{"id":"5fd11b80-4a94-48f7-b438-f80f5bd3d791","arxiv_id":"2412.17341","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"HDTSA is a new R package that bundles six published methods for modeling and testing high-dimensional time series into one integrated tool.","lead":"This paper introduces HDTSA, an R package for high-dimensional time series analysis available on CRAN. It bundles factor models, PCA, CP decomposition, cointegration, and two high-dimensional tests into one tool.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Package correctness is unvalidated; a Monte Carlo size/power check of the bootstrap tests and a ground-truth recovery check for CP_MTS would settle whether the claimed implementations are correct.","rationale":"The reader's weakest assumption is exactly that the code correctly implements the published algorithms, and the paper provides no direct evidence for this beyond the existence of the package and illustrative outputs. That is the most load-bearing premise: if a coding error exists in any of the six functions, the central claim that HDTSA provides a general framework implementing these methods is compromised. The paper's examples are not sufficient validation because they report estimated quantities without comparing them to known ground truth (e.g., Example 3 reports a segmentation but not whether it matches the true latent groups, and Example 6 reports a single p-value rather than repeated null-distribution behavior). The proposed Monte Carlo check is cheap, decisive, and directly tied to the paper's own simulation setups. Since the package is on CRAN, reproducible code exists, and the statistical methods themselves are established, this concern does not justify rejection; it strengthens the case for the reader's CONDITIONAL verdict, which asks for additional validation in documentation or future revisions.","tokens_in":30113,"tokens_out":8666,"duration_ms":89070,"concrete_test":"Run a validation suite using the paper's own DGPs: (1) for WN_test() and MartG_test(), generate n=200, p=10 i.i.d. N(0,I_p) data, run B=200 bootstrap replications, repeat 1000 times, and check that empirical rejection rates at α=0.05 lie in [0.03, 0.07]; (2) for CP_MTS(), replicate Example 4 with known d=d1=d2=3 over 100 seeds and check that the estimated ranks are correct and the canonical angles between hatA, hatB and the true loading spaces are near 0. A failure in either part means the corresponding function does not implement the published method as claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that HDTSA correctly implements four modeling and two inference methods from the cited literature. This requires the R/C++ code to match the published algorithms. The paper provides no test suite, no comparison against reference implementations, and no Monte Carlo evaluation; the examples only show that functions run and produce plausible numbers. For WN_test and MartG_test (§2.5–2.6), the bootstrap critical values depend on several error-prone steps: forming the long vector f_t, generating η_i from N(0,Θ), multiplying by I_K⊗hatΩ, and aggregating block maxima. A sign, lag, or dimension error in any step would make every p-value wrong. For CP_MTS (§2.3), the generalized eigenequation K2,q b = λ K1,q b with truncated tilde-K1,q is numerically delicate; the paper even notes the rank-reduced solver is complex. Without ground-truth recovery checks, code existence does not establish implementation correctness. This is a correctness risk, not a disagreement with the underlying statistical theory.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HDTSA, an R package available on CRAN that claims to provide a general framework for high-dimensional time series analysis. It implements four dimension-reduction methods (factor models, PCA, CP-decomposition, and cointegration analysis) and two high-dimensional inference tests (white noise and martingale difference), with computational acceleration via RcppEigen. The manuscript gives a concise review of the underlying methodology, illustrates each main function with seeded simulated examples, and reports two real-data applications (Fama-French returns and U.S. Industrial Production indices). The central claim is that HDTSA correctly and efficiently implements the published algorithms from the cited literature.","tokens_in":30277,"tokens_out":4451,"duration_ms":48731,"significance":"If the implementation is correct, the package fills a clear gap: no existing R package integrates these six high-dimensional time-series tools, and several of the methods are recent with theoretical guarantees. The paper's strengths include a CRAN-available package, reproducible examples with set.seed, and coverage of both vector and matrix time series. However, the manuscript does not provide quantitative evidence that the implementations are correct: the examples only show that the functions run and produce plausible numbers. A Monte Carlo study with known ground truth and a comparison against reference implementations would materially raise the value of the paper. The significance of the contribution is therefore conditional on such validation.","major_comments":[{"comment":"The paper provides no Monte Carlo or ground-truth validation that the implementations match the published algorithms. For WN_test and MartG_test, the bootstrap critical values depend on several error-prone steps: forming f_t, generating η_i from N(0,Θ), multiplying by I_K⊗Ω̂, and