Pith. sign in

REVIEW 4 major objections 6 minor 34 references

From Local Patterns to Global Understanding: Cross-Stock Trend Integration for Enhanced Predictive Modeling

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a federated-learning-inspired training strategy, CSTI, improves stock price prediction by merging per-stock models into a global model and fine-tuning it back to each stock, beating the sequential baseline in most…

desk verdict Standard FedAvg+fine-tuning applied to stocks, with a pseudocode that doesn't actually implement the iterative merging it claims. read the letter →

arxiv 2505.16573 v1 pith:CNS6DPBB submitted 2025-05-22 cs.CE cs.AI

Yi Hu , Hanchi Ren , Jingjing Deng , Xianghua Xie This is my paper · ORCID
classification cs.CEcs.AI
keywords cross-stocktrendintegrationfederatedlearningstockpricepredictionmodelmergingfine-tuningFNSPIDtimeseriesforecastingparameteraveraging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a federated-learning-style training strategy, called Cross-Stock Trend Integration (CSTI), makes stock price predictors better by letting each model borrow patterns from other stocks. The strategy trains a separate model on each stock's history, averages the model parameters into a shared global model, repeats this merge, and finally fine-tunes the global model on each stock with a penalty that keeps it close to the shared parameters. On the FNSPID dataset, across seven forecasting backbones and group sizes of 5, 25, and 50 stocks, CSTI beats the paper's 'normal training' baseline in most of the reported configurations, with the largest gains for DLinear, PaiFilter, and TexFilter. The paper also reports that CSTI does not always win: normal training keeps the edge for TimesNet and PatchTST at 25 and 50 stocks. If correct, the method offers a model-agnostic way to add cross-stock information to existing forecasters without sharing raw price data.

What carries the argument

The load-bearing object is the CSTI training loop itself, an adaptation of federated averaging with a local-adaptation phase. Phase one trains each stock model $f_k(\cdot;\theta_k)$ in parallel on its own data using mean squared error. Phase two repeatedly forms a global model by parameter averaging, $\theta_g = \frac{1}{K}\sum_{k=1}^{K} w_k \theta_k$ (with $w_k=1$ in all experiments), allowing shared cross-stock patterns to accumulate. Phase three re-initializes every local model at $\theta_g$ and fine-tunes it with the regularized loss $L'_k = \frac{1}{T}\sum_{t=1}^{T}(f_k(X_{k,t};\theta_k)-y_{k,t})^2 + \lambda\|\theta_k-\theta_g\|^2$ and a learning rate scaled by $\alpha$, so each final model stays close to the shared representation while recovering stock-specific behavior. The L2 penalty and the scaled learning rate are the controls that balance global generalization against local fit.

What would settle it

Re-run the reported experiments with a strictly per-stock baseline in which each of the seven models is trained only on its own stock's data for the same number of update steps, and compare on the same FNSPID splits; if CSTI no longer beats that baseline in the majority of configurations, the paper's central claim is not supported.

Watch

Extended reading notes

Core claim

The authors establish that a training procedure modeled on federated averaging can inject cross-stock information into stock price predictors without changing the underlying architecture. In CSTI, each stock gets its own model trained on local price and sentiment features; after one epoch the models are merged into a global model by averaging parameters, the merge being repeated so shared trends accumulate; finally the global model is fine-tuned per stock under a penalty that keeps it from drifting away from the shared parameters. On the FNSPID dataset this procedure beats the sequential fine-tuning baseline in most configurations across seven backbones and three stock-group sizes, with the sharpest gains for DLinear, PaiFilter, and TexFilter; for PaiFilter at 25 stocks, the reported $R^2$ rises from about 0.44 with normal training to 0.91 with CSTI. The paper also reports cases where it does not win, notably TimesNet and PatchTST at 25 and 50 stocks, so its claim is superior performance in most, but not all, of the tested settings.

Load-bearing premise

The comparison assumes that the 'normal training' baseline, defined as sequentially fine-tuning a single model from one stock's data to the next, is a fair representative of ordinary single-stock learning rather than training each stock model independently.

