{"id":"49273b9b-0356-4e73-a220-e1e95e05240a","arxiv_id":"2607.19005","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Applying the Observable Matrix Dynamics toolkit to S&P 500 data yields crisis-specific correlation geometries, a market that never settles into a stable structure, and a weak episodic time-asymmetry in the volatility ranking.","lead":"This paper tracks the S&P 500 through three crashes using fixed-size matrices built from stock correlations and rankings, finding that 2008 and 2020 collapsed market structure while the 2001 bust was a slow, spread-out unwind. It also reports a weak arrow of time in the volatility ranking that flares during market stress, matching a known finance effect.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'market never learns' claim depends on interpreting the spectrum via calibrations from the author's own preprints; an independent re-derivation or external calibration is needed.","rationale":"The reader identified the same weakest assumption: the benchmark distinguishing 'no relaxation' from wash-out/noise is not independently calibrated. The paper's own caveats in Section 3 confirm the risk. The central claim's external validity hinges on the author's own I-BBS finite-N correction and OMD regime definitions. A concrete independent check on the finite-N correction would settle whether the observed beta<1 is meaningful. Until then, conditional acceptance with the request for external calibration is appropriate. No ad hominem; the critique is on the argument's load-bearing calibration.","tokens_in":23365,"tokens_out":1063,"duration_ms":9251,"concrete_test":"Independently re-derive or numerically simulate the I-BBS finite-N correction Delta_beta(N,d) for N≈244–309 and d∈[2,10], without relying on the author's preprint [8]. If the corrected beta for a uniformly sampled latent manifold is consistent with the observed 0.7, then the spectral data cannot distinguish 'never learned' from 'finite-N artifact'.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that the market 'never learns' relies on comparing the empirical distance-matrix spectrum to the un-relaxed/relaxed regimes defined in the author's own OMD and I-BBS preprints [7, 8]. The paper explicitly lists wash-out effects (finite T of order N, small N with sqrt(N)≈16, few-factor-plus-noise cross section) that could explain the smooth spectrum without implying no relaxation. The measured beta≈0.7 is below the BBS unit threshold, and the finite-N correction to beta is taken from the same author's preprint [8]. If that correction is inaccurate, or if the wash-out effects fully account for the smooth spectrum, the distinction between 'never-learning' and 'relaxed-but-noisy' collapses. The claim is not internally inconsistent, but its external validity depends on unverified calibration from the author's own prior work, and the paper itself states the absence of fine BBS structure is 'an open point'. Since the paper's main empirical findings (spectral collapse, sector rotation, ranking-chain persistence) are plausible and clearly presented, the load-bearing weakness is specifically the benchmark calibration for the central 'never learns' interpretation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper applies the Observable Matrix Dynamics (OMD) framework to three S&P 500 crisis periods (2001, 2008, 2020), using a fixed-universe panel and three fixed-size matrix observables: an arccos distance matrix built from rolling return correlations, and two rank-ordered Markov transition matrices for performance and volatility. The distance-matrix analysis reports a collapse of the effective factor count at the 2008 and 2020 crises, a dispersed unwind in 2001, a coherent sector-eigenvector rotation after market-factor removal, and a causal trailing-window ROC that finds early-warning skill only for the 2008 crisis. The ranking-chain analysis reports persistence and mixing-time differences between the return and volatility rankings, near-reversibility of the return chain, a weak episodic arrow of time in the volatility chain, and sector-level transfer-entropy leadership, including a utilities-led performance channel and a financials-led risk channel. The paper interprets the persistent, shoulder-free distance-matrix spectrum with beta near 0.7 as evidence that the market never learns or relaxes to a stationary low-dimensional geometry.","tokens_in":23513,"tokens_out":6862,"duration_ms":67124,"significance":"If the central interpretation is accepted, the paper makes a substantive empirical contribution. It shows that trajectory-level diagnostics (Perron scale, participation ratio, eigenbasis rotation, commutator norm) add information beyond a single correlation matrix, separates correlated crashes from dispersion events, and provides crisis-specific sector composition