aggregating block maxima. A sign, lag, or dimension error in any of these steps would make every printed p-value wrong, yet the paper reports only one seeded example for each test. Please add a simulation study reporting empirical size and power for the tests (e.g., under the null and under linear and nonlinear alternatives) and, for the modeling functions, recovery rates for ranks and estimation errors for loadings over repeated replications.","section":"§3 (Examples 1–6) and §2.5–2.6"},{"comment":"The CP_MTS example reports that d̂=3 for the direct and refined methods and d̂1=d̂2=d̂=3 for the unified method, all from a single seed. The paper itself notes that solving the generalized eigenequation with rank-reduced matrices is numerically delicate, and the truncated matrix rK1,q in (15) is central to the direct method. Because the output shows only the estimated rank, not the accuracy of the estimated loading matrices  and B̂ or the latent factors, the example does not establish that the code correctly implements the CP-decomposition algorithms. Please include repeated-simulation summaries of rank-selection accuracy and estimation error (e.g., subspace distances for the loading spaces) for all three methods, including the rank-deficient case d1≠d or d2≠d.","section":"§3.3, Example 4 (CP_MTS)"},{"comment":"The weak-factor demonstration reports that the standard procedure gives r̂=2 and the two-step procedure gives r̂=3 for a single generated dataset. As printed, the claim 'we may get r̂=2' is seed-dependent and could be a property of that particular draw rather than a systematic behavior of the estimator. Please provide a Monte Carlo summary over seeds, with the proportion of replications in which each method recovers the true number of factors, to support the claim that the two-step procedure is needed and effective.","section":"§3.1, Example 1"}],"minor_comments":[{"comment":"The package name 'HDRF A' (in the list of existing packages) appears to be a typo; the CRAN package is 'HDRFA'.","section":"§1, footnote on package names"},{"comment":"The kernel matrix Θ is described via its (i,j)-th element K((i−j)/b_n), but the kernel function K is not explicitly defined. Please give the exact forms used for the QS, Parzen, and Bartlett kernels or refer to the specific equations in Andrews (1991).","section":"§2.5, after Eq. (23)"},{"comment":"The text says the eigenvectors solve 'ˆK2,q b = λ ˜K1,q b', but the truncated matrix defined in (15) is denoted rK1,q. Please align the notation to avoid confusion between the truncated and untruncated matrices.","section":"§2.3, after Eq. (15)"},{"comment":"In the code block, the initial assignment 'X <- mat.or.vec(n, r)' is immediately overwritten by 'X <- t(cbind(x1, x2, x3))'; the first assignment is unnecessary and may mislead readers about the intended dimension of X.","section":"§3, Example 1 code"},{"comment":"The argument description for WN_test states that 'pre' uses time series PCA but does not state that the PCA-transformed series are used in the test statistic; please clarify the exact role of the PCA preprocessing.","section":"§3.5, WN_test and MartG_test documentation"},{"comment":"The forecasting comparison in Section 4.1 is qualitative: the paper shows one-step-ahead predictions from three methods but does not evaluate them against the actual January 2022 outcome or report any holdout error metric. Adding a small out-of-sample comparison would support the claim that the package 'can help users ... make reliable predictions'.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is essentially a software description with illustrative examples; the methodological theory is already published, mostly by the same authors. The central risk is implementation correctness, and the current version does not provide enough evidence to rule out coding errors in the bootstrap procedures or the generalized-eigenvalue computations. I would not reject the paper, but the revision should include a proper validation section with Monte Carlo simulations and ground-truth recovery checks. I would also encourage the authors to make available a test suite or unit tests in the package repository so that the correctness claims are verifiable by users."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real R package on CRAN that puts six high-dimensional time series methods (factor models, PCA, CP-decomposition, cointegration, white noise test, martingale difference test) under one interface. That integration is new; no other R package spans this exact set. The paper documents it clearly and the examples are reproducible with seeds. The weak spot is that none of the implementations are validated numerically: no Monte Carlo size/power, no ground-truth recovery, no comparison against existing packages. We know the functions run; we don't know the bootstrap p-values are right.\n\nCredit where due: the package is well structured, uses RcppEigen for the heavy matrix work, and ships with two plausible real-data applications. The methods all come from the authors' own published papers; that is an inward-looking citation pattern, but in a software paper it is not a flaw – implementing your own prior methods is the point. The paper even flags limitations in the text, like the rank condition in CP models, which is a good sign.