Pith. sign in

REVIEW 4 major objections 5 minor 43 references

When Dimensionality Hurts: The Role of LLM Embedding Compression for Noisy Regression Tasks

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper argues that the optimal embedding dimension for LLM text regression is set by task noise: compressing to an 8-dimensional autoencoder latent improves financial return prediction, while compression only degrades review and essay…

desk verdict A useful empirical observation about embedding compression and noisy regression, but the signal-to-noise explanation is underdetermined by the cross-task comparison. read the letter →

arxiv 2502.02199 v1 pith:YM5VRIV5 submitted 2025-02-04 cs.CL cs.CEcs.LGq-fin.CP

Felix Drinkall , Janet B. Pierrehumbert , Stefan Zohren This is my paper · ORCID
classification cs.CLcs.CEcs.LGq-fin.CP
keywords embeddingcompressiondimensionalityreductionautoencodersignal-to-noiseratioregressionfinancialreturnpredictionLLMembeddingsregularisation
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 how much you should compress a text embedding depends on how much noise sits between the text and the regression target. In a stock-return prediction task, where news articles are only weakly tied to price moves, compressing a $768$-dimensional LLM embedding through the hidden layer of an autoencoder improves out-of-sample performance, with the best results near $d_z = 8$. In review-scoring and essay-grading tasks, where the text is causally tied to the label, compression never helps and full embeddings are best. The paper further argues that sentiment and emotion features, which have been reported to work well in finance, do not beat autoencoder latents of the same dimension, so their success may be a regularisation effect of compression rather than a special property of the features.

What carries the argument

The central object is an autoencoder trained on the $768$-dimensional all-mpnet-base-v2 embeddings with a mean-squared-error reconstruction loss; its encoder maps $v_i \in \mathbb{R}^{d_{LLM}}$ to a latent $z_i = E(v_i) \in \mathbb{R}^{d_z}$, and varying $d_z$ is the compression dial. The evidence is the shape of the Huber-loss curve (with threshold $\delta = 1$) of a random forest regressor as $d_z$ is varied, compared across a noisy financial task and high-signal review and essay tasks. The paper also defines the 'intrinsic dimension' of a task as the smallest latent dimension that reaches within 10% of the minimum loss, which it estimates at $8$ for reviews and $32$ for essay writing.

What would settle it

Construct variants of one dataset that differ only in label noise -- for instance, replace a controlled fraction of the financial-news target returns with random values or add known signal -- and check whether the optimal autoencoder dimension $d_z$ moves monotonically with that noise fraction; if it does not, the signal-to-noise explanation is wrong.

Watch

Extended reading notes

Core claim

The paper's central claim is that the optimal dimensionality of a text embedding is not a property of the text itself but of the signal-to-noise ratio of the downstream regression task. On a noisy task -- predicting daily stock returns from financial news -- compressing the $768$-dimensional all-mpnet-base-v2 embedding through the learned hidden layer of an autoencoder improves regression performance, with the best results at a latent dimension of $d_z = 8$, and no statistically significant difference for $d_z \in \{4, 16, 32\}$. On high-signal tasks (Yelp and Amazon review scoring, English essay grading), the same compression monotonically hurts: performance is best at the uncompressed dimensionality and degrades as the latent gets smaller. The authors further find that sentiment and emotion probability vectors perform no better than autoencoder latents of the same dimension, and conclude that the reported success of such interpretable features in finance is largely a regularisation effect of dimensionality reduction rather than evidence that the features carry unique signal.

Load-bearing premise

The central claim treats the stock-returns task and the review/essay tasks as differing mainly in signal-to-noise ratio; if their very different training-set sizes, target scales, or task difficulties produce the different curve shapes, the conclusion would not follow.

Editorial extensions

If this is right

  • For noisy regression targets, expecting a convex loss-vs-dimensionality curve means hyperparameter searches should include low latent dimensions, not just the full embedding.
  • For high-signal tasks, dimensionality reduction below the task's intrinsic dimension (about $8$ for reviews, $32$ for essays) sacrifices accuracy; full embeddings or mild compression are preferable.
  • Since performance is statistically indistinguishable for $d_z \in \{4, 16, 32\}$ on the financial task, practitioners have flexibility in choosing a latent size, but extreme values ($1$, $2$, $128$, $256$, $512$) significantly degrade results.
  • Any newly proposed interpretable compressed feature set should be benchmarked against an uninterpretable autoencoder latent of the same size; otherwise its apparent advantage may be pure regularisation.
  • The existence of an intrinsic dimension well below $768$ across all three domains means computation-heavy downstream architectures can operate on compressed inputs without much loss, provided the compression is not too aggressive.

