Pith. sign in

REVIEW 4 major objections 7 minor 54 references

Interpretable Company Similarity with Sparse Autoencoders

T0 review · 4 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Sparse autoencoder features extracted from an LLM's internal representation of company reports cluster companies more coherently by stock-return co-movement than SIC codes, broader industry codes, or text embeddings, and the clusters are…

desk verdict Promising first use of SAEs for company similarity, with real code/data contributions; but the 'strictly out-of-sample' GCDR result is tainted by PCA fit on the full sample and the headline metric is tuned on itself, so the abstract overstates the evidence. read the letter →

arxiv 2412.02605 v3 pith:QDDACYBK submitted 2024-12-03 cs.CL cs.LGecon.GNq-fin.EC

Marco Molinari , Victor Shao , Luca Imeneo , Mateusz Mikolajczak , Vladimir Tregubiak , Abhimanyu Pandey , Sebastian Kuznetsov Ryder Torres Pereira This is my paper · ORCID
classification cs.CLcs.LGecon.GNq-fin.EC
keywords companysimilaritysparseautoencodersinterpretabilityclusteringpairstradingsectorclassificationcointegrationLLMactivations
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 sets out to show that sparse autoencoder (SAE) features, taken from a large language model's internal representation of a company's annual report, give a more useful and more interpretable measure of company similarity than static sector codes or text embeddings. The evidence is that companies grouped by SAE features have more highly correlated monthly stock returns than companies grouped by SIC codes, broader industry codes, or BERT/SBERT/PaLM embeddings, and that pairs-trading strategies built on SAE clusters earn higher risk-adjusted returns in backtests. Because SAE features are sparse, each cluster can be explained by a small set of meaningful features, which removes the main barrier to using machine-learned similarity in high-stakes financial decisions. If the claim holds, risk managers and portfolio managers would have a granular, updateable, explainable alternative to industry classifications.

What carries the argument

The central object is the summed sparse feature vector. A sparse autoencoder writes each token's dense activation as a sparse linear combination of interpretable features; the paper sums those feature activations across all tokens in a company's annual report to make one vector per company-year. The summed vector is compressed with principal component analysis (4000 components, capturing 89.92% of variance), and company-year vectors are clustered by building a minimum spanning tree with normalized cosine distance and pruning edges above a cutoff threshold, so connected components become the clusters. The cutoff is set by temporal cross-validation for the main variant and by a five-year rolling lookback for the out-of-sample variant. This mechanism converts a sparse, variable-length representation into a fixed-length, clusterable space whose clusters can be scored against return co-movement and explained by a few important features.

What would settle it

Recompute the rolling out-of-sample clusters with the dimension-reduction transform refit only on the preceding five years; if the mean intra-cluster correlation falls from roughly 0.39 to the embedding baseline (~0.22) or the SIC baseline (~0.23), the claimed superiority rests on lookahead rather than on the sparse features. Alternatively, regress monthly returns on market and sector factors and check whether SAE clusters still explain residual co-movement; if they do not, the metric is capturing shared momentum or liquidity, not fundamentals.

Watch

Extended reading notes

Core claim

The paper's central claim is that decomposing an LLM's internal representation of a company description with a sparse autoencoder yields a feature space whose clusters match stock-return co-movement better than any of the compared baselines. The evidence is a mean intra-cluster Pearson correlation of logged monthly returns: 0.359 for the main SAE clusters and 0.385 for the rolling out-of-sample variant (0.391 in the appendix), versus 0.198–0.219 for embedding-based clusters and 0.187–0.231 for SIC/BISC codes. In an out-of-sample pairs-trading backtest, the SAE-based clusters produce a Sharpe ratio of 12.18 (9.69 for the rolling variant) versus 7.58–10.57 for the alternatives. The authors also report that the features most important for the clusters are interpretable at 77–80% under an automated fuzzing test, so the similarity signal is not a black box.

