Pith. sign in

REVIEW 4 major objections 4 minor 85 references

The paper claims that replacing fixed linear frequency-alignment rules with shared deep ReLU networks—a depth-separable neural network (DSNN)—can improve mixed-frequency forecasting, and it proves a non-asymptotic excess-risk bound for the

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A parameter-shared, three-stage deep ReLU architecture for mixed-frequency alignment and forecasting, with approximation theory and non-asymptotic excess-risk bounds, outperforms MIDAS-style baselines in simulations and on US macro data.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection The DSNN architecture is worth discussing, but the paper's central theorem is not proven as written—the proof of Proposition 1 ignores the stacking structure, so the claimed rate does not follow. the 4 major comments →

arxiv 2607.14771 v1 pith:YU65NFQC submitted 2026-07-16 stat.ME

Mixed-Frequency Time Series Forecasting via Depth-Separable Neural Networks

Yize Wang , Qianqian Zhu , Guodong Li This is my paper
classification stat.ME
keywords mixed-frequency time seriesdepth-separable neural networksfrequency alignmentnon-asymptotic prediction error boundReLU networkshierarchical composition modelsparameter sharingleast squares estimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that the frequency-alignment step in mixed-frequency forecasting should be learned, not prespecified. It proposes a depth-separable neural network that sequentially aligns high-frequency data to medium, then to low frequency, using shared deep ReLU networks, before a final neural network forecasts the low-frequency response. The theoretical core is a non-asymptotic bound on prediction error: under hierarchical composition assumptions, the excess risk decays at a near-optimal rate with sample size. Empirical results on simulations and on forecasting U.S. quarterly macro variables with monthly and daily predictors show lower forecast errors than existing linear and nonlinear mixed-frequency methods. If correct, this would give practitioners a flexible, scalable, and theoretically grounded alternative to MIDAS-style alignment.

Core claim

The central discovery is that the two alignment functions—high-to-medium and medium-to-low—can be implemented as parameter-shared deep ReLU networks within a composed architecture, and this composition has provable approximation and generalization properties. The paper establishes that the least-squares DSNN estimator achieves a non-asymptotic excess-risk bound of order (log^5(T)/T^{ν0})^{2γ*/(2γ*+1)} plus a weak-dependence term, provided the true regression function lies in a hierarchical composition class with the same factor dimensions and stacking structure. Empirically, the DSNN reports lower RMSFE than VARX, RNN/LSTM/DNN baselines, and several MIDAS variants across all simulation setti

What carries the argument

The key machinery is the depth-separable composition m = g3,NN ∘ g2,NN ∘ g1,NN, where each g_j,NN is a deep ReLU network and the composition is defined through stacking operations. Parameter sharing across the many series in each alignment stage reduces the effective number of parameters, making the architecture scalable to large sets of high-frequency predictors. The main theoretical tool is an error-propagation bound that expresses the total approximation error as a sum of per-stage errors weighted by factor dimensions and weight-norm products, which leads to the overall excess-risk rate.

Load-bearing premise

The entire theory rests on the assumption that the true regression function is exactly a hierarchical composition m0 = g3 * g2 * g1 with the same stacking layout and the same factor dimensions (r1, r2) as the fitted DSNN; if the data-generating process does not match this structure, the approximation bound and the excess-risk rate are not guaranteed.

