{"id":"57253529-67f5-48ed-9a3b-38516f4d263c","arxiv_id":"2608.07349","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A typed residual algebra with field-preserving local correction, one identity-erasing aggregation, and fresh-residual closure improves a frozen stock-return prediction pipeline by 5.58 pp net return.","lead":"This paper introduces a representation-learning framework, residual algebra, that keeps track of which representation produced which prediction error, and applies it to stock return prediction. On Chinese A-share data, the proposed FPRC-PQ architecture raises net-of-cost return from 13.52% to 19.10% and Sharpe from 1.42 to 2.09 versus a simple average of three factor fields.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 5.2's audits select (9)–(11) using 2023–2026 outcomes, so the headline gains are post-selection; the frozen-protocol claim is not credible and the 5.58pp/0.67 Sharpe advantage may be optimistic.","rationale":"The reader's weakest assumption was that no 2023–2026 label influenced any architectural choice, and I agree this is the load-bearing point. The paper's own Section 5.2 and Table 2 describe audits run on the strict book that selected structures in (9)–(11). Even with preregistered decision rules, using the evaluation window to choose the architecture makes the headline comparisons conditional on the test data. The reported bootstrap intervals are computed for the selected model only and do not account for selection, so the 5.58 pp improvement and 0.67 Sharpe gain are not independent evidence for the residual-ownership explanation. I considered other possible concerns—such as the absence of released code or the idealization in the control-variate argument—but those are secondary verification or exposition issues rather than threats to the central causal claim. The paper's algebra appears internally consistent and the controls are thoughtful, but the empirical attribution depends on the frozen-protocol guarantee, and that guarantee is in direct tension with the audit procedure. The appropriate verdict remains conditional: the paper should be accepted only if the authors can show the audit decisions were fixed before seeing 2023–2026 outcomes or can produce a nested holdout evaluation that confirms the selected architecture's advantage on data untouched by selection.","tokens_in":17038,"tokens_out":6766,"duration_ms":66819,"concrete_test":"Hold out the 2025–2026 portion of the strict book (or, failing that, 2026). Run audits A1–A5 using only the 2023–2024 window to select the architecture and freeze it. Then evaluate the selected architecture on the untouched 2025–2026 window against the pooled-field baseline and the matched-capacity direct control. If the selected architecture's net-return delta on untouched data is not close to the reported 5.58 pp (or its bootstrap CI includes zero), the headline gain is at least partly a selection artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the 5.58 pp net return and 0.67 Sharpe gain over the pooled-field baseline come from residual ownership and composition, not from features or capacity. This requires that the architecture (9)–(11) was fixed before any 2023–2026 label was seen. Section 4 states that 'no 2023–2026 label influenced any choice' and Section 3.2 calls the audits 'preregistered,' but Section 5.2 and Table 2 report five audits whose key contrasts are evaluated on the strict 2023–2026 book and whose results 'selected (9)-(11) over its neighbors' (e.g., A1, A4, A5). Even if each decision rule was fixed in advance, applying those rules to evaluation-window outcomes makes every subsequent comparison—including the headline FPRC-PQ versus pooled fields and versus matched-capacity direct—conditional on the test set. The bootstrap CIs and p-values do not account for this selection step, so they overstate confidence. Thus the strongest claim rests on an assumption that the paper itself appears to violate unless the audit decisions can be shown to have been made without consulting 2023–2026 returns.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces 'residual algebra' as a representation-learning principle: each Fold field is a 10x10 point-in-time conditional-mean table that owns both a coordinate system and the residual it leaves; FPRC-PQ composes a typed relaxation, an equal-mean aggregation boundary, and a fresh-residual shared closer. The authors derive an exact telescoping decomposition, a control-variate interpretation, a Neyman-style coupled-refit orthogonality, and an analytical 'reflective rumination' extension (explicitly not evaluated). On 3.67M Chinese A-share stock-day rows (2023-2026) under a claimed frozen protocol, FPRC-PQ earns 19.10% net return and Sharpe 2.09 versus 13.52% and 1.42 for the pooled-field baseline, with matched-capacity, unified-residual, identity-free, pairwise-only, and side-information controls trailing. The central empirical claim is that the gain is caused by explicit residual ownership and ordered composition, not by more features or more trees.","tokens_in":17392,"tokens_out":7289,"duration_ms":67470,"significance":"Conditional on the frozen-protocol claim being true, the paper is a valuable and unusually disciplined contribution: the algebraic identities are clean and exactly stated; the matched-capacity direct control, the unified-residual