Editorial extensions

If this is right

  • CSTI is architecture-agnostic: it changes the training procedure, not the model, so it can be layered onto existing forecasting backbones.
  • Parallel per-stock training before merging reduces wall-clock training time relative to serial single-stock pipelines as the number of stocks grows.
  • The global model alone tracks broad price trends, suggesting it can serve as a pretrained initialization for stock-specific fine-tuning.
  • Sentiment information continues to help under CSTI, consistent with the finding that sentiment-enriched inputs outperform price-only inputs.
  • CSTI is not uniformly superior: with TimesNet and PatchTST at 25 and 50 stocks, the normal sequential baseline wins, so the benefit depends on the backbone and group size.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors leave implicit that the choice of baseline matters: their 'normal training' is sequential fine-tuning from stock to stock, not independent per-stock training, so CSTI's margin could change against a strictly isolated baseline.
  • The global merged model is never evaluated as a reusable pretrained forecaster on unseen stocks, a use that the reported trend-following behavior suggests and that the paper does not test.
  • Because the experiments set all aggregation weights $w_k=1$, the proposed adaptive weighting by volatility, market capitalization, or validation accuracy remains untested; the current gains are a lower bound on what the mechanism might deliver.
  • A sector-sorted stress test would clarify whether the benefit comes from shared fundamental drivers: CSTI should help more when merged stocks belong to the same sector than when they are randomly grouped.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes Cross-Stock Trend Integration (CSTI), a training strategy for stock price prediction. CSTI is described in three phases: train a model per stock, repeatedly merge the local models into a global model in a federated-learning-inspired manner, and then fine-tune the global model on each stock with an L2 penalty toward the global parameters. The authors evaluate CSTI on the FNSPID dataset with seven backbone models (Transformer, TimesNet, DLinear, FreTS, PaiFilter, TexFilter, PatchTST) and report MAE, MSE, and R2 for training on 5, 25, and 50 stocks, with and without sentiment features. They claim that CSTI outperforms a "normal training" baseline in most settings and enhances the predictive ability of state-of-the-art forecasting models.

Significance. If the central claim held, the paper would offer a simple, model-agnostic training wrapper that lets per-stock forecasting models exploit cross-stock information, with a publicly available implementation and parallel-training efficiency. The evaluation on a large, realistic financial dataset with multiple backbone models is a strength, and the authors do acknowledge some settings where CSTI is not better. However, the paper's significance as written is undermined by a load-bearing discrepancy in Algorithm 1, by a baseline that is not the isolated per-stock training the introduction criticizes, and by the absence of any uncertainty quantification. The manuscript's own table contradicts the claim that sentiment features consistently help. These issues must be resolved before the contribution can be assessed.