Load-bearing premise

The main result stands on two assumptions: that the dimension-reduction step is fit without using the years being scored, and that stocks whose monthly returns move together are fundamentally similar; if either fails, the reported out-of-sample advantage is not a clean measure of fundamental similarity.

Editorial extensions

If this is right

  • Risk management and hedging could use SAE clusters instead of waiting for classification agencies to update sector codes, since the measure can be recomputed whenever a new annual report is filed.
  • Pairs-trading strategies that select co-integrated pairs from SAE clusters should suffer fewer non-convergence losses, because the clusters are chosen to align with return co-movement and the backtest Sharpe ratios are higher.
  • Because the clusters come with a small set of important sparse features, investment decisions based on similarity can be explained to clients or regulators in terms of concrete textual concepts.
  • The method works for documents of arbitrary length, removing a practical constraint that embedding-based similarity measures face when reports vary widely in size.

Reading between the lines

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

  • If the PCA is refit on a strict rolling basis using only past years, the out-of-sample correlation may fall below the reported 0.385/0.391, so the size of the advantage over embeddings should be treated as an upper bound until that check is done.
  • The intra-cluster return correlation metric may partly capture shared sector momentum or liquidity shocks rather than firm-level fundamentals; controlling for market and sector factors would show whether the residual co-movement remains.
  • The same feature-summing pipeline could transfer to other high-stakes domains, such as grouping clinical trial descriptions or legal contracts by interpretable characteristics.
  • The observed exponential decay in feature activation frequencies suggests a much smaller set of features may carry most of the signal, which could lead to cheaper inference and simpler explanations than the full feature space.
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 / 7 minor

Summary. The paper proposes using sparse autoencoder (SAE) features extracted from Llama 3.1 8B activations on SEC company descriptions to measure company similarity. The authors sum SAE feature activations across tokens, apply PCA, and cluster the resulting vectors via minimum spanning trees with a cosine-distance cutoff threshold. They compare these clusters (GCD and its rolling variant GCDR) against SIC/BISC codes and three embedding baselines (BERT, SBERT, PaLM-gecko) using the mean intra-cluster Pearson correlation of logged monthly returns (MC) and Sharpe ratios from a pairs-trading backtest. They report that SAE-based clusters achieve higher MC (0.359 for GCD, 0.385 for GCDR) and higher Sharpe ratios than all baselines, and they provide an interpretability analysis showing that a small set of features explains clusters. Code and data are released.

Significance. If the claims hold, this paper makes a valuable contribution by showing that interpretable SAE features can outperform both discrete sector classifications and dense embeddings for company similarity, with potential applications in risk management and portfolio construction. The authors provide an open dataset and code, which is a strength. The interpretability evaluation, including the fuzzing-based assessment, is a useful addition to the literature. However, the significance is tempered by the methodological concerns below, particularly the questionable out-of-sample validity of the main evaluation and the selection of hyperparameters on the evaluation metric.

