Pith. sign in

REVIEW 4 major objections 4 minor 74 references

Pre-training Time Series Models with Stock Data Customization

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that three stock-specific pre-training tasks—identifying a stock's code, its sector, and the moving average of a masked price window—give a simple two-layer transformer an initialization that beats the market and prior…

desk verdict A genuinely useful pre-training idea for stock selection, but the MAP task likely leaks its target and the empirical claims outrun the evidence. read the letter →

arxiv 2506.16746 v1 pith:TLFOY6CJ submitted 2025-06-20 cs.CE

Mengyu Wang , Tiejun Ma , Shay B. Cohen This is my paper · ORCID
classification cs.CE
keywords stockselectionpre-trainingtransformermovingaveragepredictioncodeclassificationsectorSharperatiorepresentationlearning
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

Stock selection models are usually trained directly to predict profits, and pre-training recipes are imported from NLP and vision without adapting to how stock prices behave. This paper argues that the right pre-training objectives should exploit two properties of financial data: each stock has a persistent identity (its code and sector leave statistical fingerprints in the price series), and prices are non-stationary, so precise masked-value prediction is the wrong target. The authors propose three pre-training tasks—stock code classification, stock sector classification, and moving average prediction—and show that a standard two-layer transformer pre-trained on them, then fine-tuned to rank stocks, outperforms the market and previous methods in cumulative return and Sharpe ratio on NASDAQ, NYSE, TOPIX-100, FTSE-100, and recent NASDAQ data. If the claim holds, it means the latent statistical structure of price series is a reusable source of signal that improves downstream stock selection without new model architectures.

What carries the argument

The load-bearing machinery is the set of three pre-training objectives, applied to sliding-window price slices before the stock-selection fine-tuning step. Stock code classification (SCC) trains the model to predict which of N stocks a normalized price slice came from; stock sector classification (SSC) predicts the sector; moving average prediction (MAP) masks part of a window and predicts the average close price over the window, a smoother target than the exact masked values. The three losses are combined as $L_{\mathrm{pt}} = \alpha L_{\mathrm{scc}} + \beta L_{\mathrm{ssc}} + \gamma L_{\mathrm{map}}$ during pre-training; fine-tuning then replaces the task heads with a return-ratio head trained with a regression plus ranking loss, with some transformer parameters optionally frozen. The simulations use a Wiener-process model $S(t+\Delta t)=S(t)\exp((\mu-\sigma^2/2)\Delta t+\sigma\sqrt{\Delta t}\,Z_t)$ to test whether the classification tasks detect differences in the statistical parameters of the generating process.

What would settle it

Run MAP pre-training on the NASDAQ data twice, once with the 5, 10, 20, and 30-day moving-average features present and once with them removed or masked, holding everything else fixed. If the downstream stock-selection Sharpe ratio and the pre-training MSE are essentially unchanged when the moving-average features are removed, the MAP gains are robust; if the gains collapse, the reported benefit of MAP comes from copying the target from the input.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that pre-training a stock-selection model on tasks that are not directly profit-related—guessing which stock a price slice came from, which sector it belongs to, and the average close price of a partially masked window—produces representations that transfer to the ranking objective better than training from scratch or importing generic time-series pre-training. The authors implement this in SSPT, a two-layer transformer, and report that pre-training with the classification tasks (optionally combined with moving average prediction) raises cumulative investment return ratio and Sharpe ratio relative to the market and to strong baselines such as hypergraph and spatio-temporal transformers across five datasets spanning four markets and two time periods. Simulated data experiments with geometric Brownian motion support the mechanism: classification accuracy rises when simulated series differ in drift or volatility, indicating the tasks capture distinguishing statistical features rather than memorized labels.

Load-bearing premise

The moving-average prediction task assumes its target—the window's average close price—cannot be nearly read off from the input features, since the inputs already contain 5-, 10-, 20-, and 30-day moving averages and the paper does not say these are masked; if they remain, the task reduces to copying an available feature.