and name-level attribution. The ranking-space analysis offers a new market-neutral view with entropy production and transfer entropy that reproduces volatility clustering and the Zumbach effect in a discrete representation. The empirical execution is careful in several respects: fixed-universe panels keep the cross section stable, the lookback windows are chosen to keep N < L, the ROC analysis uses trailing data only, an external stress-event calendar is used for validation, and the limitations are stated explicitly. The principal weakness is that the load-bearing 'never learns' interpretation depends on benchmarks calibrated in the author's own companion preprints, and the paper's own caveats permit a relaxed-but-noisy alternative.","major_comments":[{"comment":"The central claim that the market 'never learns' or 'never relaxes to an equilibrium M-matrix' depends on identifying the measured delocalised exponent beta near 0.7 with the 'un-relaxed, pre-learning regime' whose benchmark range beta in [0.65, 0.81] is taken from the OMD preprint [7], and on the I-BBS finite-N correction from preprint [8] to explain why the measured beta lies below the BBS threshold beta = 1. Because both sources are the author's own preprints, the calibration is not independently verifiable from this manuscript. The paper itself lists three effects that plausibly wash out the strict BBS fine structure - finite lookback with T of order N, small universe with sqrt(N) about 16 to 18, and a few-factor-plus-noise cross section - and calls the absence of a shoulder 'an open point.' Under these qualifications, the observed smooth spectrum is equally consistent with a relaxed-but-noisy system of small size. Please add an independent calibration: for example, simulate stationary factor-plus-noise returns with the same N and L and known latent dimension, and show that the resulting spectrum and beta distribution differ from the market's, or provide external replication of the I-BBS finite-N correction.","section":"Section 3, Figure 2"},{"comment":"The claim of no stationary limit sits in tension with the paper's own evidence that the observables revert after each crisis. Figure 4 shows the Covid trajectories plateau for approximately the lookback length and then 'steps back down', and Section 4.8 describes the MDS cloud re-expanding to 'its calm size' with 'the market back to normal.' A process that repeatedly returns to a pre-crisis baseline is mean-reverting and non-stationary, but that is not the same as never relaxing to a stationary geometry. The statement that the market 'never relaxes' needs an operational definition (for example, no convergence of the full spectral shape to a fixed attractor over decade-long windows, or beta never crossing a threshold) together with a control showing what observed trajectories would look like under a relaxational alternative. As written, the claim is stronger than the evidence.","section":"Section 4.1 and Section 4.8"},{"comment":"The power-law exponent beta is fit on at most sqrt(N) - 2, or about 14 to 16, eigenvalues per date for N between 244 and 309, and no confidence intervals or goodness-of-fit diagnostics are reported for the fitted beta. Given that the diagnostic distinction between the market's beta near 0.7 and the benchmark un-relaxed range near 0.81 is numerically narrow, the absence of error bars makes it difficult to assess whether the market's trajectory is genuinely distinguishable from any of the reference regimes. Please report standard errors or bootstrap intervals for beta at each date, and ideally for the Perron eigenvalue as well.","section":"Section 3, Eq. (2)"}],"minor_comments":[{"comment":"Please state how missing CRSP returns, trading halts, and any suspended or delisted names are treated in the fixed-universe panel; the current description of a 'fixed universe' does not fully specify the handling of incomplete daily records.","section":"Section 2"},{"comment":"The random-subspace null for the projector drift uses the assumption that the two K-frames are independent and uniformly random; this should be stated explicitly, and at the smallest short-lookback universe with N = 94 a finite-N correction to the expectation E||V1^T V2||^2_F = K^2/N may be worth quantifying.","section":"Section 4.4"},{"comment":"The detailed-balance surrogate used for the entropy-production z-scores is described only as 'a reversible chain matched to the block's marginals and sample size'; the exact construction (for example, a Metropolis-filtered symmetrized transition matrix) should be specified so that the reported z-scores in Figure 17 are reproducible.","section":"Section 5.1"},{"comment":"The per-period AUC values (0.46 to 0.72) are in-sample and based on only three crisis periods, so the conclusion of regime-dependent forecastability