What would settle it

Simulate data from a true function that is not of the hierarchical composition form—for instance, give each high-frequency predictor its own distinct nonlinear alignment function that cannot be represented by a single shared g1—and then evaluate whether the empirical excess risk of the least-squares DSNN stays below or near the bound, or instead degrades noticeably as the sample size grows.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If DSNN is correct as claimed, forecasters can replace pre-specified linear lag polynomials in mixed-frequency settings with adaptive nonlinear alignments, potentially capturing threshold and interaction effects that linear methods miss.
  • The proven non-asymptotic rate suggests that, with sufficient sample size, the prediction error of DSNN approaches near-optimal nonparametric rates even under weak temporal dependence, providing a theoretical safety net for its empirical use.
  • Parameter sharing across alignment networks implies the method can handle many high-frequency predictors—such as dozens of daily financial series—without the curse of dimensionality that plagues simple stacking.
  • The architecture unifies existing approaches: with linear constraints it recovers MIDAS regressions, and with identity alignment it reduces to stacked-series regression, offering a single flexible framework.
  • If the bound holds in practice, it opens the door to hyperparameter selection based on the rate, such as choosing factor dimensions r1 and r2 to balance approximation and generalization terms.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The theory depends crucially on the assumption that the true function has exactly the same hierarchical composition structure and factor dimensions as the fitted DSNN; if the true alignment differs across predictors, the approximation bound may not hold and the excess-risk rate could degrade—this is a testable limitation.
  • An obvious extension, not explored in the paper, is to estimate the factor dimensions r1 and r2 adaptively from data rather than by cross-validation, perhaps through a sparsity-penalized variant of the least-squares objective.
  • The same depth-separable alignment idea could be ported to other mixed-frequency problems, such as daily-to-monthly volatility forecasting, where one might compare against GARCH-MIDAS models to see whether the nonlinear alignment gains carry over.
  • The paper's empirical comparison leaves open a distribution-free or quantile version of DSNN, which would be useful for asymmetric loss functions in macroeconomic risk forecasting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper proposes a depth-separable neural network (DSNN) for forecasting a low-frequency target from medium- and high-frequency predictors. The architecture uses shared ReLU alignment networks g1 and g2 followed by a prediction network g3, with the star product notation in Eq. (2.6) denoting stacking and blockwise repetition. The paper claims an approximation bound (Proposition 1), a generalization bound (Proposition 2), and a non-asymptotic excess-risk bound (Theorem 1). It then reports simulations and an application to forecasting U.S. quarterly macroeconomic variables. The central theoretical claim is the non-asymptotic rate for the DSNN estimator, which is asserted to be the first of its kind for a neural mixed-frequency alignment architecture.

Significance. If Theorem 1 were correct, the paper would make a genuine theoretical contribution by giving the first non-asymptotic prediction-error rate for a neural mixed-frequency alignment-and-forecast architecture. The paper is also strong in scope: it provides a full Appendix with proofs, a detailed simulation study, and an empirical application with a public code link. However, the main theorem is currently not supported by the proof as written. The approximation proof of Proposition 1 does not account for the blockwise, stacked structure of the DSNN, and the proof of Theorem 1 makes a separability assumption that is not justified by the stated architecture choices. The empirical superiority claim is based on point forecasts without uncertainty quantification. These issues are load-bearing for the paper's central claims, so substantial revision is needed.

