{"id":"e892014d-ba96-423c-8d1e-6c1f001a7846","arxiv_id":"2411.14999","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The EE-classifier classifies functional data by plotting each curve's modified epigraph and hypograph indexes relative to the two classes and applying standard classifiers to the plot.","lead":"This paper presents a new classifier for functional data that turns each curve into two extremality scores, one relative to each class, and then applies off-the-shelf classifiers in that two-dimensional space. The method is tested on synthetic and real datasets, including a small S&P 500 up/down prediction task.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evaluation protocol in §4 never states that test-curve EE features are computed against training-only reference samples; if pooled references are used, reported accuracies are optimistically biased and the central empirical claim is unsupported.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the paper never states whether test curves contribute to the reference sample used to compute extremality indexes. This is the right concern because the central claim is empirical, not purely methodological. If the evaluation leaks label information through the index computation, then the high accuracies in Figures 4, 6, 8, 10, and 12 and in Table 2 do not support the conclusion that the EE-classifier generalizes. The concern is concrete and testable: a nested protocol that restricts reference samples to training folds would settle it. I considered the incorrect algebraic relation MHI(x) - MEI(x) = 1/n in Section 2.2. That is a genuine mathematical error (for continuous curves one expects MHI + MEI ≈ 1, not a constant difference of 1/n), but it is secondary: even if that line is corrected, the classifier's reported performance is unaffected. The S&P 500 interpretation is also overreaching, but it is an application rather than the central methodological claim. The paper has no machine-checked proofs or released code, so the empirical protocol is the main evidentiary load-bearing element. If the authors provide code or clarify that reference samples are training-only, the concern may be resolved without changing the conditional verdict.","tokens_in":14848,"tokens_out":6107,"duration_ms":64930,"concrete_test":"Re-run the synthetic and real-data experiments with a nested train/test protocol: split into training and test sets; compute MHI/MEI for training curves using only the training class samples as reference; for each test curve, compute MHI/MEI with respect to each class's training sample, never including any test curve in the reference empirical distributions; then fit the same classifiers on training features and score test features. Compare the resulting accuracies with Figures 4, 6, 8, 10, and 12 and Table 2. If the corrected accuracies drop by more than a few points, or if a label-permutation null under the original protocol yields above-chance accuracy, the reported results are leakage-inflated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the EE-classifier achieves high classification accuracy on synthetic and real functional data. That claim rests entirely on the reported cross-validation and test accuracies. Section 4.1 says only that the sample is split into 80% training and 20% testing, and that 100-fold cross-validation is used as train control; it never specifies the reference sample used to compute each curve's modified epigraph/hypograph index. Equation (12) defines the EE-plot with points x ∈ {X ∪ Y}, which is natural for a visualization but dangerous for evaluation: if a test curve is included in the reference empirical distribution used to compute its own MHI or MEI, then its feature is mechanically biased toward the label of the reference class. That is a direct label leak, and it would inflate both the cross-validation curves in Figures 4, 6, 8, 10, and 12 and the test accuracies in Table 2. The method itself may be sound, but the paper's empirical demonstration is not interpretable until this protocol is specified or corrected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a supervised classifier for functional data based on the modified epigraph index (MEI) and modified hypograph index (MHI). The method, called the EE-classifier, maps each curve to a two-dimensional space whose coordinates are the curve's extremality indexes with respect to the two class distributions, producing an 'EE-plot', and then applies standard classifiers (LDA, QDA, kNN, SVM, random forest) in that space. The authors compare the EE-classifier with the depth-based DDG-classifier on six simulated experiments and three real functional data sets, and