would be strengthened by bootstrap confidence intervals and by a statement of how sensitive the AUC is to the drawdown threshold and the 63-day horizon.","section":"Section 4.7, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the self-referential benchmark: the 'un-relaxed' signature and the finite-N correction that together support the paper's headline claim are drawn from two companion preprints by the same author ([7] and [8]) that are not yet peer-reviewed and not included with the manuscript. I would encourage the editor to ask that the essential calibration material, or an independent replication, be added to the paper or a public supplementary file. The empirical work is otherwise well executed and clearly written, and the paper fits the scope of q-fin.ST; the issue is specifically the strength of the 'never learns' interpretation relative to the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing. First, the empirical work here is honest, careful, and mostly convincing: fixed-universe panels, L>N lookbacks, surrogate tests against a reversible null, and a public code repository. The trajectory-level diagnostics are genuinely new in this context, and the main factual findings—correlation collapse in 2008 and 2020, the dispersed 2001 unwind, sector rotation under market-factor removal, persistence of the volatility ranking, and the weak episodic arrow of time in volatility—are plausible and consistent with what the literature already knows. Second, the headline claim that the market never learns or relaxes is not established. It depends on interpreting the smooth spectrum through the un-relaxed regime and the finite-N correction from the author's own preprints, and the paper itself lists three wash-out effects (finite T of order N, small N, few-factor-plus-noise cross section) that could produce the same smooth spectrum without implying no relaxation. The abstract sells this claim harder than Section 3 supports; Section 3 even calls the absence of fine BBS structure an open point. That tension is the paper's load-bearing weakness.\n\nThe rest of the paper does not depend on that claim, which is why I would still send it to a serious referee. The spectral collapse, the 2001-versus-2008/2020 distinction, the eigenbasis rotation diagnostics, the name-level attribution, and the ranking-chain results are all worth having, and the finite-sample controls are above average for this literature. The code is public, the analysis is reproducible, and the limitations section is candid.\n\nSoft spots, in proportion. The no-relaxation benchmark needs external calibration or an independent re-derivation; as it stands, the central claim is circular in a narrow but real way. The headline diagnostics (beta, mixing times, z-scores, AUCs) mostly lack error bars or bootstrap intervals; the per-period ROC is explicitly in-sample. And the stress-event selectivity of the volatility arrow is post-hoc: a pre-specified definition of sustained directional versus symmetric shocks would make it a usable regime indicator. These are fixable, and none of them undermines the core empirical findings.\n\nWho should read this: statistical finance and random-matrix people, anyone working with correlation geometry or rank-based market dynamics, and followers of the OMD/I-BBS program. Practitioners should not expect a tradeable signal. Recommend peer review with a request for an independent benchmark on the no-learning claim and bootstrap intervals on the headline diagnostics.","headline":"A careful, reproducible empirical toolkit for reading equity correlation geometry and rank dynamics, but the central 'market never learns' claim needs an independent benchmark before it should be taken as established.","tokens_in":24121,"tokens_out":1552,"would_cite":true,"duration_ms":16421,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that the equity market's correlation geometry never relaxes: across three crisis decades the distance-matrix spectrum stays in the un-relaxed, pre-learning regime, collapsing in dimension at crises without ever forming a…","keywords":["Observable Matrix Dynamics","distance-matrix spectrum","random matrix theory","participation ratio","entropy production","transfer entropy","volatility clustering","market crises"],"falsifier":"Recompute the distance-matrix spectrum on a much larger universe, with thousands of names, using a lookback long enough that $T\\gg N$ and a noise-cleaned correlation estimator; if a shoulder appears near $K\\approx\\sqrt{N}$, quasi-multiplets form, or $\\beta$ rises above one, the market would be relaxing to a manifold and the never-learns claim would be false.","tokens_in":23018,"feed_emoji":"📉","tokens_out":9882,"duration_ms":80526,"temperature":0.7,"pith_summary":"This paper argues that the large-cap US equity market, watched through