major comments (4)
  1. [Algorithm 1, lines 10-11; Section III-B] The pseudocode does not implement the claimed iterative cross-stock refinement. In Phase 2, the global model is computed as an average of local models, but the loop at lines 10-11 repeats "Phase 1 & 2" without ever broadcasting θg back to the local models θk. The only assignment θk ← θg occurs in Phase 3, after the merging loop has finished. Consequently, during the repeated merging rounds, each local model either continues from its own parameters (if Phase 1 resumes from line 4) or is re-initialized (if Phase 1 restarts at line 3); in neither case does the global model influence local training. The global model at each round is therefore just an average of independently trained local models, and the paper's central mechanism—iterative integration of cross-stock patterns—is absent from the described algorithm. If the actual implementation does broadcast θg to local models, that step must be added to the pseudocode and reconciled with Section III-A's claim that local training avoids cross-stock contamination; if it does not, the method's description and its claimed interpretation must be revised.
  2. [Section IV-B, Table I] The "normal training" baseline is not the isolated per-stock training that the introduction criticizes. It is defined as training on one stock and then sequentially fine-tuning on the next stock, following reference [26]. Sequential fine-tuning is itself a cross-stock transfer procedure, and it may either help or hurt relative to independent per-stock training. Without an isolated per-stock baseline or published state-of-the-art numbers, the reported comparisons cannot support the claim that CSTI improves upon standard single-stock learning. Please add an independent per-stock baseline and, if available, results from the original backbone papers or other published models on the same data.
  3. [Table I; Section IV-B] No error bars, repeated seeds, or significance tests are reported, and several differences are very small. For example, in the 50-stock w/o-sentiment column, DLinear shows MAE 0.0237 (normal) versus 0.0234 (CSTI), and FreTS shows R2 0.7959 versus 0.8000; these differences are within the range one would expect from random initialization. Since the paper's main conclusion is that CSTI "outperforms" in most experiments, the authors need to report means and standard deviations over multiple random seeds and perform a paired significance test, or at least identify which cells differ by more than the noise level.
  4. [Section IV-B, Table I] The text contains claims that are directly contradicted by the paper's own table. First, the statement that "our method does not always yield better results" is correct, but it is in tension with the abstract's unconditional claim that CSTI "outperforms benchmark models and enhances the predictive capabilities of state-of-the-art approaches." Second, the sentence "models trained on data with sentiment information consistently outperformed those trained on data without sentiment information" is false in Table I: for example, in the 5-stock rows, TimesNet has MAE 0.1158 with sentiment versus 0.0459 without, DLinear has 0.0424 versus 0.0359, and FreTS has 0.0361 versus 0.0308. Please either correct the claim or provide an analysis that explains these discrepancies.
minor comments (6)
  1. [Section IV-A] The text says "In out experiments" (typo for "our experiments") and later "In out experiments, we used two sets of data features." Please proofread.
  2. [Section IV-A] The code link appears as a footnote with an asterisk but the footnote text is not visible in the manuscript; please ensure the repository URL is displayed and stable.
  3. [Section IV-A] The description of the normal training protocol in Section IV-B should be moved to Settings, and it should specify how the sequential fine-tuning is ordered across the randomly selected stocks and how many epochs are spent per stock; without this, the comparator is not fully reproducible.
  4. [Section III-C, Eq. (7)] The regularization weight λ and adaptation coefficient α are introduced but no values or selection procedure are reported in Section IV; since these are free parameters, their chosen values should be stated for reproducibility.
  5. [Table I] The caption says "results highlighted in red are the better ones," but the table in the manuscript does not visibly highlight cells in red; please add the highlighting or change the caption.
  6. [Figure 4] The loss-curve discussion describes only a single Transformer run on AMD with 50 stocks; please specify which stock and seed are shown, and clarify whether this is representative of other models and stocks.

Circularity Check

1 steps flagged · score 4.0 of 10

No self-citation-driven circularity, but the claimed 'iterative' cross-stock merging never feeds the global model back to local models, so the global understanding is defined by construction as a plain average.

  1. self definitional [Algorithm 1, lines 2-14; Section III-B, Eq. (4)]
    "Phase 1 ... Initialize model fk(·; θk) ... Update θk by minimizing Lk ... Phase 2 ... Aggregate global model: θg = 1 K PK k=1 wkθk ... for e = 1 to E do Refine θg by repeating Phase 1 & 2 ... Phase 3 ... Initialize θk ← θg"

    The only mechanism claimed to produce cross-stock understanding is the iterative merging loop, but Algorithm 1 never copies θg back into any local θk before the next local update. Repeating Phase 1 & 2 therefore re-trains each θk on its own stock only, and θg at every iteration is just the arithmetic mean of independently trained local models (wk=1 in all experiments). The paper's central assertion that iterative merging lets the global model 'progressively capture shared patterns and relationships among stocks' is thus a property attributed to a quantity that is, by construction, a one-shot parameter average; the feedback step that would realize the claimed cross-stock learning appears only after the loop ends (line 14), in the fine-tuning phase.

full rationale

The paper contains no closed-form derivation, so the classic circularity failure of fitting a parameter and then calling the fit a prediction does not arise. The empirical results on FNSPID are, in principle, external evidence, and there is no load-bearing self-citation chain: the FL inspiration is attributed to external works such as McMahan et al. The main circularity concern is narrower and definitional: Algorithm 1's 'iterative merging' is implemented without any broadcast of the global model to the local models, so the global model reduces by construction to the arithmetic mean of independently trained stock models. The property that this average 'captures shared patterns and relationships among stocks' is not derived from any cross-stock learning mechanism in the pseudocode; it is asserted of a simple average. Separately, the 'normal training' baseline is defined by the authors following [26] as sequential fine-tuning from one stock to the next, not the isolated single-stock training criticized in the introduction; that makes the reported advantage partly self-referential as a benchmark choice, but it is a validation weakness rather than a derivation-level circularity. The score reflects the reduction of the central global-model claim to plain averaging while acknowledging that the empirical performance numbers themselves are not forced by any fitted parameter.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on unstated hyperparameters (alpha, lambda), an assumed meaningful aggregate of heterogeneous stock models, and a self-defined baseline. No free parameters are fitted in a theoretical derivation; the listed items are the choices the experiments depend on.