Editorial extensions

If this is right

  • Pre-training on stock identity and sector classification reliably improves downstream stock selection even when the fine-tuned model's parameters are all unfrozen, giving a stronger alternative to training from scratch.
  • Moving average prediction beats traditional masked-value prediction as a pre-training target for volatile price series, with lower pre-training error and better Sharpe ratio on NASDAQ.
  • The classification tasks' accuracy, though far below perfect, is well above random guessing and rises when simulated price series differ more in volatility, indicating price slices carry distinguishing statistical features.
  • A practical recipe emerges: use all available price features during pre-training, learning rate $10^{-3}$ for classification tasks and $10^{-4}$ for MAP, freeze only the embedding layer for SSC, and combine SCC and SSC with balanced coefficients for dependable gains.
  • Outperformance holds across market regimes, including the COVID-19 period in NASDAQ-recent, suggesting the initialization generalizes across time and geography.

Reading between the lines

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

  • Beyond the paper: the MAP task's reported gains may be partly inflated by feature leakage—the inputs already include 5, 10, 20, and 30-day moving averages, and the paper never states that these features are masked during MAP; a clean test would remove or mask them and re-measure the pre-training and downstream gains.
  • Beyond the paper: the identity and sector classification tasks could serve as a general representation-learning prior for any non-stationary financial series where individual assets have persistent statistical profiles, such as cryptocurrencies or commodities.
  • Beyond the paper: the Sharpe-ratio differences between SSPT and the strongest baseline on NASDAQ are substantial, but the paper does not report variance or significance across random seeds; a natural extension is a paired statistical test across re-runs.
  • Beyond the paper: the finding that MAP is a good initialization but transfers poorly when parameters are frozen suggests a testable hypothesis—progressively unfreezing layers during fine-tuning may unlock more of MAP's learned structure.
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 / 4 minor

Summary. The paper proposes three pre-training tasks for stock price series—stock code classification (SCC), stock sector classification (SSC), and moving average prediction (MAP)—and a two-layer transformer, SSPT, that is pre-trained with these tasks and fine-tuned for stock selection. Experiments on NASDAQ, NYSE, TOPIX-100, FTSE-100, and NASDAQ-recent report IRR and Sharpe ratio improvements over market and baselines, plus simulated-data analyses that attribute the classification tasks' benefit to distinguishing statistical features. The authors also explore fine-tuning/freezing strategies and loss-coefficient combinations.

Significance. If the results hold, the paper contributes a lightweight, data-efficient pre-training recipe for stock selection and provides some mechanistic evidence via controlled simulations. Strengths include the public code, the multi-market/multi-period evaluation, and the explicit ablation of task combinations. The paper is also candid in Section 6.2 about MAP's fragility in combined pre-training, which is a point in its favor. However, the central empirical claim currently rests on single test-year numbers without variance estimates, and the MAP task has a concrete leakage problem that may invalidate one of the three proposed mechanisms. The significance is therefore conditional on a successful revision.