Reading between the lines

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

  • Beyond the paper, the noise-dependence claim predicts a monotone experiment: take one dataset and corrupt a controlled fraction of its targets with random noise, and the optimal $d_z$ should fall as noise rises; no such test appears in the paper.
  • Beyond the paper, because the three task families also differ in training-set size (roughly $30$k financial samples versus $568$k-$650$k reviews) and in target distribution, the reported curve shapes could partly reflect sample size; subsampling the review data to $30$k would separate these explanations.
  • Beyond the paper, the regularisation reinterpretation of sentiment suggests that PCA or even random projection of embeddings to the same dimension would reproduce much of sentiment's benefit on noisy tasks; this is directly testable on the released financial-news dataset.
  • Beyond the paper, a practical extension would be adaptive dimensionality selection: estimate the validation noise of a task and pick $d_z$ accordingly before training a regression head, rather than fixing dimension by model convention.
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 / 5 minor

Summary. The paper studies whether reducing the dimensionality of LLM text embeddings improves regression performance when the input-target relationship is noisy. Using all-mpnet-base-v2 embeddings and a random forest regressor, the authors train an autoencoder to compress embeddings to various latent dimensions dz and evaluate on a newly curated financial news return-prediction dataset (30,115 training articles), three review-scoring datasets (Yelp, App Reviews, Amazon), and an English Language Learning essay-scoring dataset. They report a U-shaped (convex) performance curve on financial returns with a best dimension of dz=8, in contrast to approximately monotone curves on review/writing tasks, which they attribute to different signal-to-noise ratios. They also compare autoencoder latent features with sentiment/emotion class-probability features and argue that the previously reported success of such interpretable features in finance is largely a regularization effect.

Significance. If the main claim held—that optimal embedding dimensionality is governed by the task's signal-to-noise ratio—it would be practically important: it would give a principled reason to compress embeddings in finance and would reinterpret sentiment/emotion features as regularizers rather than as semantically privileged. The paper's strengths are its cross-domain comparison, its release of a financial news dataset, its use of a simple default random forest, and its attempt to attach significance tests to the financial results. However, the central comparison currently confounds noise with sample size, target type, and task difficulty, and the statistical evidence identifies a plateau rather than a sharp optimum. The paper is therefore a useful contribution to an active discussion, but the headline conclusion needs substantially stronger evidence before publication.

