{"id":"b54f5671-c723-4056-98cb-bfdb4b3cc494","arxiv_id":"2607.08588","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Multi-horizon kNN forecast-error growth that fits Mittag–Leffler better than exponential is a preliminary diagnostic of fractional-memory dynamics in scalar time series.","lead":"A kNN multi-horizon forecast-error curve can flag fractional-memory dynamics when Mittag–Leffler growth fits better than exponential growth. The method needs only a scalar time series and may help justify fractional models before full system identification.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Missing integer-order negative control leaves the diagnostic's specificity untested.","rationale":"The Reader correctly isolates the specificity gap as the load-bearing weakness: the paper never shows that the ML-vs-exp preference fails on integer-order chaos or on classical long-memory mimics under the same pipeline. All positive results (58% RMSE cut, 100% bootstrap Δ>0, stable-contraction improvement) are obtained exclusively on fractional generators, so they cannot distinguish “fractional memory” from “any non-exponential geometry that an extra free parameter can fit.” The local-slope argument is model-independent but still only demonstrates non-exponentiality, not fractional origin. The paper’s own careful disclaimer that α_fit is merely an effective shape parameter further underscores that the diagnostic is under-specified without negative controls. Adding the integer-order control (and ideally an ARFIMA or structural-break series) is the minimal concrete check that would decide whether the claim can be upgraded from “consistent with fractional systems” to “diagnostic of fractional dynamics.” Until that is done, CONDITIONAL remains the appropriate verdict; no stronger rejection is warranted because the positive synthetic results and bootstrap stability are real, and the positioning as a preliminary indicator is appropriately modest.","tokens_in":14483,"tokens_out":658,"duration_ms":6298,"concrete_test":"Simulate the integer-order (α=1) version of system (3.1) with the same parameters, Δt, stride, embedding, K, and H_max as Table 3; recompute the normalized G(h) curve and fit both exponential and free-order ML models exactly as in §2.5. If Δ=RMSE_exp−RMSE_ML remains positive and of similar magnitude (or α_fit≪1), the fractionality interpretation is not specific; if Δ≤0 and α_fit≈1, the concern is largely resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that superior Mittag–Leffler (vs exponential) fit of multi-horizon kNN forecast-error growth is a diagnostic indicator of fractional-memory dynamics. That claim requires that the same pipeline not produce a comparable ML preference on non-fractional systems that still generate non-exponential error curves. Section 3 evaluates only Caputo fractional systems (chaotic system (3.1) with α_true=0.916; stable relaxation (3.3) with α=0.70). No matched integer-order chaotic map or ODE is run under identical embedding (m=8, τ_embed=2), K=3, horizon schedule, geometric-mean error, and free-order ML fit. The free-order model has an extra parameter α_fit; globally it collapses to α_fit≈0.007 (Table 4), far from α_true, so the RMSE gain (58%, Δ>0 in 100% of bootstraps) may simply reflect flexible curvature rather than fractional memory. The Discussion itself flags long-memory mimics and structural breaks but does not test them. Without the negative control, the reported preference is necessary but not shown to be specific.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a data-driven preliminary diagnostic for fractional-memory signatures in scalar time series: multi-horizon out-of-sample kNN forecast-error growth (geometric-mean absolute error after delay embedding) is fitted by classical exponential and free-order Mittag–Leffler models, with local log-slope behavior used as supporting evidence. On a Caputo fractional chaotic system the Mittag–Leffler model reduces RMSE by ~58% relative to exponential (Δ>0 in 100% of 500 bootstrap replicates); analogous gains appear for analytical and kNN-based stable fractional relaxation/contraction. The authors correctly treat the fitted order α_fit as an effective shape parameter of the error curve rather than a recovery of the true system order, and position the pipeline as an intermediate check before full fractional identification.","tokens_in":14854,"tokens_out":1051,"duration_ms":19552,"significance":"If the diagnostic is shown to be reasonably specific, it supplies a practical, equation-free, time-domain pre-model test that could curb unjustified fractionalization of complex systems when only scalar observations are available. Strengths that should be retained include the nonparametric kNN pipeline, the explicit refusal to equate α_fit with the true order, the dual evaluation on chaotic growth and stable contraction, the bootstrap stability analysis, and the connection to the authors’ earlier forecast-error LLE estimator. These elements make the contribution more than a pure curve-fitting exercise and give it clear applied value in dynamical systems and fractional modeling.","major_comments":[{"comment":"Section 3 evaluates only Caputo fractional systems (chaotic system (3.1) with α_true=0.916 and stable relaxation (3.3) with α=0.70). The central claim that superior Mittag–Leffler versus exponential fit of G(τ) is a diagnostic of fractional-memory dynamics therefore lacks a matched integer-order negative control run under identical embedding (m=8, τ_embed=2), K=3, horizon