major comments (4)
  1. [A.1.1, Eq. (A.1.1)] The proof of Proposition 1 treats the DSNN as a simple composition of three functions. It sets y1 = g1(x) in R^{r1} and y1' = g1_NN(x) in R^{r1}, then propagates the error through a single g2_NN: R^Q -> R^{r2}. But the DSNN defined in Eq. (2.6) applies g1_NN separately to each of D1 = Q d3 high-frequency sub-series, producing an r1*D1-dimensional output, and then applies g2_NN separately to each of D2 = d2 + r1 d3 medium-frequency stacked vectors. Thus Eq. (A.1.1) does not describe the actual architecture. The factors sqrt(r1) and sqrt(r2) in bounds (I) and (II) are therefore unjustified; a correct proof must track block-diagonal repetitions and the stacking operators, which would introduce dependence on D1, D2, Q, and S. Consequently, the dimension factors in Theorem 1 and the claimed rate are not established.
  2. [Definition 3] The target class F* is defined as m0 = g3 * g2 * g1, where g1: R^S -> R^{r1}, g2: R^Q -> R^{r2}, and g3 is a function on R^{K D3}. However, the class of functions to be approximated by a DSNN is much larger: the true m0 must map a collection of d3 high-frequency series, d2 medium-frequency series, and lags to the low-frequency response. The star product in Eq. (2.6) includes explicit stacking and D1,D2 copies of g1_NN and g2_NN; this structure is absent from Definition 3. Unless F* is defined as the class of functions admitting exactly the same blockwise decomposition, Proposition 1's approximation statement is ambiguous. The simulation DGP in Section 4 uses additive component functions, but the theorem needs an explicit target class that matches the DSNN architecture.
  3. [A.1.3, proof of Theorem 1] The proof of Theorem 1 contains an unjustified separability step. After deriving the bound with the factor L* = L1 + L2 + L3 multiplying the entire complexity term, the text states 'Without loss of generality, we assume L1, L2 and L3 grow at the same rate, then L* can be treated as a multiple of L1, L2 or L3.' But the theorem then chooses Lbar1*wbar1, Lbar2*wbar2, and Lbar3*wbar3 at different rates, depending on D1, D2, r1, r2, d0, M2, and M3. The optimization problem is not separable in the displayed bound because L* couples the three stages. The final rate is therefore not derived from the preceding bound.
  4. [Section 5, Table 2] The abstract and Section 5 claim that the DSNN 'achieves significant predictive gains' over existing methods. Table 2 reports only point RMSFE and MAFE for a single test window, with no standard errors, confidence intervals, or significance tests. Several margins are small (e.g., Period I, S3: DSNN 0.523 vs. DNN 0.540). The applied claim is thus stronger than the evidence supports. The simulation study also generates the true function from the same hierarchical composition class, so it does not test the method under misspecification of the structural assumption.
minor comments (4)
  1. [Abstract/Section 2] Typographical issues: 'preforms' should be 'performs'; 'gradient decent' should be 'gradient descent'; 'efficient' should be 'efficient'.
  2. [Section 3, after Proposition 2] The text refers to 'The generalization error bound in Theorem 2' but the statement is Proposition 2; the reference should be updated.
  3. [Table 1] The missing MIDAS1 entry at T=300 is explained in the text, but a table note would help the reader.
  4. [Section 2.2 and Definition 3] The star notation * is used informally. It is defined only in Eq. (2.6) with an accompanying diagram; the target class in Definition 3 should restate the block sizes and stacking explicitly.

Circularity Check

0 steps flagged

No significant circularity: the theory is derived from external approximation and weak-dependence results, and the central claims are not equivalent to their assumptions by construction.

full rationale

The paper's derivation chain is not circular. The target class ℱ* (Definition 3) is an explicit structural assumption on the true function m0, stated independently of the DSNN parameterization; it is not defined in terms of the fitted networks. Proposition 1's approximation bound is obtained by combining an external approximation result (Fan et al., 2024, Proposition 3.4, via Lemma A.1) with a Lipschitz error-propagation argument. Proposition 2's generalization bound is derived from covering-number arguments and a weak-dependence concentration inequality (Lemma A.5, following Alquier et al.), again external. Theorem 1 is a standard excess-risk decomposition combining these two bounds; no fitted parameter is relabeled as a prediction, and no hyperparameter or simulation output is inserted into the theorem. The empirical hyperparameters are selected by validation and reported as such, not claimed as theoretical predictions. The simulation DGP is drawn from the same hierarchical composition class used in the theory; this is an experimental design choice, not circular reasoning. The skeptical concern that the proof of Proposition 1 ignores the stacking/block-diagonal structure is a possible correctness or rigor gap, not a circularity: it does not show that the theorem's conclusion is equivalent to its inputs. No self-citation is load-bearing; the only apparent self-citation (Wang et al., 2020, including co-author Li) is cited for the architectural analogy to depth-separable networks, not as the basis of any theorem or uniqueness claim. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

No new physical or latent entities are postulated; the 'factors' are learned network outputs. The theory rests on the hierarchical composition class, boundedness, weak dependence, and the cited ReLU approximation theorem. The empirical claims additionally depend on hand-chosen and cross-validated hyperparameters.