major comments (4)
  1. [§4.2, Eq. (3), with §5.1] The MAP objective is not well-defined as a pre-training task unless the moving-average input features are masked. X_{i,t} contains, for every day in the window, the close price plus 5-, 10-, 20-, and 30-day moving averages of close (§5.1). The MAP target is the ΔT-day average close. If any close in the window is masked, the surrounding MA features are linear functions of that masked close (e.g., MA5 at the next day uses the masked close along with four observed closes), so the target is nearly a linear function of features that remain visible. The paper never states that MA features are removed or masked when a close price is masked. The MSE improvement in §6.1 (5.5e-6 to 2.3e-6 when all price values are added) is exactly what one would expect from this leakage. Please specify the masking protocol; if MA features are not masked, the MAP task must be rerun with them excluded or masked, and the claims about MAP's contribution re-evaluated.
  2. [§6.2, Figure 6] The paper's own ablation undercuts the claim that all three pre-training tasks are effective. The text states that two of the three combinations including MAP are worse than no pre-training and that MAP's contribution is sensitive to the loss coefficient; only the validation-selected γ=10 case beats SCC+SSC. Given that the abstract presents MAP as one of three novel tasks that 'consistently outperforms' the market, the evidence for MAP is not robust. Either report a stable recipe for including MAP with pre-registered coefficients, or revise the contribution claim to SCC+SSC with MAP treated as a conditional or optional task.
  3. [§5.4, §6.3, Tables 2–3] All reported IRR/SR values are single numbers from one test year with no error bars, multiple seeds, or significance tests. The evaluation is further softened by validating hyperparameters on the same validation period used for early stopping and then selecting the 'broader range of settings' for the comparison (§6.3). Because stock-selection metrics are noisy and the market baseline is itself variable, the headline claim of consistent outperformance requires variance estimates or repeated test-year evaluations. Please add at least repeated-seed results and, if possible, additional test years or a block-bootstrap confidence interval.
  4. [§6.3, Table 2] Baseline comparisons are not fully reproducible as reported. The text says the comparative models are selected from 'a broader range of settings based on validation results,' but the exact hyperparameter ranges and selection protocol are not given. Since many baselines are from prior papers with their own tuning, it is unclear whether the gains reflect the pre-training tasks or a more favorable search budget. Please provide the full hyperparameter grid and selection rule, or the code used, so the comparison is apples-to-apples.
minor comments (4)
  1. [Figure 5] The caption says 'The Sharpe ratio results of the stock selection task,' but the y-axis is labeled 'Cumulative IRR'; the figure appears to show IRR, while Section 6.1 describes SR as the reported metric. Please correct this mismatch or the associated discussion.
  2. [Appendix C, Table 4] The MVP vs. MAP comparison is not a clean ablation because both tasks may be affected by the same moving-average feature leakage discussed above; this table should be re-run after the masking protocol is clarified.
  3. [Table 2] Several entries in the table have formatting problems, such as '0.711.27' for ALSP-TF and '0.792.14' for CI-STHPAN, which should be separated into distinct numbers for readability.
  4. [Throughout] There are several typos and minor wording issues, including 'higer' in Section 6.1, 'no paramters' in Figure 5, 'NASDAQ2tracks' in Appendix A, and 'FSTE-100' in the caption of Figure 8. These should be corrected in a final pass.

Circularity Check

1 steps flagged · score 6.0 of 10

MAP pre-training target is recoverable from the moving-average features included in the input, so one of the three proposed pre-training predictions reduces by construction.

  1. self definitional [Section 4.2, Eq. (3); Section 5.1 data features]
    "We then mask a portion of the prices within each window and train the model to predict these average prices from the partially masked data. ... The model is trained with the loss L_map, defined as: L_map = ( f(X_i,t ; w_f, w_map) - (1/Delta T) sum_{j=1}^{Delta T} p_{i,t-j+1} )^2. ... The historical data comprises five daily values: open price, high price, low price, close price, and trading volume. Following previous works, we augment these values with 5, 10, 20, and 30-day moving averages."

    The MAP target is the average close over the look-back window, but the input X_{i,t} already contains, at every time step in that window, 5-, 10-, 20-, and 30-day moving averages of close prices. These features are linear functions of the close prices, so the window average (and the masked close prices themselves) can be recovered from unmasked moving-average features; e.g., p_t = 5*MA5_t - (p_{t-4}+...+p_{t-1}) when those four closes are observed, and longer-window targets are linearly expressible through the included MA features. The paper never states that the moving-average features are masked or removed during MAP pre-training. The MAP task is therefore self-predictive: the target is a function of the input features by construction rather than an independent quantity to be learned.

full rationale

The circularity is confined to the moving-average prediction pre-training task. SCC and SSC use stock code and sector as external labels, not as inputs, so those tasks are not circular. The paper's self-citations are background/baseline references and are not load-bearing. The final SSPT comparisons against market benchmarks and baseline methods are empirical and not forced by the MAP reduction. Nonetheless, because one of the three proposed pre-training objectives is, as specified, nearly copyable from the augmented input features, the paper's claim that MAP 'mitigates price volatility and non-stationarity' and contributes learned knowledge is not supported as stated. This is a partial circularity (one prediction reduces by construction), warranting a score of 6 rather than a higher score.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central empirical claims rest on a standard deep learning pipeline with many hyperparameters selected via validation. The most fragile unstated premise is that the MAP target (window average close price) is not directly recoverable from the precomputed moving-average input features.