a rolling distance-matrix trajectory, is a permanently non-equilibrium object: its correlation geometry collapses at the 2008 and 2020 crises and never settles onto a low-dimensional manifold, the state a learning system occupies before learning rather than after. The evidence is the spectrum of the arccos distance matrix, which keeps a smooth power-law decay with exponent near 0.7 and no shoulder or quasi-multiplets across three crisis decades. The same fixed-size matrices also build two market-neutral Markov chains on daily return and volatility rankings: the return chain mixes in about a week and stays near time-reversible, while the volatility chain mixes in a month or more and carries a weak, episodic arrow of time that flares at market stress. If the paper is right, crises are at once a collapse of effective dimension and a burst of rank migration, and the reason an endogenous crisis like 2008 was forecastable while an exogenous shock like 2020 was not becomes a readout of the same geometry.","feed_headline":"Stock market never learns its correlation geometry","feed_subtitle":"Across three crisis decades the spectrum stays un-relaxed: dimension collapses at 2008 and 2020, never settling.","key_machinery":"The central object is the trajectory of the fixed-size arccos distance matrix $M_{ij}(t)=\\arccos C_{ij}(t)$, built from rolling return correlations, together with two rank-ordered transition matrices $P^R(t)$ and $P^V(t)$ for daily return and volatility rankings. The distance-matrix spectrum is read through the Perron eigenvalue, the delocalised power-law slope $\\beta$ from BBS theory, and the participation ratio of the correlation matrix, while the ranking chains are read through their second-eigenvalue mixing times, entropy production against a reversible null, and transfer entropy. The machinery's work is to compactify a high-dimensional evolving cross section into fixed-size objects whose spectral and geometric changes mark crises and distinguish co-movement from dispersion.","core_discovery":"The central claim is that the market never learns its own correlation structure. Read against distance-matrix spectra from learning systems at initialisation, the market's spectrum stays in the un-relaxed regime: a smooth power-law decay with delocalised exponent $\\beta\\approx 0.7$, no quasi-multiplets, no shoulder at $K\\approx\\sqrt{N}$, and no relaxation to an equilibrium $M$-matrix. Supporting observations come from three fixed-size observables: the arccos distance matrix $M_{ij}(t)=\\arccos C_{ij}(t)$ shows the participation ratio collapsing from roughly nine effective factors to four at Covid and from eight to five at 2008, while the 2001 bust runs the other way as a dispersed, decorrelated unwind; removing the market factor leaves a coherent sector rotation led by utilities; and the return and volatility ranking chains show fast, near-reversible performance dynamics beside slow, stress-flaring volatility dynamics. The forecastability result follows from the same machinery: fragility signals built only from trailing returns predict the endogenously built-up 2008 crisis with a per-period area-under-the-ROC-curve near 0.72 at the 63-day horizon, but not the exogenous 2020 shock or the 2001 unwind.","pith_inferences":["If the claim holds, the same spectrum can serve as a generic test for whether any complex system has a learnable low-dimensional structure: a distance-matrix spectrum that never develops a shoulder or quasi-multiplets would mark a system with no relaxed geometry.","The forecastability contrast between 2008 and 2020 suggests a crisis taxonomy by origin that could guide when to trust fragility indicators: endogenous build-ups leave a spectral trail, exogenous shocks do not, and the two should not be pooled into one average skill score.","The volatility-chain arrow of time could be developed into a real-time regime indicator that ignores brief symmetric spikes and fires only on sustained directional reshuffling; testing it on intraday data or other asset classes would show whether the selectivity is general.","If markets are truly non-equilibrium, structural changes such as new trading mechanisms or regulatory constraints should transiently steepen the delocalised exponent toward or above one if any learning-like relaxation were possible; a null result would reinforce the never-learns claim."],"forward_implications":["If the market never relaxes, covariance and risk models that assume a stationary long-run correlation structure are misspecified; the correlation geometry should be treated as a non-equilibrium trajectory.","The spectral signature separates crisis types: correlated crashes such as 2008 and 2020 collapse the effective factor count and raise the market-factor share, while the 2001 dot-com period is a dispersed unwind with the