free parameters (4)
  • r1 (first-stage factor dimension) = 5 (Period I), 5 (Period II); CV-selected in simulations
    Chosen by rolling-window validation (Section 2.3); controls whether the first alignment expands or compresses information and enters the theoretical constants.
  • r2 (second-stage factor dimension) = 2 (Period I), 4 (Period II); CV-selected
    Chosen by validation; period-dependent value is used to argue the method adapts (r2>Q=3 in Period II).
  • K (lag order) = 1 in both empirical periods; 2–5 in simulations
    Selected by validation; defines the feature window for the prediction network.
  • Network depth/width, learning rate, batch size, epochs = L1=L2=2, L3=2, w1=w2=64, w3=128, lr=1e-5, batch=64, epochs=200
    Chosen by hand in Section 5; these choices affect the empirical comparison and are not determined by the theory.
axioms (6)
  • domain assumption m0 belongs to the hierarchical composition class F* with fixed levels and factor dimensions (Definition 3; Section 2.2).
    All approximation and excess-risk guarantees are relative to this class; the paper does not test this structural premise on real data.
  • domain assumption Assumption 1: bounded support of x, bounded m0, bounded network weights with product bound M_j.
    Used throughout the approximation, covering-number, and error-propagation proofs.
  • domain assumption Assumption 2: stationarity, ergodicity, and theta_infinity weak dependence with finite coefficients.
    Needed for the Hoeffding-type concentration inequality and the generalization bound.
  • standard math The deep ReLU approximation rate (Lw)^{-2gamma*} for hierarchical composition classes (Fan et al. 2024, Prop 3.4).
    Lemma A.1 extends it to vector outputs but does not re-derive the rate.
  • domain assumption Errors epsilon[t] are i.i.d., zero mean, and independent of past predictors (model (2.1)).
    Used to identify excess risk with squared approximation error.
  • domain assumption The least-squares estimator m_hat_T is the global minimizer over the DSNN function class (Eq. 3.2), although Algorithm 1 is stochastic gradient descent.
    The theorem is about the population minimizer of empirical risk, not necessarily about the local optimum returned by Adam.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Mixed-Frequency Time Series Forecasting via Depth-Separable Neural Networks." pith.science (2026). https://pith.science/paper/YU65NFQC

@misc{pith2026260714771,
  author       = {Pith},
  title        = {Pith review of: Mixed-Frequency Time Series Forecasting via Depth-Separable Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YU65NFQC}},
  note         = {Machine review of arXiv:2607.14771}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

To better forecast mixed-frequency time series, it is the key to choose a suitable way for frequency alignment. However, the existing methods are all limited to linear transformations, and this may overlook the possible nonlinearity, leading to a worse prediction. We alternatively consider a deep neural network for each frequency alignment, and hence a depth-separable neural network. Moreover, a parameter-sharing mechanism is adopted across the alignment at each stage, making possible a deeper network for a large set of higher-frequency predictors. This paper establishes an approximation theory for the proposed depth-separable network, and a non-asymptotic prediction error bound is also derived. Simulation studies demonstrate the finite-sample performance of the proposed method, and an empirical application to forecasting U.S. quarterly macroeconomic variables using monthly and daily indicators, highlights its superior predictive accuracy over existing mixed-frequency methods.

Figures

Figures reproduced from arXiv: 2607.14771 by Guodong Li, Qianqian Zhu, Yize Wang.

Figure 1
Figure 1. Figure 1: Illustration of the proposed Depth-Separable Neural Network (DSNN). The [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

85 extracted references · 2 linked inside Pith

  1. [1]

    2023 , doi =

    Fan, Jianqing and Gu, Yihong , journal =. 2023 , doi =

    2023

  2. [2]

    and Ng, Serena , journal =

    McCracken, Michael W. and Ng, Serena , journal =. 2015 , doi =

    2015

  3. [3]

    2020 , doi =

    McCracken, Michael and Ng, Serena , month =. 2020 , doi =

    2020

  4. [4]

    2022 , doi =

    Xu, Xiaoxiang and Liao, Mingqiu , journal =. 2022 , doi =

    2022

  5. [5]

    2020 , doi =

    Wen, Long and Liu, Chang and Song, Haiyan and Liu, Han , journal =. 2020 , doi =

    2020

  6. [6]

    Journal of Business & Economic Statistics , volume=

    A general multivariate threshold GARCH model with dynamic conditional correlations , author=. Journal of Business & Economic Statistics , volume=. 2011 , publisher=

    2011

  7. [7]

    2016 , publisher=

    Support Vector Machines and Perceptrons: Learning, Optimization, Classification, and Application to Social Networks , author=. 2016 , publisher=

    2016

  8. [8]

