract's statement that 'restricting them to the routing gate improves accuracy and Value-at-Risk calibration' overreaches the evidence. Please either rerun the VaR and stress analyses for RG-ResMoE(-z), or explicitly rephrase the claim so that these results are attributed to the RG-ResMoE architecture as a whole, not to the gate pathway alone.","section":"§7, Tables 7 and 8"}],"minor_comments":[{"comment":"Figure 2 plots per-window ΔIC and ΔRMSE for the input and gate pathways, but it provides no confidence bands, error bars, or significance information. Given that the aggregate gate-pathway effect is tiny, the visual impression of 'stability' should be backed by a per-window statistical summary or at least by reporting the across-window mean and standard error of the gate-pathway deltas.","section":"§6.1, Figure 2"},{"comment":"The description of MLP-S as 'matching an individual block in RG-ResMoE' is ambiguous, since RG-ResMoE contains a base block plus four expert blocks. Please clarify whether MLP-S matches the width of a single block but uses only one block, and how its parameter count compares to one expert block.","section":"§4.2, MLP-S sizing"},{"comment":"Table 4 mixes models with and without the `ref.` marker, which is slightly confusing because some rows are reference rows and others are comparison rows. Consider adding a separate column for p-values (which is currently missing) and clearly labeling all rows that are not tested against a reference.","section":"Table 4, layout"},{"comment":"The Japanese replication paragraph reports p-values for IC, RMSE, and QLIKE for RG-ResMoE vs MLP-L, but not for the RG-ResMoE vs RG-ResMoE(-z) comparison, which would parallel the main-market weakness. Please add the gate-pathway comparison for Japan as well.","section":"§8, replication paragraph"},{"comment":"The collapse threshold is described as 'a prespecified diagnostic threshold' but the actual value (mean test QLIKE > 2.0) appears only in the caption of Figure 4. Please state it explicitly in the protocol in Section 4.3.","section":"§4.3, collapse threshold"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is a genuinely careful empirical study, but the headline overclaims. The controlled comparison of integration pathways is the real contribution, and the negative result is solid: appending regime variables to the input drops IC from 0.5421 to 0.5378 for MLP-L and collapses all 30 seeds. The residual MoE with frozen base and zero-initialized experts is also a real stability win—0/30 collapses versus 24/30 for standard MoE—and that effect is statistically meaningful. Soft routing beating hard routing is clean and reproduced on Japan. The references are appropriate, and the related work is honest about the known components.\n\nThe soft spots are real but localized. The positive half of the 'pathway is decisive' claim rests on RG-ResMoE with gate versus RG-ResMoE without gate: +0.0004 IC, no Diebold-Mariano test reported, and the difference is inside the across-seed standard deviations. That is not evidence of a decisive gate contribution; it is absence of harm plus one fewer collapsed seed. The paper's own Eq. 9 could have tested this directly, and the fact that it was not applied here is conspicuous. Second, the capacity match to MLP-L is asserted via width 44, but no parameter counts are given, leaving a hole in the 'architecture, not capacity' argument. Third, the U.S. panel is survivorship-restricted to stocks listed through the end, and that is not discussed. Finally, no code or data is released, so the numbers cannot be checked.\n\nFor the right reader—someone designing compact regime-aware forecasters—the negative input-integration result and the residual-stability finding are worth citing. The 'decisive pathway' conclusion should be downgraded unless the gate/no-gate comparison gets a proper test. I would send this to peer review and ask for that test, parameter counts, and a disclosure of the universe filter. With those, the paper would be a solid contribution.","headline":"A carefully controlled empirical study whose robust negative result on input-appended regime variables is solid, but the 'decisive pathway' claim rests on a tiny untested gain.","tokens_in":12567,"tokens_out":4042,"would_cite":true,"duration_ms":33963,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"In compact neural volatility forecasters, regime information should enter through the routing gate, not the forecasting input.","keywords":["volatility forecasting","mixture of experts","regime switching","residual learning","gated routing","training stability","Value-at-Risk","cross-sectional equity"],"falsifier":"Count the trainable parameters of the large MLP and the proposed model under the reported widths; if the MLP has materially fewer parameters, retrain