free parameters (7)
  • alpha (SCC loss coefficient) = e.g., 1 or 5
    Controls weight of stock code classification loss in Eq. (4); selected via validation.
  • beta (SSC loss coefficient) = 1 (balanced)
    Controls weight of stock sector classification loss in Eq. (4); selected via validation.
  • gamma (MAP loss coefficient) = e.g., 1, 5, 10
    Controls weight of moving average prediction loss in Eq. (4); selected via validation.
  • epsilon (fine-tuning loss balance) = from {1, 5, 10}
    Balances regression and ranking losses in Eq. (5); chosen by validation.
  • learning rate = from {1e-3, 1e-4, 1e-5}
    Chosen per task and phase via validation; classification tasks prefer 1e-3, MAP prefers 1e-4.
  • look-back length (Delta T) = from {16, 32}
    Length of price series window; selected from a set based on validation.
  • mask rate = optimal around 0.3
    Fraction of prices masked in MAP; tuned via validation (Figure 9).
assumptions (4)
  • domain assumption Stock markets contain learnable, non-random signals.
    Stated in Section 1 as foundational; if false, no prediction method can work.
  • domain assumption Price series slices contain sufficient distinguishing information to classify stock code and sector.
    Assumed in Section 4.1; the paper provides evidence but it is a premise for the pre-training tasks.
  • domain assumption Moving averages provide a more stable and predictable target than raw prices.
    Motivates MAP in Section 4.2.
  • ad hoc to paper The MAP target is not directly available as an input feature during pre-training.
    The paper includes 5, 10, 20, and 30-day moving averages as input features (Section 5.1) but does not state they are masked or removed during MAP; if they are present, the task reduces to copying a feature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pre-training Time Series Models with Stock Data Customization." pith.science (2026). https://pith.science/paper/TLFOY6CJ

@misc{pith2026250616746,
  author       = {Pith},
  title        = {Pith review of: Pre-training Time Series Models with Stock Data Customization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLFOY6CJ}},
  note         = {Machine review of arXiv:2506.16746}
}
read the original abstract

Stock selection, which aims to predict stock prices and identify the most profitable ones, is a crucial task in finance. While existing methods primarily focus on developing model structures and building graphs for improved selection, pre-training strategies remain underexplored in this domain. Current stock series pre-training follows methods from other areas without adapting to the unique characteristics of financial data, particularly overlooking stock-specific contextual information and the non-stationary nature of stock prices. Consequently, the latent statistical features inherent in stock data are underutilized. In this paper, we propose three novel pre-training tasks tailored to stock data characteristics: stock code classification, stock sector classification, and moving average prediction. We develop the Stock Specialized Pre-trained Transformer (SSPT) based on a two-layer transformer architecture. Extensive experimental results validate the effectiveness of our pre-training methods and provide detailed guidance on their application. Evaluations on five stock datasets, including four markets and two time periods, demonstrate that SSPT consistently outperforms the market and existing methods in terms of both cumulative investment return ratio and Sharpe ratio. Additionally, our experiments on simulated data investigate the underlying mechanisms of our methods, providing insights into understanding price series. Our code is publicly available at: https://github.com/astudentuser/Pre-training-Time-Series-Models-with-Stock-Data-Customization.

Figures

Figures reproduced from arXiv: 2506.16746 by the authors.

Figure 1
Figure 1. An example of stock code classification. This task [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the pre-training and fine-tuning procedures for our SSPT model. The frozen parameters can be adjusted, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Classification accuracy comparison for the two pre [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: The Sharpe ratio results of the stock selection task [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 4
Figure 4. Figure 4: Evaluation results of the three pre-training tasks at [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Sharpe ratio results on the NASDAQ market for [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Classification results on 10, 50, 100, 500, and 1000 [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Classification accuracy comparison for the two pre [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Pre-training MSE and stock selection SR results [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: The cumulative IRR results of the stock selection [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Cumulative IRR results on the NASDAQ market [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Classification accuracy on 10 simulated stock price [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 59 canonical work pages