    2014 , url =

    Marsilli, Clément , title =. 2014 , url =

    2014

  9. [9]

    2021 , doi =

    Marcellino, Massimiliano and Sivec, Vasja , journal =. 2021 , doi =

    2021

  10. [10]

    2004 , url =

    Ghysels, Eric and Santa-Clara, Pedro and Valkanov, Rossen , month =. 2004 , url =

    2004

  11. [11]

    , journal =

    Ghysels, Eric and Sinko, Arthur and Valkanov, Rossen I. , journal =. 2007 , doi =

    2007

  12. [12]

    2008 , url =

    Clements, Michael and Galvão, Ana , journal =. 2008 , url =

    2008

  13. [13]

    2010 , doi =

    Rodriguez, Abel and Puggioni, Gavino , journal =. 2010 , doi =

    2010

  14. [14]

    2010 , url =

    Andreou, Elena and Ghysels, Eric and Kourtellos, Andros , journal =. 2010 , url =

    2010

  15. [15]

    2010 , doi =

    Frale, Cecilia and Monteforte, Libero , month =. 2010 , doi =

    2010

  16. [16]

    2013 , doi =

    Andreou, Elena and Ghysels, Eric and Kourtellos, Andros , journal =. 2013 , doi =

    2013

  17. [17]

    and Ghysels, Eric and Sohn, Bumjean , journal =

    Engle, Robert F. and Ghysels, Eric and Sohn, Bumjean , journal =. 2013 , doi =

    2013

  18. [18]

    2018 , doi =

    Conrad, Christian and Custovic, Anessa and Ghysels, Eric , journal =. 2018 , doi =

    2018

  19. [19]

    , journal =

    Amendola, Alessandra and Candila, Vincenzo and Gallo, Giampiero M. , journal =. 2018 , doi =

    2018

  20. [20]

    2015 , doi =

    Foroni, Claudia and Marcellino, Massimiliano and Schumacher, Christian , journal =. 2015 , doi =

    2015

  21. [21]

    , journal =

    Ghysels, E. , journal =. 2014 , doi =

    2014

  22. [22]

    2016 , doi =

    Ghysels, Eric and Plazzi, Alberto and Valkanov, Rossen , journal =. 2016 , doi =

    2016

  23. [23]

    2022 , url =

    Babii, Andrii and Ghysels, Eric and Striaukas, Jonas , journal =. 2022 , url =

    2022

  24. [24]

    2018 , doi =

    Xu, Qifa and Zhuo, Xingxuan and Jiang, Cuixia and Liu, Yezheng , journal =. 2018 , doi =

    2018

  25. [25]

    2021 , doi =

    Xu, Qifa and Liu, Shuting and Jiang, Cuixia and Zhuo, Xingxuan , journal =. 2021 , doi =

    2021

  26. [26]

    2024 , doi =

    Saputra, Wisnowan Hendy and Prastyo, Dedy Dwi and Kuswanto, Heri , journal =. 2024 , doi =

    2024

  27. [27]

    2020 , doi =

    Li, Xiangpeng and Yu, Hong and Xie, Yongfang and Li, Jie , journal =. 2020 , doi =

    2020

  28. [28]

    2024 , doi =

    Filho, Leopoldo Lusquino and De Oliveira Werneck, Rafael and Castro, Manuel and Júnior, Pedro Ribeiro Mendes and Lustosa, Augusto and Zampieri, Marcelo and Linares, Oscar and Moura, Renato and Morais, Elayne and Amaral, Murilo and Salavati, Soroor and Loomba, Ashish and Esmin, Ahmed and Gonçalves, Maiara and Schiozer, Denis José and Ferreira, Alexandre an...