apply the method to predict monthly up/down movements of the S&P 500 index. The central claim is that this extremality-based transformation yields competitive and robust classification accuracy for functional data.","tokens_in":15111,"tokens_out":4871,"duration_ms":47133,"significance":"If the empirical evaluation is valid, the EE-classifier is a simple and interpretable alternative to depth-based functional classifiers: it reuses existing extremality indexes, introduces no additional fitted parameters in the feature construction, and provides a useful visual diagnostic in the EE-plot. The paper also offers a systematic benchmark against the DDG-classifier across multiple synthetic and real settings, which is valuable for practitioners. However, the empirical claims rest on an incompletely specified cross-validation and testing protocol, and there are internal inconsistencies in the reported S&P 500 results; these issues need to be resolved before the central claim can be accepted.","major_comments":[{"comment":"The stated relation MHI(x) - MEI(x) = 1/n is incorrect. From Eq. (10), for almost every t, I(xi(t) <= x(t)) + I(xi(t) >= x(t)) = 1, so summing over i gives MHI(x) + MEI(x) = 1 (up to measure-zero tie sets), not a constant difference. The mirrored appearance of the EEE and EEH plots follows from the sum relation. This error should be corrected, and the sentence concluding that the indexes are 'linearly dependent' should be rephrased accordingly.","section":"Section 3.1, Eq. (10)"},{"comment":"The evaluation protocol does not specify the reference sample used to compute MHI and MEI for test and cross-validation curves. Equation (12) defines the EE-plot using x in X union Y, which is natural for visualization, but if a test curve is included in the reference empirical distribution used to compute its own MHI or MEI, the feature is mechanically biased toward the reference class and the reported accuracies in Figures 4, 6, 8, 10, and 12 and Table 2 are optimistically biased. The authors must state explicitly that, during cross-validation and testing, each curve's indexes are computed with respect to class-specific training reference samples only. Without this clarification, the central empirical claim is not interpretable.","section":"Section 4.1 and Eq. (12)"},{"comment":"The phrase '100-Fold cross-validation as train control' is not a standard protocol and is undefined. The authors should specify how the folds are constructed, whether the 80% training sample is further split into folds, how the 20% test sample is held out, and how hyperparameters for kNN, SVM, and random forest are selected. This information is necessary for reproducibility and for interpreting the boxplots in Figures 4, 6, 8, 10, and 12.","section":"Section 4.1, cross-validation description"},{"comment":"The reported S&P 500 test results are internally inconsistent. For a 14-month test period, an accuracy of 0.71 for RF/MEI corresponds to 10 correct predictions, which matches the statement that 10 of 14 increasing months were correctly predicted; however, the text also describes the accuracy as '~60%', and Table 2 reports 0.64 for LDA/MEI, not 0.60. Additionally, the phrase 'out of the 14 months in which the market increases' should be clarified: is the test set 14 months total, or are there 14 increasing months within a larger test set? These numbers need to be reconciled.","section":"Section 5, Table 2"}],"minor_comments":[{"comment":"The definitions of hypograph and epigraph in Eq. (8) are reversed relative to the standard conventions and to the text immediately preceding them: the display writes hyp(x) = {y > x(t)} and epi(x) = {y <= x(t)}, while the text and Eq. (9) use the conventional definitions. This display should be corrected.","section":"Section 2.2, Eq. (8)"},{"comment":"The centerline expressions 'sin(pi)' and 'x2' need clarification: as written, sin(pi) is identically zero, and x2 should be written as x^2 or x(t)^2. The domain of t should also be stated explicitly.","section":"Table 1"},{"comment":"The description of the synthetic data generation is incomplete: 'generate gauss fdata' and 'exp cov' are not fully documented, and the paper does not specify how the 100 data points per curve are obtained from the covariance operator. A brief algorithmic description or a reference to the R package documentation would improve reproducibility.","section":"Section 4.1"},{"comment":"The axis labels in Figure 11(b) and 11(c) both read 'MEI G1' and 'MEI G2', although panel (c) is described as the EEH plot. Please check the axis labels in all EEH plots for consistent labeling.","section":"Figures 11(b) and 11(c)"},{"comment":"The sentence 