schedule, geometric-mean error, and free-order ML fitting. Without that control (e.g., the logistic or other maps already used in the authors’ prior LLE work), it remains untested whether the observed preference is specific to fractional memory or simply reflects any non-exponential geometry that the three-parameter ML model can capture.","section":null},{"comment":"Table 4 and the global fit: the free-order Mittag–Leffler model has three free parameters (C_ML, λ_ML, α_fit) against two for the exponential model, yet model comparison is reported solely via RMSE (and log-RMSE). Globally α_fit collapses to ≈0.007, far from α_true=0.916, confirming that the RMSE gain is driven by flexible curvature rather than recovery of the system order. An information-criterion or nested-model penalty (as already used for the stable kNN contraction in Table 6) is needed to show that the preference survives the extra degree of freedom; otherwise the 58% reduction and 100% bootstrap Δ>0 are expected by construction and do not yet establish a fractional-specific signature.","section":null},{"comment":"Discussion correctly flags long-memory mimics (structural breaks, aggregation, low-frequency contamination) but does not subject any of them to the same multi-horizon kNN pipeline. Because the indicator is defined essentially as “ML fits G(τ) better than exponential plus decreasing local log-slope,” these classical confounders remain unexcluded; a minimal set of synthetic non-fractional long-memory or broken-trend series would make the specificity claim load-bearing rather than aspirational.","section":null}],"minor_comments":[{"comment":"Section headings contain stray spaces (“F ractional chaotic system”, “F ractional stable system”); clean for production.","section":null},{"comment":"Table 5 shows two intermediate windows with Δ≤0; a brief remark on the transitional regime would help readers interpret the windowed analysis.","section":null},{"comment":"Figures 2–3 would benefit from explicit statement of the horizon units (τ = h·Δt_obs) in the captions so that the plotted range is immediately comparable to the tables.","section":null},{"comment":"The reconstruction-parameter sweep (Table 2) is useful; stating whether the same (m,τ_embed) optimum was re-used for the stable contraction experiment would improve reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The missing integer-order negative control is the single most important fix; once supplied (and if the ML preference disappears or weakens on integer chaos), the paper becomes a solid methods contribution for math.DS / applied fractional dynamics. Novelty relative to the authors’ own 2025 LLE estimator is incremental but legitimate; the journal should not reject solely on that ground."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This paper takes the multi-horizon kNN forecast-error idea from Velichko’s LLE work and turns it into a preliminary check for fractional-memory signatures: fit exponential vs Mittag–Leffler to the geometric-mean error curve G(h), look at local log-slope, and bootstrap the preference. That is the actual novelty. On the Caputo systems they run (chaotic α=0.916 and stable relaxation/contraction), ML wins cleanly—58 % RMSE drop, Δ>0 in every bootstrap replicate, and free-order ML cuts contraction RMSE by nearly an order of magnitude. They correctly refuse to treat α_fit as the true order (it collapses to ~0.007) and call it a shape parameter. Methods are spelled out well enough to reimplement, the math is standard Caputo/Mittag–Leffler, and the citation pattern covers the right literature on unjustified fractionalization, long-memory mimics, and identification pipelines.\n\nThe soft spot is real but proportionate: they never run a matched integer-order chaotic map or ODE under the same embedding, K, and free-order ML fit. Without that negative control, superior ML fit could just be the extra free parameter capturing any non-exponential geometry. They flag structural-break and aggregation mimics in the discussion but do not test them. That leaves the diagnostic necessary but not yet shown to be specific. Code and data are not public, which is a minor practical drag.\n\nThis is for people who build or critique fractional models from scalar series and want a cheap pre-model screen before full identification. It is not a theory paper and does not claim order recovery. The central argument holds on the systems they study; the missing control is the main thing a referee should demand. I would send it to peer review rather than desk-reject, and I would read the revision if the controls appear. Worth engaging if you work on memory diagnostics or fractional system ID; otherwise it is optional.","headline":"Useful method extension of forecast-error growth to a fractional diagnostic, honest about its limits, but specificity is still untested without integer-order controls.","tokens_in":15439,"tokens_out":493,"would_cite":false,"duration_ms":19424,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37M10","26A33","34A08"],"pacs":[],"model":"grok-4.5","headline":"Multi-horizon kNN forecast-error growth can flag fractional dynamics when the error curve fits a Mittag–Leffler law better than an exponential one.","keywords":["Fractional dynamics","forecast error growth","k-nearest neighbors","Mittag–Leffler function","fractional systems","delay embedding","chaos detection"],"falsifier":"Run the identical pipeline on matched integer-order chaotic maps and on classical long-memory series that are not fractional (e.g., structural-break or aggregated processes): if Mittag–Leffler still wins RMSE and local-slope tests as often as on true fractional Caputo systems, the diagnostic does not specifically indicate fractional dynamics.","tokens_in":15356,"feed_emoji":"📈","tokens_out":1006,"duration_ms":16489,"temperature":0.7,"pith_summary":"This paper argues that you can test a scalar time series for fractional-memory signatures without knowing the governing equations. Train a simple k-nearest-neighbors forecaster on delay-embedded observations, then watch how the out-of-sample prediction error grows (or contracts) as the forecast horizon lengthens. Classical chaotic integer-order systems produce roughly exponential error growth; fractional systems are expected to produce slower, Mittag–Leffler-type growth or decay. By fitting both laws to the same empirical curve and checking that the local slope of the log-error declines, the method builds a preliminary fractionality indicator. On a known fractional chaotic system the Mittag–Leffler model cut RMSE by about 58 percent and won in every bootstrap replicate; the same advantage appears in stable fractional relaxation and in a kNN contraction test. The fitted Mittag–Leffler order is treated only as a shape parameter of the error curve, not as a recovered system order.","feed_headline":"Forecast-error curves flag fractional dynamics","feed_subtitle":"A kNN multi-horizon test reads Mittag–Leffler memory signatures from scalar series alone","key_machinery":"The multi-horizon kNN forecast-error curve: after delay-embedding a scalar series and training a K-nearest-neighbors regressor, form the geometric-mean absolute error G(h) at each horizon h, normalize by G(1), then compare exponential versus Mittag–Leffler parametric fits (plus the local slope of log G) as a fractionality indicator.","core_discovery":"The geometry of multi-horizon forecast error on a scalar series carries a usable signature of fractional dynamics: when the normalized kNN error-growth (or contraction) curve is better described by a Mittag–Leffler law than by a classical exponential law, and when the local slope of the log-error falls with horizon, the series is consistent with fractional-memory behavior. This holds for both chaotic divergence and stable relaxation, and the preference for Mittag–Leffler is stable under bootstrap resampling of the test set.","pith_inferences":["Without published negative controls on integer-order chaos and non-fractional long memory under the same pipeline, the indicator’s specificity remains unproven and should be the first follow-up experiment.","Because α_fit is only a shape parameter, pairing this diagnostic with an independent order estimator (or multivariate measurements) would be the natural next step toward full identification.","If the method survives noise and real experimental series (viscoelastic, electrochemical, biological), it could become a cheap gatekeeping step against unjustified fractionalization of models."],"forward_implications":["Scalar observations alone can supply a pre-model check before one commits to a fractional governing equation.","The same forecast-error geometry already used for chaos detection can double as a dynamical-characterization tool for memory effects.","A positive diagnostic supports trying fractional models in applications where only limited measurements exist, while leaving exact order recovery as a separate inverse problem.","Windowed and bootstrap comparisons of RMSE and log-RMSE give a practical robustness layer for the indicator."],"fun_headline_variants":["Mittag-Leffler forecast-error growth flags fractional dynamics","kNN multi-horizon errors reveal Mittag-Leffler fractional signatures","Error-growth curves diagnose fractional memory from scalar series","Multi-horizon kNN tests detect fractional dynamics via Mittag-Leffler fit","Forecast-error geometry signals fractional systems over exponential chaos"],"cache_read_input_tokens":6016,"weakest_assumption_plain":"A better Mittag–Leffler fit to the forecast-error curve is taken as evidence of fractional memory rather than of any other non-exponential geometry, extra free parameter, embedding choice, or long-memory mimic.","fun_headline_variants_meta":{"raw":{"variants":["Mittag-Leffler forecast-error growth flags fractional dynamics","kNN multi-horizon errors reveal Mittag-Leffler fractional signatures","Error-growth curves diagnose fractional memory from scalar series","Multi-horizon kNN tests detect fractional dynamics via Mittag-Leffler fit","Forecast-error geometry signals fractional systems over exponential chaos"]},"model":"grok-4.5","effort":"low","cost_usd":0.005578,"raw_usage":{"total_tokens":1569,"prompt_tokens":911,"num_sources_used":0,"completion_tokens":92,"cost_in_usd_ticks":55780000,"prompt_tokens_details":{"text_tokens":911,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":566,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":911,"tokens_out":92,"duration_ms":5234,"temperature":1.0,"reasoning_tokens":566,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T04:43:22.430539+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the identical pipeline on matched integer-order chaotic maps and on classical long-memory series that are not fractional (e.g., structural-break or aggregated processes): if Mittag–Leffler still wins RMSE and local-slope tests as often as on true fractional Caputo systems, the diagnostic does not specifically indicate fractional dynamics.","supporting_citations":[],"review_version":1}