  1. [1]

    Ayodele Ariyo Adebiyi, Aderemi Oluyinka Adewumi, and Charles Korede Ayo

    work page

  2. [2]

    Wei Bao, Jun Yue, and Yulei Rao. 2017. A deep learning framework for financial time series using stacked autoencoders and long-short term memory.PloS one 12, 7 (2017), e0180944

    work page 2017

  3. [3]

    2009.Arbitrage theory in continuous time

    Tomas Björk. 2009.Arbitrage theory in continuous time. Oxford university press

    work page 2009

  4. [4]

    William Brock, Josef Lakonishok, and Blake LeBaron. 1992. Simple technical trading rules and the stochastic properties of stock returns.The Journal of finance 47, 5 (1992), 1731–1764

    work page 1992

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. [n. d.]. Language models are few-shot learners.Advances in neural information processing systems33 ([n. d.]), 1877–1901

    work page 1901

  6. [6]

    Salvatore Carta, Anselmo Ferreira, Alessandro Sebastian Podda, Diego Reforgiato Recupero, and Antonio Sanna. 2021. Multi-DQN: An ensemble of Deep Q-learning agents for stock market forecasting.Expert systems with applications164 (2021), 113820

    work page 2021

  7. [7]

    Yingmei Chen, Zhongyu Wei, and Xuanjing Huang. 2018. Incorporating cor- poration relationship via graph convolutional neural networks for stock price prediction. InProceedings of the 27th ACM international conference on information and knowledge management. 1655–1658

    work page 2018

  8. [8]

    Bent Jesper Christensen and Morten Ørregaard Nielsen. 2007. The effect of long memory in volatility on stock market fluctuations.The Review of Economics and Statistics89, 4 (2007), 684–700

    work page 2007

Show all 74 references
  1. [9]

    1988.What moves stock prices?Vol

    David M Cutler, James M Poterba, and Lawrence H Summers. 1988.What moves stock prices?Vol. 487. National Bureau of Economic Research Cambridge, Massachusetts

    1988

  2. [10]

    Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805(2018)

    2018 arXiv

  3. [11]

    Qianggang Ding, Sifan Wu, Hao Sun, Jiadong Guo, and Jian Guo. 2020. Hier- archical Multi-Scale Gaussian Transformer for Stock Movement Prediction.. In IJCAI. 4640–4646

    2020

  4. [12]

    Fama Eugene and Kenneth French. 1992. The cross-section of expected stock returns.Journal of finance47, 2 (1992), 427–465

    1992

  5. [13]

    Eugene F Fama. 1970. Efficient capital markets: A review of theory and empirical work.The journal of Finance25, 2 (1970), 383–417

    1970

  6. [14]

    Eugene F Fama. 1995. Random walks in stock market prices.Financial analysts journal51, 1 (1995), 75–80

    1995

  7. [15]

    Alan Fan and Marimuthu Palaniswami. 2001. Stock selection using support vector machines. InIJCNN’01. International Joint Conference on Neural Networks. Proceedings (Cat. No. 01CH37222), Vol. 3. IEEE, 1793–1798

    2001

  8. [16]

    Jinyong Fan and Yanyan Shen. 2024. StockMixer: a simple yet strong MLP-based architecture for stock price forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 8389–8397

    2024

  9. [17]

    Fuli Feng, Huimin Chen, Xiangnan He, Jie Ding, Maosong Sun, and Tat-Seng Chua. 2019. Enhancing Stock Movement Prediction with Adversarial Training.. InIJCAI, Vol. 19. 5843–5849

    2019

  10. [18]

    Fuli Feng, Xiangnan He, Xiang Wang, Cheng Luo, Yiqun Liu, and Tat-Seng Chua

  11. [19]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick

  12. [20]

    Min Hou, Chang Xu, Yang Liu, Weiqing Liu, Jiang Bian, Le Wu, Zhi Li, Enhong Chen, and Tie-Yan Liu. 2021. Stock trend prediction with multi-granularity data: A contrastive learning approach with adaptive fusion. InProceedings of the 30th ACM International Conference on Informat...