'we anticipate a definite market increase for the years 2021-2023' is not supported by the probabilistic classifier outputs and appears to be a forward-looking assertion. It should be replaced with a factual statement about the test-period predictions.","section":"Section 5"},{"comment":"There are several typos and inconsistent notations, e.g., 'coned' for 'coined', 'mithochondiral' for 'mitochondrial', and 'Wavelenghts' in Figure 7. A careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from an explicit statement about data and code availability; no code repository is mentioned. The novelty relative to the DDG-classifier is incremental but reasonable for a methods paper. The main concern is not the methodological idea itself but the evaluation protocol and the incorrect mathematical relation in Section 3.1; both are fixable within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the DD-classifier recipe with depth swapped for modified epigraph/hypograph indexes. That is a real but incremental step, and the paper is honest about it. The EE-plot is a natural visualization, and the comparison against the DDG-classifier across synthetic and three real datasets is sensible. Credit where due: the literature review is competent, the experimental design (six synthetic settings, Berkeley, Tecator, MCO) is thorough enough for a methods note, and the authors are careful to benchmark against an existing classifier rather than claim superiority on all fronts.\n\nThe problems are real but fixable. First, the stated relation MHI(x) - MEI(x) = 1/n is simply false. From equation (10), MHI + MEI equals 1 plus a tie term, so the difference is not constant. The authors even rely on this to justify that EEH and EEE are mirror images; that part holds qualitatively but the formula needs correcting.\n\nSecond, and more seriously, the evaluation protocol never specifies how test curves enter the MHI/MEI computation. Equation (12) defines the EE-plot with x in {X union Y}, which is fine for visualization, but if a test curve is included in the reference empirical distribution used to compute its own index, that is direct label leakage and the accuracies in Table 2 and Figures 4, 6, 8, 10, and 12 are optimistic. The paper must state that test features are computed against training-only reference samples. Without that, the main empirical claim is uninterpretable.\n\nThird, the S&P application is overinterpreted. Fourteen months of test data is tiny; Table 2 shows most methods around 50%, with RF at 0.71. The text smoothly shifts from 'most metrics hover around ~50%' to '~60% is a good result' and then 'we anticipate a definite market increase.' That is not supported. The authors should either present it as an illustrative case or drop the predictive language.\n\nThe circularity burden is low and the method is nonparametric, so the conceptual core survives. If the authors clarify the CV protocol, fix the math relation, and tone down the financial claims, this is a solid incremental contribution to functional data classification. It deserves a serious referee; I would send it to review with those requests rather than desk reject.","headline":"A reasonable depth-classifier analog whose empirical claims rest on an underspecified cross-validation protocol; fix that and the math slip, and it's a publishable incremental contribution.","tokens_in":15586,"tokens_out":2807,"would_cite":false,"duration_ms":25952,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62R10","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that extremality indexes—how often a curve lies above or below other curves—can replace depth as the basis for classifying functional data, with accuracy competitive against the DDG-classifier and around 60% out-of-sample…","keywords":["functional data","supervised classification","extremality","epigraph index","hypograph index","EE-plot","functional depth","S&P 500"],"falsifier":"Re-run the synthetic and real experiments under a strict protocol in which the modified epigraph and hypograph indexes are recomputed using only the curves in the current training fold, with test curves scored against that same training reference, and compare the resulting accuracy distributions with those reported. A concrete check is to compute test-curve index values once against the training curves only and once against the pooled training-plus-test curves; if the second version systematically raises accuracy, the reported advantage is driven by information leakage.","tokens_in":14676,"feed_emoji":"📊","tokens_out":10785,"duration_ms":99718,"temperature":0.7,"pith_summary":"This paper sets out to show that extremality indexes can do for