\n\nThe substantive concern the reader raises is correct: the white noise and martingale difference bootstrap critical values pass through several delicate steps (forming f_t, multiplying by I_K⊗Ω, aggregating block maxima), and a dimension or sign error would silently corrupt every p-value. CP_MTS's generalized eigenequation with a truncated matrix is likewise numerically delicate. Without a test suite or Monte Carlo checks, the package's correctness remains an assertion, not a demonstrated fact. This is a gap, not a fatal flaw: the examples show the functions at least execute and produce sensible rank estimates (r=3, d=1, etc.). Still, for a software paper of this scope, a validation section is standard and its absence is the main reason this is conditional, not accept.\n\nMinor issues: a couple of code typos in the real-data section (res_factors$A should be res_cp$A in the CP block) and a few spots where the prose overstates compatibility claims. Easy fixes.\n\nBottom line: this paper is for R users and applied statisticians in finance and economics, and it is a useful artifact. It deserves a serious referee; send it out with a request for Monte Carlo validation and a correction of the typos. I would engage with it again after those revisions.","headline":"Useful integration of six high-dimensional time series methods in an R package; correctness is plausible but unvalidated.","tokens_in":30795,"tokens_out":3967,"would_cite":true,"duration_ms":39152,"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":"The paper presents HDTSA, an R package that assembles four dimension-reduction methods and two high-dimensional inference tests into a single framework, and argues it is the first such integrated toolbox for analyzing time series whose…","keywords":["CP-decomposition","cointegration analysis","factor model","high-dimensional time series","martingale difference test","principal component analysis","R","white noise test"],"falsifier":"Reproduce the paper's simulated examples with the stated seeds: Example 1 must return 3 factors with twostep = TRUE, Example 4 must return d1 = d2 = d = 3, and Example 5 must return cointegration rank 3. A mismatch in any of these printed outputs would show that the package does not implement the published algorithms as claimed.","tokens_in":29910,"feed_emoji":"📦","tokens_out":6951,"duration_ms":65263,"temperature":0.7,"pith_summary":"HDTSA is an R package that unifies four dimension-reduction methods—factor models, time-series PCA, CP-decomposition of matrix time series, and cointegration analysis—with two high-dimensional inference tests: a white noise test and a martingale difference test. The paper's central claim is that this is the first R package to integrate these six tools, so that users can fit latent structure and run diagnostics on vector or matrix time series with dimension p much larger than sample size n. The package uses C++ matrix operations to keep large multiplications fast, and the paper demonstrates each function on simulated data and on real financial and economic series. If the package works as described, it closes a practical gap: previously these methods were scattered across separate packages or not available in R at all.","feed_headline":"One R package now unites six high-dimensional time-series tools","feed_subtitle":"HDTSA wraps factor, PCA, CP-decomposition, cointegration, and two high-dimensional tests into one toolbox.","key_machinery":"The carrying object is the package's eigenanalysis of lagged autocovariance matrices, which underlies all four modeling tools. For factor and PCA methods, the p-by-p matrix $\\hat{W} = \\sum_{k=1}^K \\hat{\\Sigma}_y(k)\\hat{\\Sigma}_y(k)^\\top$ is thresholded and its eigenvectors span the estimated loading space, with ranks chosen by ratios of successive eigenvalues; CP-decomposition applies generalized eigen-solvers $\\hat{K}_{2,q} b = \\lambda \\hat{K}_{1,q} b$, or a refined d-dimensional version, to estimate the two loading matrices; cointegration estimates the cointegration space from eigenvectors of $\\check{W}$ associated with the smallest eigenvalues. For inference, the two tests are max-type statistics over lagged cross-correlations ($T_{n,\\mathrm{WN}}$) and a sum-of-max statistic over lagged moments ($T_{n,\\mathrm{MDS}}$), with critical values obtained by a Gaussian bootstrap governed by a kernel-estimated long-run covariance matrix. C++ integration via RcppEigen supplies the fast matrix kernel that makes the repeated high-dimensional computations practical.","core_discovery":"The paper's contribution is the HDTSA package and the claim that it fills an existing gap: no R package, and to the authors' knowledge no package outside R, integrates high-dimensional time-series dimension reduction and statistical inference. Four modeling functions—Factors() and HDSReg() for factor models with and without observed regressors, PCA_TS() for time-series PCA segmentation, CP_MTS() for matrix time-series CP-factor estimation via direct, refined, and unified methods, and Coint() for cointegration rank selection—are paired with WN_test() and MartG_test(), bootstrap-based tests for white noise and martingale difference properties that remain valid when p can be much larger than n. For each method, the paper connects the R function to a specific published estimator and states the theoretical guarantee that justifies it.","pith_inferences":["Editorial inference: the absence of numerical accuracy checks means the paper's examples are demonstrations, not verification; a user relying on HDTSA for research should run their own