    2021

  13. [21]

    Wenpeng Hu, Mengyu Wang, Bing Liu, Feng Ji, Haiqing Chen, Dongyan Zhao, Jinwen Ma, and Rui Yan. 2019. Transformation of dense and sparse text repre- sentations.arXiv preprint arXiv:1911.02914(2019)

    2019 arXiv

  14. [22]

    2016.Options, futures, and other derivatives

    John C Hull and Sankarshan Basu. 2016.Options, futures, and other derivatives. Pearson Education India

    2016

  15. [23]

    Narasimhan Jegadeesh and Sheridan Titman. 1993. Returns to buying winners and selling losers: Implications for stock market efficiency.The Journal of finance 48, 1 (1993), 65–91

    1993

  16. [24]

    2013.Stock Market Trend Analysis Using Hidden Markov Models

    G Kavitha, A Udhayakumar, and D Nagarajan. 2013.Stock Market Trend Analysis Using Hidden Markov Models. Technical Report. arXiv. org

    2013

  17. [25]

    Raehyun Kim, Chan Ho So, Minbyul Jeong, Sanghoon Lee, Jinkyu Kim, and Jaewoo Kang. 2019. Hats: A hierarchical graph attention network for stock movement prediction.arXiv preprint arXiv:1908.07999(2019)

    2019 arXiv

  18. [26]

    Kelvin JL Koa, Yunshan Ma, Ritchie Ng, and Tat-Seng Chua. 2024. Learning to generate explainable stock predictions using self-reflective large language models. InProceedings of the ACM on Web Conference 2024. 4304–4315

    2024

  19. [27]

    Tong Li, Zhaoyang Liu, Yanyan Shen, Xue Wang, Haokun Chen, and Sen Huang

  20. [28]

    Wei Li, Ruihan Bao, Keiko Harimoto, Deli Chen, Jingjing Xu, and Qi Su. 2021. Modeling the stock relation with graph network for overnight stock movement prediction. InProceedings of the twenty-ninth international conference on interna- tional joint conferences on artificial in...

    2021

  21. [29]

    Weixian Waylon Li, Yftah Ziser, Yifei Xie, Shay B Cohen, and Tiejun Ma. 2024. TSPRank: Bridging Pairwise and Listwise Methods with a Bilinear Travelling Salesman Model.arXiv preprint arXiv:2411.12064(2024)

    2024 arXiv

  22. [30]

    Jintao Liu, Hongfei Lin, Xikai Liu, Bo Xu, Yuqi Ren, Yufeng Diao, and Liang Yang

  23. [31]

    Yang Liu, Qi Liu, Hongke Zhao, Zhen Pan, and Chuanren Liu. 2020. Adaptive quantitative trading: An imitative deep reinforcement learning approach. In Proceedings of the AAAI conference on artificial intelligence, Vol. 34. 2128–2135

    2020

  24. [32]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692 (2019)

    2019 arXiv

  25. [33]

    Andrew W Lo and A Craig MacKinlay. 1990. When are contrarian profits due to stock market overreaction?The review of financial studies3, 2 (1990), 175–205

    1990

  26. [34]

    Di Luo, Weiheng Liao, Shuqi Li, Xin Cheng, and Rui Yan. 2023. Causality-Guided Multi-Memory Interaction Network for Multivariate Stock Price Movement Pre- diction. InProceedings of the 61st Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Paper...

    2023

  27. [35]

    In Proceedings of the first workshop on financial technology and natural language processing

    Transformer-based capsule network for stock movement prediction. In Proceedings of the first workshop on financial technology and natural language processing. 66–73

  28. [36]

    Tao Ma and Ying Tan. 2022. Stock ranking with multi-task learning.Expert Systems with Applications199 (2022), 116886

    2022

  29. [37]

    Amil Merchant, Elahe Rahimtoroghi, Ellie Pavlick, and Ian Tenney. 2020. What happens to BERT embeddings during fine-tuning?arXiv preprint arXiv:2004.14448 (2020)

    2020 arXiv

  30. [38]