identity control, and the eight side-information channels are the right kinds of falsifiable contrasts; paired 21-day block-bootstrap inference is appropriate; the reflective rumination is honestly scoped as not part of the empirical headline; and the bit-exact reproduction audit is a strong reproducibility promise. The main significance risk is that the architecture-selection audits in §5.2 are reported on the 2023-2026 evaluation window, which, if they were used to choose the architecture, makes the headline comparisons post-selection and undermines the causal attribution. The contribution's value therefore depends on resolving that contradiction.","major_comments":[{"comment":"The manuscript asserts in §4 that 'no 2023–2026 label influenced any choice' and in §3.2 that five preregistered audits selected (9)–(11), but Table 2 reports the audit contrasts as paired block-bootstrap deltas on the strict 2023–2026 book (e.g., A1: typed local-PQ vs. pooled fields +4.78 pp; A4: local-only vs. dual-layer Q +0.30 pp; A5: self-state vs. empty −0.45 pp). These realized-window outcomes are used to select the architecture, so the headline comparison FPRC-PQ versus pooled fields and versus matched-capacity direct is post-selection. The bootstrap CIs in Tables 1 and 9 do not condition on this selection and therefore overstate confidence. This directly undermines the paper's strongest claim in the Abstract and §5.1 that the gain is due to residual ownership rather than features or capacity. Please either (i) show that every A1–A5 decision was made from data through 2022-12-31 alone, with the evaluation-window tables provided only as retrospective illustrations, or (ii) treat the audits as exploratory and re-estimate the headline comparisons under a genuinely fixed architecture with selection-adjusted inference.","section":"§5.2, Table 2; §3.2; §4"}],"minor_comments":[{"comment":"Several numeric columns run together in the printed table (for example, '5.012.0921.67' and '2.5719.10'), making the row values difficult to read; the table should be reformatted with explicit column spacing.","section":"Table 1"},{"comment":"The notation bE_PIT and the purge-gap label availability 'through t−h−1' are used before being defined; please define these terms at first use in Section 2 rather than only in the appendix.","section":"§2, Eq. (2)"},{"comment":"The term 'strict-OOF modular-FPRC experts' is introduced without definition; please clarify the out-of-fold construction used to score all 1,330 candidate triples.","section":"Appendix A"},{"comment":"The reported Pr[Δ>0] values are one-sided probabilities; please state explicitly which direction each probability refers to so that a reader cannot mistake them for two-sided p-values.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's frozen-protocol narrative is central to its contribution, but the Table 2 audits appear to use the evaluation window for architecture selection. If the authors cannot demonstrate otherwise, this is a serious correctness-risk concern for the empirical headline even though the algebraic framework may still be sound. The reproducibility statement promises future code/data release; the actual artifacts are not available at submission, which is worth monitoring in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a serious architecture study with a clean algebraic framing, but its headline empirical claim is compromised by a test-set selection problem that the paper doesn't properly acknowledge.\n\nThe genuinely new thing is FPRC-PQ: typed residual ownership per field, a single identity-erasure boundary at the aggregation mean, and a shared closer that only sees the fresh aggregate residual. The math is correct—the telescoping identity and the control-variate interpretation are true by construction, and the reflective rumination extension is honestly flagged as not part of the reported results. The empirical work is unusually careful: matched-capacity, unified-residual, identity-free, pairwise-only, and learned-aggregation controls, plus paired block-bootstrap inference. That's real effort and real credit.\n\nThe soft spot is load-bearing. Section 4 states every choice was frozen on data through 2022 and that no 2023–2026 label influenced anything. But Section 5.2 and Table 2 describe five audits run on the strict 2023–2026 book, with paired deltas on that window, that \"selected (9)–(11) over its neighbors.\" Even if each decision rule was preregistered, applying those rules to the evaluation window and then reporting the winner's performance on the same window makes the headline 5.58 pp net return and 0.67 Sharpe improvement post-selection. The bootstrap CIs and p-values don't account for that selection step. The paper needs to clarify this: either the audits were done on a pre-2023 validation set and the evaluation window was truly untouched, or the current numbers should be treated as exploratory.\n\nMinor but worth noting: no code or data are released yet, so the backtest can't be independently checked. One market, one asset class, and only a 42-month window, so external validity is limited. The paper's qualitative message—that placing capacity where the residual lives matters more than raw capacity—may survive even if the exact deltas don't, but the specific numbers are probably optimistic.