functional-data classification what depth has already been shown to do. The proposed EE-classifier assigns to each curve a pair of numbers—its modified epigraph index or modified hypograph index computed with each of two reference groups as the reference—and plots those points in a two-dimensional EE-plot; standard classifiers such as LDA, QDA, kNN, SVM, and random forests then separate the groups in that planar space. The authors argue that the visual patterns in the EE-plot encode distributional similarity, with a diagonal cluster for identical distributions, chromosome-like shapes for overlapping sinusoids, and L-shapes for non-overlapping curves. They report accuracy on six synthetic experiments and three real functional datasets that is competitive with the depth-based DDG-classifier, often with tighter interquartile spreads, and an application to S&P 500 direction prediction with reported out-of-sample accuracies ranging from about 0.43 to 0.71 across configurations.","feed_headline":"Extremality indexes classify functional data as well as depth methods","feed_subtitle":"Mapping each curve into a two-index square lets off-the-shelf classifiers separate functional groups, rivaling depth-based benchmarks.","key_machinery":"The central objects are the modified epigraph index and modified hypograph index. For curves $x_1,\\dots,x_n$ on an interval $I$, the modified hypograph index $MHI(x) = \\frac{1}{n\\lambda(I)}\\sum_{i=1}^n \\lambda\\{t\\in I: x_i(t)\\le x(t)\\}$ measures the average fraction of time other sample curves fall at or below $x$, and $MEI(x) = \\frac{1}{n\\lambda(I)}\\sum_{i=1}^n \\lambda\\{t\\in I: x_i(t)\\ge x(t)\\}$ measures the average fraction of time they fall at or above $x$. Each index lies in $[0,1]$, and the two are linearly dependent, $MHI(x)-MEI(x)=1/n$. The EE-classifier computes one of these indexes twice for every curve—once with the first group as reference and once with the second—to produce the two-dimensional EE-plot, and that reduction to a planar space is what lets ordinary classifiers such as LDA, QDA, kNN, SVM, and random forests be applied directly.","core_discovery":"On the paper's own terms, the discovery is that the ordering information contained in the modified epigraph index and modified hypograph index is sufficient for supervised classification of functional data, without relying on a notion of depth. For two groups with distributions F and G, every curve is mapped to a point such as (MEI_F(x), MEI_G(x)) or (MHI_F(x), MHI_G(x)); because the two indexes are linearly dependent (MHI(x) − MEI(x) = 1/n for a sample of size n), the epigraph and hypograph plots are mirrored versions of each other. The resulting EE-plot turns functional classification into a planar classification problem, and the paper finds that off-the-shelf classifiers achieve high accuracy in controlled experiments using curves with different centerlines, amplitudes, and dispersions, as well as on the Berkeley growth, Tecator, and MCO datasets. The comparison with the DDG-classifier is meant to show that extremality indexes are a practical alternative, matching the benchmark on most tasks and often showing smaller spreads in cross-validation accuracy.","pith_inferences":["Beyond the paper: the reported EE-plot shape patterns—diagonal, chromosome-like, and L-shaped—suggest the same construction could serve as a simple visual homogeneity diagnostic for two functional samples, complementing existing rank tests.","Beyond the paper: because MHI and MEI differ by the constant $1/n$, the two-dimensional EE-plot is essentially a one-dimensional signal plus its reflection; a natural extension is to test whether a single coordinate based on the diagonal or the difference between the indexes suffices for classification.","Beyond the paper: the paper develops only the two-class case; a strict multiclass generalization would need a strategy such as one-vs-rest index pairs, and the current results give no direct evidence about how accuracy would degrade in that setting.","Beyond the paper: the S&P 500 test window is short, and the paper itself notes that expanding the training history improved metrics; an external check would be to benchmark the EE-classifier against a majority-class or no-change predictor on the same months, since directional stock prediction is difficult to beat."],"forward_implications":["Functional-data classification can be built on extremality alone: no depth notion is required, and the input to the classifier is just a pair of index values per curve.","Because the EE-plot lives in two dimensions, any off-the-shelf classifier can be applied, so the method inherits the flexibility of planar classification