    Tobias J Moskowitz and Mark Grinblatt. 1999. Do industries explain momentum? The Journal of finance54, 4 (1999), 1249–1290

    1999

  31. [39]

    1999.Technical analysis of the financial markets: A comprehensive guide to trading methods and applications

    John J Murphy. 1999.Technical analysis of the financial markets: A comprehensive guide to trading methods and applications. Penguin

    1999

  32. [40]

    Yongen Luo, Jicheng Hu, Xiaofeng Wei, Dongjian Fang, and Heng Shao. 2014. Stock trends prediction based on hypergraph modeling clustering algorithm. In2014 IEEE International Conference on Progress in Informatics and Computing. IEEE, 27–31

    2014

  33. [41]

    G Preethi and B Santhi. 2012. STOCK MARKET FORECASTING TECHNIQUES: A SURVEY.Journal of Theoretical & Applied Information Technology46, 1 (2012)

    2012

  34. [42]

    Yao Qin, Dongjin Song, Haifeng Chen, Wei Cheng, Guofei Jiang, and Garrison Cottrell. 2017. A dual-stage attention-based recurrent neural network for time series prediction.arXiv preprint arXiv:1704.02971(2017)

    2017 arXiv

  35. [43]

    Ramit Sawhney, Shivam Agarwal, Arnav Wadhwa, Tyler Derr, and Rajiv Ratn Shah. 2021. Stock selection via spatiotemporal hypergraph attention network: A learning to rank approach. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 497–504

    2021

  36. [44]

    2005.Stochastic calculus for finance I: the binomial asset pricing model

    Steven Shreve. 2005.Stochastic calculus for finance I: the binomial asset pricing model. Springer Science & Business Media

    2005

  37. [45]

    Rudra Kalyan Nayak, Debahuti Mishra, and Amiya Kumar Rath. 2015. A Naïve SVM-KNN based stock market trend reversal analysis for Indian benchmark indices.Applied Soft Computing35 (2015), 670–680

    2015

  38. [46]

    Heyuan Wang, Shun Li, Tengjiao Wang, and Jiayi Zheng. 2021. Hierarchical Adaptive Temporal-Relational Modeling for Stock Trend Prediction.. InIJCAI. 3691–3698

    2021

  39. [47]

    Heyuan Wang, Tengjiao Wang, Shun Li, and Shijie Guan. 2022. HATR-I: Hierar- chical adaptive temporal relational interaction for stock trend prediction.IEEE Transactions on Knowledge and Data Engineering35, 7 (2022), 6988–7002

    2022

  40. [48]

    Heyuan Wang, Tengjiao Wang, Shun Li, Jiayi Zheng, Shijie Guan, and Wei Chen

  41. [49]

    Jung-Hua Wang and Jia-Yann Leu. 1996. Stock market trend prediction using ARIMA-based neural networks. InProceedings of International Conference on Neural Networks (ICNN’96), Vol. 4. IEEE, 2160–2165

    1996

  42. [50]

    Chaojie Wang, Yuanyuan Chen, Shuqi Zhang, and Qiuhui Zhang. 2022. Stock market index prediction using deep Transformer model.Expert Systems with Applications208 (2022), 118128

    2022

  43. [51]

    Mengyu Wang and Tiejun Ma. 2024. MANA-Net: Mitigating Aggregated Senti- ment Homogenization with News Weighting for Enhanced Market Prediction. arXiv preprint arXiv:2409.05698(2024)

    2024 arXiv

  44. [52]

    Mengyu Wang, Yijia Shao, Haowei Lin, Wenpeng Hu, and Bing Liu. 2022. Cmg: A class-mixed generation approach to out-of-distribution detection. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 502–518

    2022

  45. [53]

    Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y Zhang, and Jun Zhou. 2024. Timemixer: Decomposable multiscale mixing for time series forecasting.arXiv preprint arXiv:2405.14616(2024)

    2024 arXiv

  46. [54]

    Adaptive Long-Short Pattern Transformer for Stock Investment Selection.. InIJCAI. 3970–3977

  47. [55]