major comments (4)
  1. [Section 4.1, Figure 2] The central claim that the optimal embedding dimensionality depends on the signal-to-noise ratio is underdetermined by the evidence presented in Figure 2. The three settings differ simultaneously in training-set size (30,115 for financial returns versus 650k for Yelp, 568k for Amazon, 288k for App Reviews, and roughly 3.3k for ELL), in target type (continuous returns versus integer 1-5 ratings), and in the causal proximity between text and target; signal-to-noise is never measured. With a default random forest on 768 raw features, a 30k-sample noisy task is far more exposed to variance than a 650k-sample task, so the convexity of the financial curve could reflect sample size or target scaling instead of label noise. The paper should support the SNR claim with a within-task manipulation of noise (e.g., injecting label noise into the review tasks at matched sample sizes) or with a direct estimate of per-task SNR.
  2. [Section 4, Figure 1] The paper's own significance test contradicts the precision of the stated optimum: the text reports no statistically significant difference between dz=8 and dz in {4,16,32}, so the result identifies a plateau, not a unique optimum, and the convexity claim should be softened accordingly. Furthermore, the t-tests compare every dimension to the best observed value without any multiple-comparison correction; the authors should report corrected p-values, confidence intervals, or effect sizes for the financial task.
  3. [Section 3.2] The autoencoder is the central method, but Section 3.2 does not specify its architecture (depth, layer widths, activation), optimizer, learning rate, batch size, number of random seeds, or the data split used to train it. It is also not stated whether the autoencoder was fit only on the training inputs or on the full corpus including the test set; if the latter, the evaluation is transductive and the comparison with sentiment/emotion features is not apples-to-apples. The dataset is released, but the code and full training configuration are not; without these, the main experimental result is not reproducible.
  4. [Section 3.2, Figure 1] The claim that sentiment and emotion features perform no better than autoencoder latent features at their respective dimensionalities cannot be verified because the number of classes (and hence the input dimensionality) of the fine-tuned sentiment and emotion classifiers is never stated. The paper should report these dimensionalities explicitly and compare sentiment/emotion features against autoencoder latent features at exactly matched dimensions, including a statement of whether the class-probability vectors were passed through any extra preprocessing such as standardization.
minor comments (5)
  1. [Figure 2] The caption says the curve is 'averaged over Review, English Writing, and Financial Returns tasks,' but Appendix B reports individual review datasets; clarify which datasets compose the 'Review' curve, how the per-task normalization is averaged (mean or median), and what the shaded region or error bars represent.
  2. [Section 4.1] The 'intrinsic dimension' is defined as the dimension at which performance reaches 10% of the minimum loss, but no justification is given for the 10% threshold; at minimum, label this as a heuristic and state its sensitivity to the chosen threshold.
  3. [Appendix C] The MLP equations are given, but the hidden dimension d_mlp, the dropout probability p, and the optimization details are omitted, so the reported negative MLP result is not reproducible; either provide the full configuration or clearly mark the section as preliminary.
  4. [Appendix A.1] The temporal split is described as 'the last 10% of data between 2017 and 2022', but the exact cutoff date is not stated and it is unclear whether articles are deduplicated by company; please specify the exact date boundaries and any deduplication procedure.
  5. [Section 3.3] The phrase 'default parameters' for the random forest is not an exact specification; list the actual default values (e.g., number of trees, max features, minimum samples per leaf) or cite the software version so that the regression setup is reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the autoencoder compression is unsupervised and the regression target never enters the compression objective.

full rationale

Walking the derivation chain: text inputs are embedded by a fixed all-mpnet-base-v2 model; an autoencoder is trained with reconstruction loss LAE = (1/N) sum ||vi - vhat_i||^2, which does not involve the regression target yi. The downstream random forest is then trained on the compressed latent zi with target yi. Thus the reported performance differences across latent dimensions are empirical outcomes of a supervised downstream model applied to an unsupervised compression, not consequences of fitting the target into the compression. The sentiment/emotion comparisons use class probabilities from externally fine-tuned DistilRoBERTa models, which are also not fitted to the financial target, so the claim that interpretable features' gains may be a regularization effect is an interpretation of the comparison, not a definitional identity. The paper's self-citations (Drinkall et al. 2022, 2024, 2025) are contextual examples of prior uses of compressed text representations or generative-model limitations; none is load-bearing for the main derivation. The term 'intrinsic dimension' is introduced as a descriptive criterion (dimension at which performance reaches 10% of minimum loss), not as an input assumption smuggled into the results. The main weakness is that the cross-task comparison in Section 4.1 varies sample size, target distribution, and task difficulty alongside the intended signal-to-noise axis, so the SNR conclusion is underdetermined; however, this is a validity or confound concern, not circularity. No equation or fitted parameter is shown to be equivalent to an output by construction, so the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. Its central claim relies on domain assumptions about which tasks are noisy, the sufficiency of reconstruction loss for downstream tasks, and the validity of the random forest setup. The chosen latent dimension is read off the data, which makes it a free parameter in the empirical sense.

free parameters (3)
  • Autoencoder latent dimension d_z = 8 (d_z in {4,16,32} statistically indistinguishable)
    The paper reads the optimal dimension off Figure 1 on the financial task; the central claim depends on this empirical choice.
  • Huber loss threshold delta = 1
    Chosen as a standard robust loss; not central to the claim.
  • Intrinsic dimension threshold = 10% of minimum loss
    Ad hoc threshold used to define intrinsic dimension in Section 4.1; affects the statement that near-optimal performance is reached at low dimensions.
assumptions (4)
  • domain assumption News articles in the financial dataset have low signal-to-noise relative to returns.
    Section 2.1. The entire noisy versus high-signal comparison rests on this classification.
  • domain assumption Reviews and essays have a direct causal link between text and target score.
    Section 2.2. Used to define the high-signal baseline; no direct measurement of causality is given.
  • domain assumption The autoencoder reconstruction objective preserves information relevant to the downstream regression.
    Section 3.2. The compressed latent is used as features, but the paper does not verify that lost dimensions are irrelevant to the task.
  • domain assumption Random forest with default parameters is an appropriate regressor for all tasks.
    Section 3.3. No hyperparameter tuning is performed; different tasks may require different settings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Dimensionality Hurts: The Role of LLM Embedding Compression for Noisy Regression Tasks." pith.science (2026). https://pith.science/paper/YM5VRIV5