free parameters (5)
  • Adaptation coefficient alpha = not reported
    Introduced in Equation (6) to scale the global learning rate during fine-tuning; no value or tuning procedure is given in Section IV.
  • Regularization weight lambda = not reported
    Introduced in Equation (7) to control deviation from the global model during fine-tuning; no value or ablation is reported.
  • Local learning rate and momentum = 0.01 and 0.9
    Fixed in Section IV-A for all local and fine-tuning updates; no sensitivity analysis is provided.
  • Local epochs before aggregation = 1
    Section IV-A states local models train for 1 epoch before merging during the first 50 epochs; no ablation on aggregation frequency is given.
  • Global aggregation and fine-tuning epochs = 50 and 50
    Section IV-A splits 100 epochs into 50 merging rounds and 50 fine-tuning rounds to match the normal training update count; no sensitivity analysis is reported.
assumptions (5)
  • domain assumption Historical price and sentiment features in FNSPID contain sufficient signal to predict future stock prices under the chosen loss.
    Section IV-A selects Open, Close, and Sentiment features without predictive power analysis or leakage checks.
  • domain assumption Weight-space averaging of independently trained stock models produces a meaningful global model that captures shared cross-stock patterns.
    Equation (4) in Section III-B assumes FL aggregation transfers under non-IID financial data; no convergence or representation analysis is provided.
  • ad hoc to paper The 'normal training' baseline from [26] is a fair representative of single-stock learning.
    Section IV-B defines normal training as sequential fine-tuning across stocks, which is not the isolated per-stock training the paper motivates.
  • ad hoc to paper Equal total weight-update iterations make the normal and CSTI comparison fair.
    Section IV-A asserts equal update counts without reporting total iterations or complete optimizer details.
  • domain assumption A temporal train/test split exists and prevents leakage from the global aggregation stage.
    The paper never states how FNSPID is split; the evaluation depends on this unstated assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Local Patterns to Global Understanding: Cross-Stock Trend Integration for Enhanced Predictive Modeling." pith.science (2026). https://pith.science/paper/CNS6DPBB

@misc{pith2026250516573,
  author       = {Pith},
  title        = {Pith review of: From Local Patterns to Global Understanding: Cross-Stock Trend Integration for Enhanced Predictive Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CNS6DPBB}},
  note         = {Machine review of arXiv:2505.16573}
}
read the original abstract

Stock price prediction is a critical area of financial forecasting, traditionally approached by training models using the historical price data of individual stocks. While these models effectively capture single-stock patterns, they fail to leverage potential correlations among stock trends, which could improve predictive performance. Current single-stock learning methods are thus limited in their ability to provide a broader understanding of price dynamics across multiple stocks. To address this, we propose a novel method that merges local patterns into a global understanding through cross-stock pattern integration. Our strategy is inspired by Federated Learning (FL), a paradigm designed for decentralized model training. FL enables collaborative learning across distributed datasets without sharing raw data, facilitating the aggregation of global insights while preserving data privacy. In our adaptation, we train models on individual stock data and iteratively merge them to create a unified global model. This global model is subsequently fine-tuned on specific stock data to retain local relevance. The proposed strategy enables parallel training of individual stock models, facilitating efficient utilization of computational resources and reducing overall training time. We conducted extensive experiments to evaluate the proposed method, demonstrating that it outperforms benchmark models and enhances the predictive capabilities of state-of-the-art approaches. Our results highlight the efficacy of Cross-Stock Trend Integration (CSTI) in advancing stock price prediction, offering a robust alternative to traditional single-stock learning methodologies.

Figures

Figures reproduced from arXiv: 2505.16573 by the authors.