    2024

  29. [29]

    2025 , doi =

    Wang, Chunzi and Xie, Fusheng and Yan, Junpeng and Xia, Yiqing , journal =. 2025 , doi =

    2025

  30. [30]

    Statistica Sinica , volume=

    Mallows model averaging estimator for the MIDAS model with Almon polynomial weight , author=. Statistica Sinica , volume=. 2022 , publisher=

    2022

  31. [31]

    International Journal of Forecasting , volume=

    Changes in predictive ability with mixed frequency data , author=. International Journal of Forecasting , volume=. 2013 , publisher=

    2013

  32. [32]

    Oxford Bulletin of Economics and Statistics , volume=

    Regression models with variables of different frequencies: The case of a fixed frequency ratio , author=. Oxford Bulletin of Economics and Statistics , volume=. 2010 , publisher=

    2010

  33. [33]

    Journal of Forecasting , volume=

    Forecasting inflation rates using daily data: A nonparametric MIDAS approach , author=. Journal of Forecasting , volume=. 2015 , publisher=

    2015

  34. [34]

    Electronic Journal of Statistics , volume=

    Fully NonParametric MIDAS: A new approach for nonparametric mixed frequency time series regression , author=. Electronic Journal of Statistics , volume=. 2025 , publisher=

    2025

  35. [35]

    , journal =

    Bai, Jennie and Ghysels, Eric and Wright, Jonathan H. , journal =. 2013 , doi =

    2013

  36. [36]

    2024 , url =

    Glišić, Iva , title =. 2024 , url =

    2024

  37. [37]

    and Kim, Tae Bong and Seoane, Hernán D

    Eraker, Bjørn and Chiu, Ching Wai and Foerster, Andrew T. and Kim, Tae Bong and Seoane, Hernán D. , journal =. 2015 , doi =

    2015

  38. [38]

    and Marcellino, Massimiliano , journal =

    Carriero, Andrea and Clark, Todd E. and Marcellino, Massimiliano , journal =. 2015 , doi =

    2015

  39. [39]

    2016 , doi =

    Qian, Hang , journal =. 2016 , doi =

    2016

  40. [40]

    2021 , url =

    McCracken, Michael and Owyang, Michael and Sekhposyan, Tatevik , journal =. 2021 , url =

    2021

  41. [41]

    and Murasawa, Yasutomo , journal =

    Mariano, Roberto S. and Murasawa, Yasutomo , journal =. 2002 , doi =

    2002

  42. [42]

    2007 , doi =

    Schumacher, Christian and Breitung, Jörg , month =. 2007 , doi =

    2007

  43. [43]

    2008 , doi =

    Giannone, Domenico and Reichlin, Lucrezia and Small, David , journal =. 2008 , doi =

    2008

  44. [44]

    European Central Bank , year=

    Nowcasting with daily data , author =. European Central Bank , year=

  45. [45]

    2022 , doi =

    Peng, Siyang and Guo, Shaojun and Long, Yonghong , journal =. 2022 , doi =

    2022

  46. [46]

    2016 , doi =

    Marcellino, Massimiliano and Porqueddu, Mario and Venditti, Fabrizio , journal =. 2016 , doi =

    2016

  47. [47]

    , month =

    Guo, Zichuan and Cucuringu, Mihai and Shestopaloff, Alexander Y. , month =. 2025 , url =

    2025

  48. [48]

    2024 , doi =

    Fan, Jianqing and Gu, Yihong and Zhou, Wenxin , journal =. 2024 , doi =

    2024

  49. [49]

    2020 , doi =

    Schmidt-Hieber, Johannes , journal =. 2020 , doi =

    2020

  50. [50]

    2021 , doi =

    Kohler, Michael and Langer, Sophie , journal =. 2021 , doi =

    2021

  51. [51]

    2021 , doi =

    Lu, Jianfeng and Shen, Zuowei and Yang, Haizhao and Zhang, Shijun , journal =. 2021 , doi =

    2021

  52. [52]

    2021 , doi =

    Shen, Zuowei and Yang, Haizhao and Zhang, Shijun , journal =. 2021 , doi =

    2021

  53. [53]

    2022 , url =

    Caner, Mehmet and Daniele, Maurizio , month =. 2022 , url =

    2022

  54. [54]

    and Liang, Tengyuan and Misra, Sanjog , journal =

    Farrell, Max H. and Liang, Tengyuan and Misra, Sanjog , journal =. 2021 , doi =

    2021

  55. [55]

    2018 , url =

    Suzuki, Taiji , month =. 2018 , url =

    2018

  56. [56]