\n\nThis paper deserves a serious referee. The architecture is interesting and the controls are a model for the field, but the selection issue must be resolved before the quantitative claims can be trusted. I'd send it to review and ask for a clear pre-registration timeline and ideally the code.","headline":"A thoughtful architecture study with clean math and careful controls, but its headline gains are compromised by architectural choices made on the evaluation window.","tokens_in":17867,"tokens_out":2987,"would_cite":false,"duration_ms":28132,"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":"Typing residuals by their generating representation and composing them in a fixed order—relax, aggregate, close—lifts net return from 13.52% to 19.10% on a frozen 2023–2026 Chinese A-share test.","keywords":["residual algebra","typed residuals","factor fields","point-in-time evaluation","cross-sectional return prediction","control variates","gradient boosting","representation-preserving learning"],"falsifier":"A preregistered replication on a different market that fixes all architectural choices using only data before the evaluation window and then runs the same matched controls—direct, unified-residual, identity-free two-stage, pairwise-only—would settle the claim, since the central claim predicts FPRC-PQ should beat those controls out of sample and a null or reversal would falsify it.","tokens_in":16836,"feed_emoji":"📈","tokens_out":13042,"duration_ms":105075,"temperature":0.7,"pith_summary":"Conventional learning from heterogeneous representations flattens them into one feature vector, and in doing so loses the identity of the representation that caused each error. The paper argues that the residual should be typed: each representation owns its own coordinate system and the residual it leaves unresolved, and learning should be an ordered composition of operators that preserve or deliberately erase that type. The evaluated instance, FPRC-PQ, instantiates the algebra as relax–aggregate–close on three $10\\times10$ rank-grid fields, and on 3.67 million Chinese A-share stock-day observations (2023–2026) under a frozen point-in-time protocol it raises net-of-cost return from 13.52% to 19.10% and Sharpe from 1.42 to 2.09. Matched-capacity, unified-residual, identity-free two-stage, and pairwise-only controls all trail it, which the paper takes as evidence that the gain comes from residual ownership and composition order, not from more features or more trees.","feed_headline":"Residual ownership beats feature pooling by 5.58 points","feed_subtitle":"With the protocol frozen, relax–aggregate–close lifts net return from 13.5% to 19.1% and Sharpe from 1.42 to 2.09.","key_machinery":"The central object is a typed residual: a residual is owned by the representation that generated it and can be corrected only in that representation's own coordinates. Fold realizes representations as $10\\times10$ point-in-time conditional-mean grids; FPRC-PQ realizes the algebra as relax–aggregate–close. The load-bearing identities are the telescoping decomposition $\\hat S=\\bar F+\\hat\\rho+\\widehat{(\\rho-\\hat\\rho)}$, which partitions capacity across stages, and the control-variate form $B_{PQ}=\\bar F+\\hat\\rho$, whose population analogue weakly reduces residual second moment. A further identity fixes the analytical rumination gain $\\lambda^\\star=\\langle e_0,q\\rangle/\\langle q,q\\rangle$ by orthogonal projection, making the reflective extension parameter-free.","core_discovery":"Formally, a Fold field $F_m$ is the point-in-time conditional-mean map over a $10\\times10$ grid of daily rank deciles, and the population residual $r-F_m$ is what that representation leaves unresolved. The algebra composes three typed operators: relaxation corrects each field in its own coordinates, aggregation $A$ takes the equal-mean interface and is the only step that erases representation identity, and closure fits the aggregate's fresh residual without rereading local state. The composition telescopes to $\\bar F+\\hat\\rho+\\widehat{(\\rho-\\hat\\rho)}$, so capacity is partitioned across stages, and the aggregate $B_{PQ}=\\bar F+\\hat\\rho$ is a learned control-variate interface. The central discovery is that this ordering, not raw capacity, is what separates the architecture from pooling: the same features, the same tree budget, and the same pairwise interaction class all fail to match FPRC-PQ, which asserts that representation-preserving residual composition is the cause of the reported 5.58 pp net-return improvement and the 0.67 Sharpe improvement.","pith_inferences":["A testable transfer consequence: the same relax–aggregate–close routing could be applied to heterogeneous representations outside finance—for example, different model families or feature blocks in a tabular benchmark—provided each operator is fitted only to its own residual and the identity-erasure boundary is fixed; the paper leaves this extension implicit.","Reflective rumination is an unevaluated theoretical member of the algebra; a natural next experiment is to fit the reconstruction $T$ and the analytical gain $\\lambda^\\star$ on data strictly before the evaluation window and compare the $\\lambda^\\star$ member against the base backbone, which would tell whether the algebra has a second working mechanism beyond the evaluated relax–aggregate–close ins","The reported margin could depend on the