Figure 1
Figure 1. Illustration of CSTI price prediction. Techniques such as Support Vector Machine (SVM) [7], Decision Tree (DT) [8], and ensemble methods like Random Forest (RF) [9] gained popularity for their ability to model nonlinear relationships in financial data. With the availability of large datasets and increased computational power, Deep Learning (DL) approaches, including Recur￾rent Neural Network (RNN) [10], Long Short-T… view at source ↗
Figure 2
Figure 2. The overview architecture of our proposed CSTI. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The visualization of the predicted regression lines for models trained on 50 stocks without sentiment information is presented. The [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Training loss comparison using Transformer pre-trained on 50 stocks data and fine-tuned on AMD stock prices. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 20 canonical work pages

  1. [26]

    Fnspid: A comprehensive financial news dataset in time series,

    Z. Dong, X. Fan, and Z. Peng, “Fnspid: A comprehensive financial news dataset in time series,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 4918– 4927

    work page 2024

  2. [1]

    Distribution of residual autocorrelations in autoregressive-integrated moving average time series models,

    G. E. Box and D. A. Pierce, “Distribution of residual autocorrelations in autoregressive-integrated moving average time series models,” Journal of the American statistical Association, vol. 65, no. 332, pp. 1509–1526, 1970

    1970

  3. [2]

    Forecasting time series using a methodology based on autoregressive integrated moving average and genetic program- ming,

    Y .-S. Lee and L.-I. Tong, “Forecasting time series using a methodology based on autoregressive integrated moving average and genetic program- ming,” Knowledge-Based Systems, vol. 24, no. 1, pp. 66–72, 2011

    work page 2011

  4. [3]

    Comparison of arima and exponential smoothing models in prediction of stock prices,

    Y . Funde and A. Damani, “Comparison of arima and exponential smoothing models in prediction of stock prices,” The Journal of Predic- tion Markets, vol. 17, no. 1, pp. 21–38, 2023

    work page 2023

  5. [4]

    Exponential smoothing: The state of the art,

    E. S. Gardner Jr, “Exponential smoothing: The state of the art,” Journal of forecasting, vol. 4, no. 1, pp. 1–28, 1985

    work page 1985

  6. [5]

    Exponential smoothing: The state of the art—part ii,

    ——, “Exponential smoothing: The state of the art—part ii,” Interna- tional journal of forecasting , vol. 22, no. 4, pp. 637–666, 2006

    work page 2006

  7. [6]

    Predicting the brazilian stock market through neural networks and adaptive exponential smoothing methods,

    E. De Faria, M. P. Albuquerque, J. Gonzalez, J. Cavalcante, and M. P. Albuquerque, “Predicting the brazilian stock market through neural networks and adaptive exponential smoothing methods,” Expert Systems with Applications, vol. 36, no. 10, pp. 12 506–12 509, 2009

    work page 2009

  8. [7]

    Sup- port vector machines,

    M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and B. Scholkopf, “Sup- port vector machines,” IEEE Intelligent Systems and their applications , vol. 13, no. 4, pp. 18–28, 1998

    1998

Show all 34 references
  1. [8]

    Decision trees,

    B. De Ville, “Decision trees,” Wiley Interdisciplinary Reviews: Compu- tational Statistics, vol. 5, no. 6, pp. 448–455, 2013

    2013

  2. [9]

    Random forests,

    L. Breiman, “Random forests,” Machine learning , vol. 45, pp. 5–32, 2001

    2001

  3. [10]

    Finding structure in time,

    J. L. Elman, “Finding structure in time,” Cognitive science , vol. 14, no. 2, pp. 179–211, 1990

    1990

  4. [11]

    Long short-term memory,

    S. Hochreiter, “Long short-term memory,” Neural Computation MIT- Press, 1997

    1997

  5. [12]

    Learning phrase representations using rnn encoder-decoder for statistical machine translation,

    K. Cho, “Learning phrase representations using rnn encoder-decoder for statistical machine translation,” arXiv preprint arXiv:1406.1078 , 2014

    2014 arXiv

  6. [13]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  7. [14]

    Filternet: Harnessing frequency filters for time series forecasting,

    K. Yi, J. Fei, Q. Zhang, H. He, S. Hao, D. Lian, and W. Fan, “Filternet: Harnessing frequency filters for time series forecasting,” arXiv preprint arXiv:2411.01623, 2024

    2024 arXiv

  8. [15]

    Frequency-domain mlps are more effective learners in time series forecasting,

    K. Yi, Q. Zhang, W. Fan, S. Wang, P. Wang, H. He, N. An, D. Lian, L. Cao, and Z. Niu, “Frequency-domain mlps are more effective learners in time series forecasting,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  9. [16]

    Are transformers effective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transformers effective for time series forecasting?” in Proceedings of the AAAI conference on artificial intelligence, vol. 37, 2023, pp. 11 121–11 128