    2019 , doi =

    Bauer, Benedikt and Kohler, Michael , journal =. 2019 , doi =

    2019

  57. [57]

    2025 , url =

    Kohler, Michael , month =. 2025 , url =

    2025

  58. [58]

    2002 , doi =

    Györfi, László and Kohler, Michael and Krzyżak, Adam and Walk, Harro , month =. 2002 , doi =

    2002

  59. [59]

    Nature , volume=

    Learning representations by back-propagating errors , author=. Nature , volume=. 1986 , publisher=

    1986

  60. [60]

    Neural Computation , volume=

    Long short-term memory , author=. Neural Computation , volume=. 1997 , publisher=

    1997

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Greedy layer-wise training of deep networks , author=. Advances in Neural Information Processing Systems , volume=

  62. [62]

    arXiv preprint arXiv:1406.1078 , year=

    Learning phrase representations using RNN encoder-decoder for statistical machine translation , author=. arXiv preprint arXiv:1406.1078 , year=

    Pith/arXiv arXiv

  63. [63]

    2014 , url =

    Kingma, Diederik and Ba, Jimmy , month =. 2014 , url =

    2014

  64. [64]

    2016 , doi =

    Ghysels, Eric , journal =. 2016 , doi =

    2016

  65. [65]

    2023 , volume=

    Chakraborty, Nilanjana and Khare, Kshitij and Michailidis, George , journal =. 2023 , volume=. doi:10.5705/ss.202021.0206 , url =

    arXiv 2023

  66. [66]

    2015 , doi =

    Schorfheide, Frank and Song, Dongho , journal =. 2015 , doi =

    2015

  67. [67]

    2019 , doi =

    Tank, A and Fox, E B and Shojaie, A , journal =. 2019 , doi =

    2019

  68. [68]

    2024 , doi =

    Lin, Jiahe and Michailidis, George , journal =. 2024 , doi =

    2024

  69. [69]

    Lecture notes in electrical engineering , month =

    Lee, Tze-Houng and Yuan, Xue-Ming , publisher =. Lecture notes in electrical engineering , month =. 2025 , doi =

    2025

  70. [70]

    José and Louhichi, Sana and Prieur, Clémentine , month =

    Dedecker, Jérôme and Doukhan, Paul and Lang, Gabriel and Rafael, León R. José and Louhichi, Sana and Prieur, Clémentine , month =. 2007 , doi =

    2007

  71. [71]

    2024 , url =

    Ou, Weigutian and Bölcskei, Helmut , month =. 2024 , url =

    2024

  72. [72]

    2019 , publisher=

    High-Dimensional Statistics: A Non-Asymptotic Viewpoint , author=. 2019 , publisher=

    2019

  73. [73]

    2019 , url =

    Chen, Minshuo and Li, Xingguo and Zhao, Tuo , title =. 2019 , url =

    2019

  74. [74]

    and Harvey, Nick and Liaw, Christopher and Mehrabian, Abbas , journal =

    Bartlett, Peter L. and Harvey, Nick and Liaw, Christopher and Mehrabian, Abbas , journal =. 2019 , url =

    2019

  75. [75]

    IEEE Transactions on Information Theory , volume=

    Minimum complexity regression estimation with weakly dependent observations , author=. IEEE Transactions on Information Theory , volume=. 2002 , publisher=

    2002

  76. [76]

    2023 , url =

    Kengne, William and Wade, Modou , month =. 2023 , url =

    2023

  77. [77]

    arXiv preprint arXiv:1211.1847 , year=

    Prediction of time series by statistical learning: general losses and fast rates , author=. arXiv preprint arXiv:1211.1847 , year=

    Pith/arXiv arXiv

  78. [78]

    Journal of Machine Learning Research , volume=

    Over-parameterized deep nonparametric regression for dependent data with its applications to reinforcement learning , author=. Journal of Machine Learning Research , volume=

  79. [79]

    2024 , url =

    Deep autoencoders for nonlinear factor models: Theory and applications , author=. 2024 , url =

    2024

  80. [80]

    Pattern Recognition , volume=

    Financial time series forecasting with multi-modality graph neural network , author=. Pattern Recognition , volume=. 2022 , publisher=

    2022

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.