    Ke Xu, Yifan Zhang, Deheng Ye, Peilin Zhao, and Mingkui Tan. 2021. Relation- aware transformer for portfolio policy learning. InProceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence. 4647–4653

    2021

  48. [56]

    Mengyu Wang, Shay B Cohen, and Tiejun Ma. 2024. Modeling News Interactions and Influence for Financial Market Prediction.arXiv preprint arXiv:2410.10614 (2024)

    2024 arXiv

  49. [57]

    Yunan Ye, Hengzhi Pei, Boxin Wang, Pin-Yu Chen, Yada Zhu, Ju Xiao, and Bo Li

  50. [58]

    Jaemin Yoo, Yejun Soun, Yong-chan Park, and U Kang. 2021. Accurate multivariate stock movement prediction via data-axis transformer with multi-level contexts. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2037–2045

    2021

  51. [59]

    Xu Yuemei, Wang Zihou, and Wu Zixin. 2021. Predicting Stock Trends with CNN- BiLSTM Based Multi-Feature Integration Model.Data Analysis and Knowledge Discovery5, 7 (2021), 126–138

    2021

  52. [60]

    Hongjie Xia, Huijie Ao, Long Li, Yu Liu, Sen Liu, Guangnan Ye, and Hongfeng Chai. 2024. CI-STHPAN: Pre-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 9187–9195

    2024

  53. [61]

    Liheng Zhang, Charu Aggarwal, and Guo-Jun Qi. 2017. Stock price prediction via discovering multi-frequency trading patterns. InProceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining. 2141– 2149

    2017

  54. [62]

    Xu Yan and Zhang Guosheng. 2015. Application of kalman filter in the prediction of stock price. In5th international symposium on knowledge acquisition and modeling (KAM 2015). Atlantis press, 197–198

    2015

  55. [63]

    Xiang Zhang, Ziyuan Zhao, Theodoros Tsiligkaridis, and Marinka Zitnik. 2022. Self-supervised contrastive pre-training for time series via time-frequency consis- tency.Advances in Neural Information Processing Systems35 (2022), 3988–4003

    2022

  56. [64]

    Yongjie Zhang, Weixin Song, Dehua Shen, and Wei Zhang. 2016. Market reaction to internet news: Information diffusion and price pressure.Economic Modelling 56 (2016), 43–49

    2016

  57. [65]

    Yu Zhang and Qiang Yang. 2021. A survey on multi-task learning.IEEE transac- tions on knowledge and data engineering34, 12 (2021), 5586–5609

    2021

  58. [66]

    Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al . 2023. One fits all: Power general time series analysis by pretrained lm.Advances in neural information processing systems36 (2023), 43322–43355

    2023

  59. [67]

    George Zerveas, Srideepika Jayaraman, Dhaval Patel, Anuradha Bhamidipaty, and Carsten Eickhoff. 2021. A transformer-based framework for multivariate time series representation learning. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2114–2124

    2021

  60. [69]

    Wentao Zhang, Yilei Zhao, Shuo Sun, Jie Ying, Yonggang Xie, Zitao Song, Xinrun Wang, and Bo An. 2024. Reinforcement Learning with Maskable Stock Represen- tation for Portfolio Management in Customizable Stock Pools. InProceedings of the ACM on Web Conference 2024. 187–198

    2024

  61. [74]

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. 2020. Deformable detr: Deformable transformers for end-to-end object detection.arXiv preprint arXiv:2010.04159(2020). A Dataset Introduction Our experiments use five datasets spanning major stock markets a...

    2020 arXiv

  62. [2014]

    Comparison of ARIMA and artificial neural networks models for stock price prediction.Journal of Applied Mathematics2014, 1 (2014), 614342

    2014

  63. [2019]

    Temporal relational ranking for stock prediction.ACM Transactions on Information Systems (TOIS)37, 2 (2019), 1–30

    2019

  64. [2020]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Reinforcement-learning based portfolio management with augmented asset movement prediction states. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 1112–1119

  65. [2022]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 16000–16009

  66. [2024]

    In Proceedings of the AAAI Conference on Artificial Intelligence, Vol

    Master: Market-guided stock transformer for stock price forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 162–170

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.