@misc{pith2026250202199,
  author       = {Pith},
  title        = {Pith review of: When Dimensionality Hurts: The Role of LLM Embedding Compression for Noisy Regression Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YM5VRIV5}},
  note         = {Machine review of arXiv:2502.02199}
}
read the original abstract

Large language models (LLMs) have shown remarkable success in language modelling due to scaling laws found in model size and the hidden dimension of the model's text representation. Yet, we demonstrate that compressed representations of text can yield better performance in LLM-based regression tasks. In this paper, we compare the relative performance of embedding compression in three different signal-to-noise contexts: financial return prediction, writing quality assessment and review scoring. Our results show that compressing embeddings, in a minimally supervised manner using an autoencoder's hidden representation, can mitigate overfitting and improve performance on noisy tasks, such as financial return prediction; but that compression reduces performance on tasks that have high causal dependencies between the input and target data. Our results suggest that the success of interpretable compressed representations such as sentiment may be due to a regularising effect.

Figures

Figures reproduced from arXiv: 2502.02199 by the authors.

Figure 1
Figure 1. Huber Loss on the noisy financial returns task [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. The normalised Huber loss of each dataset that makes up the result in Figure [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Cosine similarity between vi and vˆi on the financial returns dataset. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 19 canonical work pages

  1. [1]

    Werner Antweiler and Murray Z Frank. 2004. https://doi.org/10.1111/j.1540-6261.2004.00662.x Is All That Talk Just Noise? The Information Content of Internet Stock Message Boards . The Journal of Finance

    work page arXiv 2004

  2. [2]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. https://openreview.net/forum?id=IW1PR7vEBf LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders . In First Conference on Language Modeling

    work page 2024

  3. [3]

    Fischer Black. 1986. https://onlinelibrary.wiley.com/doi/full/10.1111/j.1540-6261.1986.tb04513.x Noise . The Journal of Finance, 41(3):528--543

    work page arXiv 1986

  4. [4]

    L. Breiman. 2001. https://doi.org/10.1023/A:1010933404324 Random Forests . Machine Learning, 45:5--32

    doi:10.1023/a:1010933404324 2001

  5. [5]

    Kaiyan Chang, Songcheng Xu, Chenglong Wang, Yingfeng Luo, Xiaoqian Liu, Tong Xiao, and Jingbo Zhu. 2024. https://arxiv.org/abs/2404.01077 Efficient Prompting Methods for Large Language Models: A Survey . Preprint, arXiv:2404.01077

    arXiv 2024

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies

    doi:10.18653/v1/n19-1423 2019

  7. [7]

    Forecasting Credit Ratings: A Case Study where Traditional Methods Outperform Generative LLMs

    Felix Drinkall, Janet B. Pierrehumbert, and Stefan Zohren. 2025. https://arxiv.org/abs/2407.17624 Forecasting Credit Ratings: A Case Study where Traditional Methods Outperform Generative LLMs . In Proceedings of the FinNLP-FNP-LLMFinLegal Workshop @ COLING 2025. Association for Computational Linguistics

    work page Pith review arXiv 2025

  8. [8]

    Felix Drinkall, Stefan Zohren, and Janet Pierrehumbert. 2022. https://doi.org/10.18653/v1/2022.naacl-main.105 Forecasting COVID -19 Caseloads Using Unsupervised Embedding Clusters of Social Media Posts . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Associ...

    work page doi:10.18653/v1/2022.naacl-main.105 2022

Show all 43 references
  1. [9]

    Bledar Fazlija and Pedro Harder. 2022. https://doi.org/10.3390/math10132156 Using Financial News Sentiment for Stock Price Direction Prediction . Mathematics, 10(13)

    2022 doi

  2. [10]

    Alex Franklin, Maggie, Meg Benner, Natalie Rambis, Perpetual Baffour, Ryan Holbrook, Scott Crossley, and Ulrich Boser. 2022. https://kaggle.com/competitions/feedback-prize-english-language-learning Feedback Prize - English Language Learning

    2022

  3. [11]

    González-Castaño, and Enrique Costa-Montenegro

    Silvia García-Méndez, Francisco de Arriba-Pérez, Ana Barros-Vila, Francisco J. González-Castaño, and Enrique Costa-Montenegro. 2023. https://doi.org/10.1007/s10489-023-04452-4 Automatic Detection of Relevant Information, Predictions and Forecasts in Financial News Through Topi...