field-selection rule in Appendix A, which orders candidate triples by pre-2023 residual-error standard deviation; a replication that randomizes field selection while fixing the algebra would separate the contribution of the selected fields from the contribution of the composition itself."],"forward_implications":["Each correction stage should fit only the fresh residual left by its predecessor; letting the shared closer reread local states or the aggregate's own rank is significantly harmful.","Capacity placement matters more than capacity size: a matched-capacity direct learner with exactly the combined tree budget trails FPRC-PQ by 2.13 pp net and loses in all four years.","Preserving representation identity through local correction is necessary: collapsing the three fields into one concatenated representation and fitting a unified residual trails by 2.27 pp in 4/4 years, and an identity-free two-stage booster trails by 1.69 pp.","The equal-mean aggregation interface is already sufficient: learned coefficient reweighting moves net return by less than 0.05 pp within noise, while doubling the anchor harms all three chains.","Side information cannot repair a mistyped residual target: eight meta-information channels all fail, several significantly, and an error-correlated drift scalar destroys 1.13 pp when exposed after identity erasure."],"supporting_citations":[{"why":"Supplies the stagewise boosting framework that the paper re-types with representation-specific residual targets and ordered composition.","marker":"[10]"},{"why":"Supplies the gradient-boosted tree method used for every fitted operator and for the direct and matched-capacity controls.","marker":"[7]"},{"why":"Supplies the point-in-time, purge-gap, and leakage-avoidance protocol that defines the frozen evidence discipline.","marker":"[24]"},{"why":"Supplies the 21-day block bootstrap used for all paired inference on net-return deltas and intervals.","marker":"[22]"},{"why":"Supplies the control-variate identity that underlies reading the aggregate $B_{PQ}=\\bar F+\\hat\\rho$ as a variance-reducing interface.","marker":"[27]"},{"why":"Supplies the Neyman orthogonality concept used in the coupled-path mean-orthogonality claim for the refitted closer.","marker":"[8]"},{"why":"Supplies the multiple-testing and backtest-overfitting caution that motivates preregistering contrasts rather than promoting a post-hoc winner.","marker":"[16]"}],"fun_headline_variants":["Residual ownership beats pooling by 5.58 points","Algebraic residual composition yields 5.58 pp gain","Representation-preserving residual algebra wins by 5.58","Typed residual composition outperforms pooling by 5.58 pp","Residual algebra: 5.58-point edge over feature pooling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the claim that no 2023–2026 label or outcome influenced any architectural choice, because if the audits used evaluation-window data to choose the structure, the compared systems are no longer independent of the test set.","fun_headline_variants_meta":{"raw":{"variants":["Residual ownership beats pooling by 5.58 points","Algebraic residual composition yields 5.58 pp gain","Representation-preserving residual algebra wins by 5.58","Typed residual composition outperforms pooling by 5.58 pp","Residual algebra: 5.58-point edge over feature pooling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000815,"raw_usage":{"total_tokens":3651,"prompt_tokens":1102,"completion_tokens":2549,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":2463}},"tokens_in":718,"tokens_out":2549,"duration_ms":16979,"temperature":1.0,"reasoning_tokens":2463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T05:45:55.473550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A preregistered replication on a different market that fixes all architectural choices using only data before the evaluation window and then runs the same matched controls—direct, unified-residual, identity-free two-stage, pairwise-only—would settle the claim, since the central claim predicts FPRC-PQ should beat those controls out of sample and a null or reversal would falsify it.","supporting_citations":[{"cited_title":"Friedman","cited_arxiv_id":null,"evidence_quote":"Supplies the stagewise boosting framework that the paper re-types with representation-specific residual targets and ordered composition."},{"cited_title":"XGBoost: A scalable tree boosting system","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-boosted tree method used for every fitted operator and for the direct and matched-capacity controls."},{"cited_title":"John Wiley & Sons, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the point-in-time, purge-gap, and leakage-avoidance protocol that defines the frozen evidence discipline."},{"cited_title":"K ¨unsch","cited_arxiv_id":null,"evidence_quote":"Supplies the 21-day block bootstrap used for all paired inference on net-return deltas and intervals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the control-variate identity that underlies reading the aggregate $B_{PQ}=\\bar F+\\hat\\rho$ as a variance-reducing interface."},{"cited_title":"Harvey, Yan Liu, and Heqing Zhu","cited_arxiv_id":null,"evidence_quote":"Supplies the multiple-testing and backtest-overfitting caution that motivates preregistering contrasts rather than promoting a post-hoc winner."}],"review_version":1}