    2023

  10. [17]

    Timesnet: Temporal 2d-variation modeling for general time series analysis,

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “Timesnet: Temporal 2d-variation modeling for general time series analysis,” arXiv preprint arXiv:2210.02186, 2022

    2022 arXiv

  11. [18]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” arXiv preprint arXiv:2211.14730, 2022

    2022 arXiv

  12. [19]

    Cross-stock momentum and factor momentum,

    J. Yan and J. Yu, “Cross-stock momentum and factor momentum,” Journal of Financial Economics , vol. 150, no. 2, p. 103716, 2023

    2023

  13. [20]

    Diffusion variational autoencoder for tackling stochasticity in multi-step regression stock price prediction,

    K. J. Koa, Y . Ma, R. Ng, and T.-S. Chua, “Diffusion variational autoencoder for tackling stochasticity in multi-step regression stock price prediction,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, pp. 1087–1096

    2023

  14. [21]

    A stock price prediction approach based on time series decomposition and multi-scale cnn using ohlct images,

    Z. Pei, J. Yan, J. Yan, B. Yang, Z. Li, L. Zhang, X. Liu, and Y . Zhang, “A stock price prediction approach based on time series decomposition and multi-scale cnn using ohlct images,” arXiv preprint arXiv:2410.19291 , 2024. 11

    2024 arXiv

  15. [22]

    A stock prediction method based on multidi- mensional and multilevel feature dynamic fusion,

    Y . Dong and Y . Hao, “A stock prediction method based on multidi- mensional and multilevel feature dynamic fusion,” Electronics, vol. 13, no. 20, p. 4111, 2024

    2024

  16. [23]

    Multi-factor stock price prediction based on gan-trellisnet,

    W. Liu, Y . Ge, and Y . Gu, “Multi-factor stock price prediction based on gan-trellisnet,” Knowledge and Information Systems , pp. 1–22, 2024

    2024

  17. [24]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” Communications of the ACM , vol. 63, no. 11, pp. 139–144, 2020

    2020

  18. [25]

    Trellis networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “Trellis networks for sequence modeling,” arXiv preprint arXiv:1810.06682 , 2018

    2018 arXiv

  19. [27]

    Effective exploitation of macroeco- nomic indicators for stock direction classification using the multimodal fusion transformer,

    T.-W. Lee, P. Teisseyre, and J. Lee, “Effective exploitation of macroeco- nomic indicators for stock direction classification using the multimodal fusion transformer,” IEEE Access, vol. 11, pp. 10 275–10 287, 2023

    2023

  20. [28]

    The evolution of studies on social media sentiment in the stock market: Insights from bibliometric analy- sis,

    K. Nyakurukwa and Y . Seetharam, “The evolution of studies on social media sentiment in the stock market: Insights from bibliometric analy- sis,” Scientific African, vol. 20, p. e01596, 2023

    2023

  21. [29]

    Survey on privacy-preserving machine learning,

    J. Liu and X. Meng, “Survey on privacy-preserving machine learning,” JCRD, vol. 57, no. 2, p. 346, 2020

    2020

  22. [30]

    Privacy-preserving deep learning on machine learning as a service—a comprehensive survey,

    H. C. Tanuwidjaja, R. Choi, S. Baek, and K. Kim, “Privacy-preserving deep learning on machine learning as a service—a comprehensive survey,” IEEE Access, vol. 8, pp. 167 425–167 447, 2020

    2020

  23. [31]

    {SWIFT}: Super-fast and robust privacy-preserving machine learning,

    N. Koti, M. Pancholi, A. Patra, and A. Suresh, “ {SWIFT}: Super-fast and robust privacy-preserving machine learning,” inSecurity Symposium, 2021

    2021

  24. [32]

    Federated learning: Strategies for improving communication efficiency,

    J. Konecn ´y, H. B. McMahan, F. X. Yu, P. Richt ´arik, A. T. Suresh, and D. Bacon, “Federated learning: Strategies for improving communication efficiency,” ArXiv, vol. abs/1610.05492, 2016

    2016 arXiv

  25. [33]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in AIS. PMLR, 2017, pp. 1273–1282

    2017

  26. [34]

    Pytorch: An impera- tive style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, and L. Antiga, “Pytorch: An impera- tive style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019. Yi Hu received the M....

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.