    2023 doi

  4. [12]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, et al. 2024. https://arxiv.org/abs/2407.21783 The Llama 3 Herd of Models . Preprint, arXiv:2407.21783

    2024 arXiv

  5. [13]

    Felix Hamborg, Norman Meuschke, Corinna Breitinger, and Bela Gipp. 2017. https://doi.org/10.5281/zenodo.4120316 news-please: A Generic News Crawler and Extractor . In Proceedings of the 15th International Symposium of Information Science, pages 218--223

    2017 doi

  6. [14]

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, et al. 2022. https://dl.acm.org/doi/10.5555/3600270.3602446 Training Compute-optimal Large Language Models . In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY...

    2022

  7. [15]

    Baizhou Huang, Shuai Lu, Xiaojun Wan, and Nan Duan. 2024. https://doi.org/10.18653/v1/2024.acl-long.78 Enhancing Large Language Models in Coding Through Multi-Perspective Self-Consistency . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguist...

    2024 doi

  8. [16]

    Peter J. Huber. 1964. https://doi.org/10.1214/aoms/1177703732 Robust Estimation of a Location Parameter . Annals of Statistics, 53(1):73--101

    1964

  9. [17]

    Stig Johan Berggren, Taraka Rama, and Lilja vrelid. 2019. https://doi.org/10.18653/v1/W19-4409 Regression or Classification? Automated Essay Scoring for N orwegian . In Proceedings of the Fourteenth Workshop on Innovative Use of NLP for Building Educational Applications, pages...

    2019 doi

  10. [18]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. https://arxiv.org/abs/2001.08361 Scaling Laws for Neural Language Models . Preprint, arXiv:2001.08361

    2020 arXiv

  11. [19]

    Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. 2024. https://arxiv.org/abs/2405.17428 NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models . Preprint, arXiv:2405.17428

    2024 arXiv

  12. [20]

    Cohn, and Janet B

    Fangru Lin, Emanuele La Malfa, Valentin Hofmann, Elle Michelle Yang, Anthony G. Cohn, and Janet B. Pierrehumbert. 2025. https://dl.acm.org/doi/10.5555/3692070.3693283 Graph-enhanced Large Language Models in Asynchronous Plan Reasoning . In Proceedings of the 41st International...

    2025

  13. [21]

    Xiao Liu, Zirui Wu, Xueqing Wu, Pan Lu, Kai-Wei Chang, and Yansong Feng. 2024. https://doi.org/10.18653/v1/2024.findings-acl.548 Are LLMs Capable of Data-based Statistical and Causal Reasoning? Benchmarking Advanced Quantitative Reasoning with Data . In Findings of the Associa...

    2024 doi

  14. [22]

    Ivanova, Idan A

    Kyle Mahowald, Anna A. Ivanova, Idan A. Blank, Nancy Kanwisher, Joshua B. Tenenbaum, and Evelina Fedorenko. 2024. https://doi.org/10.1016/j.tics.2024.01.011 Dissociating Language and Thought in Large Language Models . Trends in Cognitive Sciences, 28(6):517--540

    2024 doi

  15. [23]

    Daniel Mantilla-García and Varadharajan Vaidyanathan. 2017. https://doi.org/10.1007/s11408-017-0290-3 Predicting Stock Returns in the Presence of Uncertain Structural Changes and Sample Noise . Financial Markets and Portfolio Management, 31

    2017 doi

  16. [24]

    McAuley and J

    J. McAuley and J. Leskovec. 2013. https://arxiv.org/abs/1303.4402 From Amateurs to Connoisseurs: Modeling the Evolution of User Expertise Through Online Reviews . In Proceedings of the 22nd International Conference on World Wide Web (WWW), pages 897--908. ACM

    2013 arXiv

  17. [25]

    McInnes , J

    L. McInnes , J. Healy , and J. Melville . 2018. https://arxiv.org/abs/1802.03426 UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction . ArXiv e-prints

    2018 arXiv

  18. [26]

    Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. https://doi.org/10.18653/v1/2023.eacl-main.148 MTEB: Massive Text Embedding Benchmark . In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, Dubrov...