with parameter-matched widths and check whether the IC gap and collapse rates persist. The central claim would be falsified if a parameter-matched MLP with regime variables appended to the input matches or beats the proposed model on IC, RMSE, and collapsed seeds.","tokens_in":11460,"feed_emoji":"📈","tokens_out":9053,"duration_ms":73853,"temperature":0.7,"pith_summary":"This paper asks where regime information should enter a neural network that forecasts five-day stock volatility. Using a rolling walk-forward evaluation on 1,027 U.S. equities, with information, model capacity, hyperparameters, and random seeds matched across architectures, it proposes RG-ResMoE: a frozen base network predicts volatility from stock features, small zero-initialized experts learn residual corrections, and a soft gate uses regime variables (market volatility and idiosyncratic volatility) only to route those corrections. The central finding is that the integration pathway is decisive: the same regime variables improve accuracy and training stability when they enter through the routing gate, but degrade both when appended directly to the forecasting input of a plain feed-forward network (MLP). Soft routing beats hard routing, and the gains concentrate in high-volatility periods, with the ordering reproduced on a Japanese equity panel. A sympathetic reader would take away a design principle: in compact neural volatility forecasters, regime information should modulate how residual predictions are combined rather than directly determine the forecast.","feed_headline":"Regime data gates experts, not inputs, in volatility forecasts","feed_subtitle":"The same market-state variables help volatility models when routed through a gate and hurt when appended to inputs.","key_machinery":"The load-bearing object is the regime-gated residual mixture-of-experts layer. A base network $\\text{Block}(x;\\theta_b)$, trained first and frozen, produces the level forecast $\\hat{y}_{\\text{base}}$; $K$ zero-initialized experts each output residual corrections $r_k(x)$; and a soft gate $\\pi = \\operatorname{softmax}(g(x,z))$ combines them as $\\hat{y} = \\hat{y}_{\\text{base}} + \\sum_k \\pi_k r_k(x)$. Regime variables $z$ appear only inside the gate, never in the forecasting blocks. Two regularizers carry the argument: a shrinkage penalty $\\alpha\\left(\\sum_k \\pi_k r_k(x)\\right)^2$ that keeps corrections close to the base forecast, and a load-balancing penalty $\\lambda_{\\text{LB}} \\sum_k (\\bar{\\pi}_k - 1/K)^2$ that prevents routing collapse. This separation of the base from the corrections is what allows the controlled comparison: the same block, capacity, and data are used everywhere, with only the integration pathway of $z$ varied.","core_discovery":"RG-ResMoE consistently outperforms a capacity-matched MLP on every reported forecast metric, and the paper argues the advantage is architectural rather than merely additive capacity. It reports a daily cross-sectional rank correlation (IC) of 0.5469, RMSE of 0.2304, out-of-sample $R^2$ of 0.292, and a quasi-likelihood loss of 0.735, with zero collapsed training runs out of 30; the large MLP scores 0.5421, 0.2320, 0.282, and 2.23 with 3 collapsed runs. Appending the same two regime variables directly to the MLP input lowers its IC to 0.5378 and collapses all 30 seeds, while gate-only use raises the proposed model's IC slightly and leaves zero collapses. All hard-routing variants, including learned top-1 gates and hand-crafted volatility partitions, underperform the soft gate by 0.0021 to 0.0034 in IC, and a standard mixture-of-experts without the frozen base collapses in 24 of 30 runs. The paper concludes that in compact forecasters the value of mixture-of-experts lies less in increasing capacity than in controlling how nonstationary regime information influences prediction.","pith_inferences":["I infer the gate is learning a smooth, state-conditional interpolation among correction functions rather than selecting discrete regimes; a natural test is to constrain the gate to a small number of interpretable regime probabilities and check whether accuracy is retained.","The mechanism may generalize beyond volatility: any auxiliary nonstationary signal that destabilizes direct input integration could be moved into the routing gate, though this is not tested in the paper.","Because the paper reports no parameter counts for the large comparison MLP versus the proposed model, the architecture-versus-capacity attribution would be sharpest if replicated with exact parameter-count-matched baselines.","An untested extension is whether the same pathway principle holds for longer-context sequence models or for forecasts other than five-day realized volatility, which the paper explicitly leaves for future work."],"forward_implications":["Regime