opposite signature.","Fragility signals that use only trailing data can give months of early warning for endogenously building crises but not for exogenous shocks, so early-warning skill is regime-dependent rather than universally present or absent.","Once the market factor is removed, the sector geometry follows its own timetable, as the post-Covid sector rotation continued for a year or more past the March 2020 crash.","The volatility ranking is five to six times more persistent than the return ranking, and its weak arrow of time flares at sustained directional stress, matching the known volatility-clustering asymmetry."],"supporting_citations":[{"why":"Supplies the OMD framework and the grokking-experiment reference spectra that define the un-relaxed versus relaxed distance-matrix signatures.","marker":"[7]"},{"why":"Supplies I-BBS inference and the finite-N correction that explains why the measured beta stays below one.","marker":"[8]"},{"why":"Supplies the BBS theory of distance-matrix spectra on spheres, the benchmark the market spectrum is read against.","marker":"[9]"},{"why":"Supplies the frustrated-distance-matrix relaxation dynamics and the trajectory-level rotation diagnostics used here.","marker":"[13]"},{"why":"Supplies the noise-bulk spectrum that accounts for the smearing of the fine BBS structure.","marker":"[1]"},{"why":"Supplies the stochastic-thermodynamics entropy-production definition used for the ranking chains.","marker":"[21]"},{"why":"Supplies the volatility-clustering asymmetry that the volatility arrow of time is said to match.","marker":"[26]"},{"why":"Supplies the transfer-entropy definition used to build the directed lead-lag networks.","marker":"[27]"}],"fun_headline_variants":["Market never learns its correlation geometry","Stock correlation spectra stay un-relaxed across crises","Crisis dimension collapses but market never learns","Unlearned correlations: market's geometry never settles","Market predicts 2008, not 2020, never learns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the market never relaxes rests on treating the smooth, shoulder-free distance-matrix spectrum as a genuine property of the market rather than an artifact of the finite sample, and the paper itself lists the short lookback, the small universe, and the few-factor-plus-noise cross section as effects that could wash out relaxation in a system that actually does relax.","fun_headline_variants_meta":{"raw":{"variants":["Market never learns its correlation geometry","Stock correlation spectra stay un-relaxed across crises","Crisis dimension collapses but market never learns","Unlearned correlations: market's geometry never settles","Market predicts 2008, not 2020, never learns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":3034,"prompt_tokens":1081,"completion_tokens":1953,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":1880}},"tokens_in":697,"tokens_out":1953,"duration_ms":14152,"temperature":1.0,"reasoning_tokens":1880,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:33:29.784434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the distance-matrix spectrum on a much larger universe, with thousands of names, using a lookback long enough that $T\\gg N$ and a noise-cleaned correlation estimator; if a shoulder appears near $K\\approx\\sqrt{N}$, quasi-multiplets form, or $\\beta$ rises above one, the market would be relaxing to a manifold and the never-learns claim would be false.","supporting_citations":[{"cited_title":"Learning as Observable Matrix Dynamics: Diffusive Relaxations versus Phase Transitions","cited_arxiv_id":"2606.29679","evidence_quote":"Supplies the OMD framework and the grokking-experiment reference spectra that define the un-relaxed versus relaxed distance-matrix signatures."},{"cited_title":"I-BBS: Coordinate-Free Inference of Latent Sub-Manifolds Using Random Distance Matrix Theory","cited_arxiv_id":"2606.29675","evidence_quote":"Supplies I-BBS inference and the finite-N correction that explains why the measured beta stays below one."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the noise-bulk spectrum that accounts for the smearing of the fine BBS structure."},{"cited_title":"Seifert.Entropy Production along a Stochastic Trajectory and an Integral Fluctuation Theorem","cited_arxiv_id":null,"evidence_quote":"Supplies the stochastic-thermodynamics entropy-production definition used for the ranking chains."},{"cited_title":"Zumbach.Time Reversal Invariance in Finance","cited_arxiv_id":null,"evidence_quote":"Supplies the volatility-clustering asymmetry that the volatility arrow of time is said to match."},{"cited_title":"Schreiber.Measuring Information Transfer","cited_arxiv_id":null,"evidence_quote":"Supplies the transfer-entropy definition used to build the directed lead-lag networks."}],"review_version":1}