without needing functional versions of those algorithms.","The linear dependence between the two indexes means the epigraph and hypograph constructions carry the same information up to reflection, and the empirical results support using either, with MHI occasionally giving better accuracy.","The experiments give concrete guidance: LDA is especially effective for non-overlapping groups, while QDA is generally the strongest overall performer among the tested classifiers.","For the S&P 500 direction task, the best EE-classifier test configuration (random forest on the MEI plot) correctly identifies 10 of the 14 up months in the 2022–2023 test window, indicating that the method can extract some signal from heavily overlapping stock-value curves."],"supporting_citations":[{"why":"Supplies the depth-depth classification template that the EE-classifier adapts to extremality indexes.","marker":"[8]"},{"why":"Defines the DDG-classifier, the functional-depth benchmark used for comparison and the planar classification step the EE-classifier borrows.","marker":"[9]"},{"why":"Introduces the epigraph and hypograph indexes as an ordering for functional data, the theoretical foundation of the proposed method.","marker":"[10]"},{"why":"Defines the DD-plot, the depth-depth visualization whose construction the EE-plot generalizes.","marker":"[15]"},{"why":"Provides the modified epigraph and hypograph indexes (MEI and MHI) that the classifier uses as features.","marker":"[18]"},{"why":"Records the linear dependence between the two indexes and shows they can order closely spaced curves, supporting their use in the EE-plot.","marker":"[14]"},{"why":"Furnishes the S&P 500 up/down classification design that the case study follows.","marker":"[22]"}],"fun_headline_variants":["EE-Classifier: functional data classification without depth","Mapping curves to a 2-D square rivals depth-based classifiers","Turn functional classification into a planar problem with EE-plots","Classify functional curves using extremality, not depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that, during cross-validation and testing, each curve's modified epigraph or hypograph index is computed relative only to the training reference sample; the paper does not explicitly state this, and computing the indexes on the pooled train-plus-test sample would leak test information into the features and inflate the reported accuracies.","fun_headline_variants_meta":{"raw":{"variants":["EE-Classifier: functional data classification without depth","Mapping curves to a 2-D square rivals depth-based classifiers","Turn functional classification into a planar problem with EE-plots","Classify functional curves using extremality, not depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1368,"prompt_tokens":911,"completion_tokens":457,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":527,"tokens_out":457,"duration_ms":4831,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:37:47.354491+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the synthetic and real experiments under a strict protocol in which the modified epigraph and hypograph indexes are recomputed using only the curves in the current training fold, with test curves scored against that same training reference, and compare the resulting accuracy distributions with those reported. A concrete check is to compute test-curve index values once against the training curves only and once against the pooled training-plus-test curves; if the second version systematically raises accuracy, the reported advantage is driven by information leakage.","supporting_citations":[{"cited_title":"Springer, New York (2021) 21 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 MEI G1 MEI G2 Group 0 1","cited_arxiv_id":null,"evidence_quote":"Furnishes the S&P 500 up/down classification design that the case study follows."},{"cited_title":"TEST 26, 119–142 (2017) https://doi.org/10.1007/s11749- 016-0502-6","cited_arxiv_id":null,"evidence_quote":"Defines the DDG-classifier, the functional-depth benchmark used for comparison and the planar classification step the EE-classifier borrows."},{"cited_title":"Recent advances in functional data analysis and related topics 14, 651–676 (2011)","cited_arxiv_id":null,"evidence_quote":"Introduces the epigraph and hypograph indexes as an ordering for functional data, the theoretical foundation of the proposed method."},{"cited_title":"Compu- tational Statistics & Data Analysis 55(4), 1679–1695 (2011) https://doi.org/10","cited_arxiv_id":null,"evidence_quote":"Provides the modified epigraph and hypograph indexes (MEI and MHI) that the classifier uses as features."}],"review_version":1}