variables should be routed through a gating network rather than concatenated to the predictor inputs in compact neural volatility models.","Soft, continuous expert weighting dominates discrete regime assignment; hand-crafted partitions based on volatility quantiles, sectors, or market-by-idiosyncratic splits do not recover soft-gate performance.","The frozen base forecast and zero-initialized residual corrections are the main source of training stability; increasing expert count from two to six provides no additional accuracy and hurts stability.","The accuracy advantage over the large MLP is concentrated in elevated-volatility periods, with about 4.3 times and 6.7 times the full-sample gain in the top volatility decile and the COVID crisis, translating into better Value-at-Risk coverage.","The same ordering of results holds on an independent Japanese equity panel, suggesting the pathway principle is not an artifact of the U.S. sample."],"supporting_citations":[{"why":"Supplies the per-stock GARCH(1,1) baseline that the pooled neural models are compared against.","marker":"[3]"},{"why":"Supplies the heterogeneous autoregressive realized-volatility baseline that defines the HAR comparator.","marker":"[4]"},{"why":"Provides the predictive-accuracy test used for the paper's pairwise model comparisons.","marker":"[5]"},{"why":"Supplies the residual-learning principle behind the frozen base network and expert corrections.","marker":"[11]"},{"why":"Provides the original mixture-of-experts formulation with a learned gating network.","marker":"[12]"},{"why":"Supplies the unconditional-coverage test used to evaluate Value-at-Risk calibration.","marker":"[13]"},{"why":"Provides the autocorrelation-robust standard errors used for the overlapping five-day forecast target.","marker":"[16]"},{"why":"Supplies the quasi-likelihood loss used to compare variance forecasts.","marker":"[17]"},{"why":"Supports the zero-initialized residual branches that anchor the model to the base forecast at initialization.","marker":"[23]"},{"why":"Supports warm-starting and staying close to a stable initialization, which underlies the frozen base design.","marker":"[1]"}],"fun_headline_variants":["Route regime data, don't append it, for volatility","Regime info gates experts, not inputs, for forecasts","Gated regime experts beat capacity-matched MLP","Regime input placement decides volatility forecast quality","Regime routing beats regime feeding in volatility models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the large comparison network truly has the same number of parameters as the proposed model; the paper matches hidden-layer widths but reports no parameter counts, so if the comparison network is actually smaller, the conclusion that the architecture rather than extra capacity drives the gains would not follow.","fun_headline_variants_meta":{"raw":{"variants":["Route regime data, don't append it, for volatility","Regime info gates experts, not inputs, for forecasts","Gated regime experts beat capacity-matched MLP","Regime input placement decides volatility forecast quality","Regime routing beats regime feeding in volatility models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000445,"raw_usage":{"total_tokens":2296,"prompt_tokens":1040,"completion_tokens":1256,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":1182}},"tokens_in":656,"tokens_out":1256,"duration_ms":11026,"temperature":1.0,"reasoning_tokens":1182,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:11:19.223340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count the trainable parameters of the large MLP and the proposed model under the reported widths; if the MLP has materially fewer parameters, retrain with parameter-matched widths and check whether the IC gap and collapse rates persist. The central claim would be falsified if a parameter-matched MLP with regime variables appended to the input matches or beats the proposed model on IC, RMSE, and collapsed seeds.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the per-stock GARCH(1,1) baseline that the pooled neural models are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the heterogeneous autoregressive realized-volatility baseline that defines the HAR comparator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the predictive-accuracy test used for the paper's pairwise model comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the autocorrelation-robust standard errors used for the overlapping five-day forecast target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the quasi-likelihood loss used to compare variance forecasts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports warm-starting and staying close to a stable initialization, which underlies the frozen base design."}],"review_version":1}