simulations or cross-check against the original method papers.","Editorial inference: because CP_MTS unifies estimation of the ranks d1, d2, and d, it may also serve as a general tool for tensor factor model selection beyond the specific matrix time-series setting presented here.","Editorial inference: the user-supplied map $\\phi$ in MartG_test (for example, $\\phi(x) = \\cos(x)$) suggests the package could be used to diagnose nonlinear remaining dependence in fitted models, not just linear autocorrelation.","Editorial inference: a natural next validation step would be a Monte Carlo benchmark comparing HDTSA's outputs against reference implementations on datasets with known factor structure, which the paper does not provide."],"forward_implications":["With one installation, a user can move from fitting a factor model or matrix CP-factor model to checking model adequacy with high-dimensional white-noise and martingale-difference tests, without switching packages.","The p >> n applicability of the six methods means standard R tools that require p ≤ n—such as classical portmanteau tests—can be replaced for large-dimension data.","The C++-based matrix multiplications make repeated high-dimensional computations, including the bootstrap loops in the tests, practical on ordinary hardware.","The predict methods for fitted factor, PCA, and CP objects let forecasters build univariate or low-dimensional forecasts on latent factors and transform them back to the original high-dimensional series."],"supporting_citations":[{"why":"Establishes the latent-factor estimation via eigenanalysis of the cumulative lagged autocovariance matrix, which Factors() implements.","marker":"Lam et al. (2011)"},{"why":"Supplies the ratio-based rank selection and the two-step strong-plus-weak factor procedure used by Factors().","marker":"Lam and Yao (2012)"},{"why":"Provides the time-series PCA segmentation and the permutation tests that PCA_TS() wraps.","marker":"Chang et al. (2018)"},{"why":"Introduces the direct and refined CP-decomposition estimators implemented in CP_MTS().","marker":"Chang et al. (2023a)"},{"why":"Adds the unified estimation of ranks d1, d2, and d that CP_MTS() offers for matrix CP-factor models.","marker":"Chang et al. (2024a)"},{"why":"Gives the eigenanalysis-based cointegration rank identification that Coint() implements.","marker":"Zhang et al. (2019)"},{"why":"Defines the max-type high-dimensional white-noise test statistic and bootstrap critical values used by WN_test().","marker":"Chang et al. (2017)"},{"why":"Defines the high-dimensional martingale difference test and bootstrap procedure used by MartG_test().","marker":"Chang et al. (2023b)"},{"why":"Documents the RcppEigen linear-algebra interface that provides the package's C++ speedups.","marker":"Bates and Eddelbuettel (2013)"}],"fun_headline_variants":["R package packs six high-dimensional time series methods","Six time-series tools, one R package: HDTSA","R package bridges high-dimensional modeling and inference","All-in-one R package for high-dimensional time series","HDTSA: unified R toolkit for high-dimensional time series analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The package's outputs are only meaningful if the R and C++ code correctly implements the published algorithms, yet the paper provides no test suite, no comparison against reference implementations, and no Monte Carlo study to verify the implementation.","fun_headline_variants_meta":{"raw":{"variants":["R package packs six high-dimensional time series methods","Six time-series tools, one R package: HDTSA","R package bridges high-dimensional modeling and inference","All-in-one R package for high-dimensional time series","HDTSA: unified R toolkit for high-dimensional time series analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000972,"raw_usage":{"total_tokens":4092,"prompt_tokens":865,"completion_tokens":3227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":3149}},"tokens_in":481,"tokens_out":3227,"duration_ms":24355,"temperature":1.0,"reasoning_tokens":3149,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:33:38.865248+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the paper's simulated examples with the stated seeds: Example 1 must return 3 factors with twostep = TRUE, Example 4 must return d1 = d2 = d = 3, and Example 5 must return cointegration rank 3. A mismatch in any of these printed outputs would show that the package does not implement the published algorithms as claimed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the latent-factor estimation via eigenanalysis of the cumulative lagged autocovariance matrix, which Factors() implements."},{"cited_title":"and Yao, Q","cited_arxiv_id":null,"evidence_quote":"Supplies the ratio-based rank selection and the two-step strong-plus-weak factor procedure used by Factors()."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the time-series PCA segmentation and the permutation tests that PCA_TS() wraps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the eigenanalysis-based cointegration rank identification that Coint() implements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the max-type high-dimensional white-noise test statistic and bootstrap critical values used by WN_test()."},{"cited_title":"and Eddelbuettel, D","cited_arxiv_id":null,"evidence_quote":"Documents the RcppEigen linear-algebra interface that provides the package's C++ speedups."}],"review_version":1}