    2023 doi

  19. [27]

    Muhammad Qorib, Geonsik Moon, and Hwee Tou Ng. 2024. https://doi.org/10.18653/v1/2024.findings-acl.967 Are Decoder-Only Language Models Better than Encoder-Only Language Models in Understanding Word Meaning? In Findings of the Association for Computational Linguistics: ACL 202...

    2024 doi

  20. [28]

    Eghbal Rahimikia and Felix Drinkall. 2024. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4963618 Re (Visiting) Large Language Models in Finance . Available at SSRN

    2024

  21. [29]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/D19-1410 Sentence- BERT : Sentence embeddings using S iamese BERT -networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference...

    2019 doi

  22. [30]

    Marie-Hélène Roy and Denis Larocque. 2012. https://doi.org/10.1080/10485252.2012.715161 Robustness of Random Forests for Regression . Journal of Nonparametric Statistics, 24(4):993--1006

    2012

  23. [31]

    Ramit Sawhney, Shivam Agarwal, Arnav Wadhwa, and Rajiv Ratn Shah. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.676 Deep Attentive Learning for Stock Movement Prediction From Social Media Text and Company Correlations . In Proceedings of the 2020 Conference on Empirical Me...

    2020 doi

  24. [32]

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/c3a690be93aa602ee2dc0ccab5b7b67e-Paper.pdf MPNet: Masked and Permuted Pre-training for Language Understanding . In Advances in Neural Information Proces...

    2020

  25. [33]

    Student. 1908. https://www.jstor.org/stable/2331554 The Probable Error of a Mean . Biometrika, pages 1--25

    1908

  26. [34]

    Eric Tang, Bangding Yang, and Xingyou Song. 2024. https://arxiv.org/abs/2411.14708 Understanding LLM Embeddings for Regression . Preprint, arXiv:2411.14708

    2024 arXiv

  27. [35]

    Yingjie Tian and Yuqi Zhang. 2022. https://doi.org/10.1016/j.inffus.2021.11.005 A Comprehensive Survey on Regularization Strategies in Machine Learning . Information Fusion, 80:146--166

    2022 doi

  28. [36]

    Sonja Tilly, Markus Ebner, and Giacomo Livan. 2021. https://doi.org/10.1016/j.eswa.2021.114760 Macroeconomic Forecasting Through News, Emotions and Narrative . Expert Systems with Applications, 175:114760

    2021

  29. [37]

    Tjong Kim Sang and Fien De Meulder

    Erik F. Tjong Kim Sang and Fien De Meulder. 2003. https://www.aclweb.org/anthology/W03-0419 Introduction to the C o NLL -2003 Shared Task: Language-Independent Named Entity Recognition . In Proceedings of the Seventh Conference on Natural Language Learning at HLT - NAACL 2003

    2003

  30. [38]

    Laurens van der Maaten and Geoffrey Hinton. 2008. http://jmlr.org/papers/v9/vandermaaten08a.html Visualizing Data using t-SNE . Journal of Machine Learning Research, 9(86):2579--2605

    2008

  31. [39]

    Tongshuang Wu, Michael Terry, and Carrie Jun Cai. 2022. https://doi.org/10.1145/3491102.3517582 AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts . In Proceedings of the 2022 CHI Conference on Human Factors in Computing Syste...

    2022

  32. [40]

    Fuzhao Xue, Yao Fu, Wangchunshu Zhou, Zangwei Zheng, and Yang You. 2023. https://arxiv.org/abs/2305.13230 To Repeat or Not To Repeat: Insights from Scaling LLM Under Token-crisis . In Proceedings of the 37th International Conference on Neural Information Processing Systems, NI...

    2023 arXiv

  33. [41]

    Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. https://proceedings.neurips.cc/paper_files/paper/2015/file/250cf8b51c773f3f8dc8b4be867a9a02-Paper.pdf Character-level Convolutional Networks for Text Classification . In Advances in Neural Information Processing Systems. Curran As...

    2015

  34. [42]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  35. [43]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.