major comments (4)
  1. [Section 3.3.1] The statement 'We fit PCA globally across 1996–2020 for consistent eigenvectors' means that the feature representation for each year's clusters is estimated using data from all years, including the evaluation year. Consequently, the GCDR results in Table 1 and Appendix D, which are described as 'strict out-of-sample' and 'forward-looking,' are not clean out-of-sample estimates; the clusters can encode information from outside the evaluation window. This is load-bearing because the superiority of the SAE features over GPaLM (0.219) and GSIC (0.231) is the central claim, and the robustness argument rests on the OOS variant. Please re-fit PCA on an expanding or rolling window that excludes the evaluation year, or otherwise demonstrate that the eigenvectors are stable across years, and re-report the rolling OOS results.
  2. [Section 3.3.1 and Appendix B] The evaluation metric MC is used as the objective function for selecting the MST cutoff threshold θ (Section 3.3.1) and for tuning the UMAP and spectral clustering hyperparameters of the embedding baselines (Appendix B). Thus the reported MC differences are partly a result of model selection on the test metric, not a prediction of generalization. The paper should use a nested validation scheme (e.g., selecting hyperparameters on a training period and reporting MC on a holdout period) or quantify the selection bias by reporting results with fixed or default hyperparameters. Without this, the 'surpass' claim is not fully supported.
  3. [Table 1 vs Appendix D] There is a numerical inconsistency: Table 1 reports MC(GCDR)=0.385, while Appendix D states 'a higher overall mean correlation of MC(GCDR) = 0.391.' The same quantity should not have two different values in the same paper. Please reconcile the discrepancy and ensure consistency across the main text, tables, and appendices.
  4. [Section 4 results] The paper reports differences in MC and Sharpe ratios without confidence intervals, significance tests, or any uncertainty quantification. For example, the difference between GCD (0.359) and GPaLM-gecko (0.219) may or may not be statistically significant given the number of clusters and years. The paper's own limitation section (Section 4.4) acknowledges that Sharpe ratios may be sensitive to θ, slippage, regime shifts, and finite-sample bias, yet no such analysis is provided. Please add bootstrap or other resampling-based uncertainty estimates for the mean correlations and Sharpe ratios, or at least a sensitivity analysis to the chosen hyperparameters.
minor comments (7)
  1. [Section 3.2] The feature-summing aggregation is motivated but not ablated against alternatives such as max pooling or mean pooling; reporting such an ablation would strengthen the claim that summing is the appropriate choice.
  2. [Figure 2 caption] The caption states that PaLM and SIC-codes are used because they have the highest MC among the embedding-based and traditional benchmark groups, but it would be helpful to also show the other baselines in the figure or explain why they are omitted.
  3. [Section 4.3 / Table 2] The interpretability comparison with Gemma 2 9B on Red Pajama and The Pile is not apples-to-apples because the features are from a different model and corpus; please state this caveat explicitly in the text.
  4. [Section 3.4] The pairs-trading backtest assumes zero transaction costs and ignores slippage; the limitation section mentions this, but the main text should remind the reader when presenting the Sharpe ratios.
  5. [Equation (3)] The TopK activation function is not formally defined; please specify the value of k (the text mentions top 128) and clarify how ties are handled.
  6. [Appendix B] The description of temporal cross-validation for hyperparameter optimization is vague; please specify whether the Optuna search uses only past folds or the full dataset, as this affects the interpretation of the baseline results.
  7. [Figure 6] The caption uses 'MCOOS_y' while the text uses 'MC(y)(θ*_y)'; please unify the notation.

Circularity Check

2 steps flagged · score 4.0 of 10

The headline MC metric is partly selected rather than predicted: the MST cutoff is tuned on MC, and the 'strict out-of-sample' rolling variant still uses a PCA representation fit on the full 1996–2020 sample; independent pairs-trading and fixed SIC/BISC baselines keep the central claim from being fully circular.

  1. fitted input called prediction [Section 3.3.1 (Cut-off θ calibration) and Table 1]
    "To determine the MST cut-off threshold θ for GCD, we initially apply a two-fold temporal cross-validation scheme: θ is chosen to maximize the average intra-cluster correlation across two time periods covering 25% and 50% of our dataset. We define this as GCD."

    The reported headline metric MC(GCD)=0.359 in Table 1 is the same intra-cluster return correlation used as the objective to select θ. The non-rolling GCD result is therefore a selected optimum of the evaluation metric on the data used for calibration, not an independent prediction. The abstract claims that SAE features 'surpass sector classifications and embeddings... in correlating logged monthly returns' based partly on this selected value. The rolling GCDR variant mitigates the selection for θ, but the table and abstract do not separate the in-sample-selected GCD result from the out-of-sample variant when making the central claim.

  2. fitted input called prediction [Section 3.3.1 (PCA sentence) and Appendix D]
    "We fit PCA globally across 1996–2020 for consistent eigenvectors, ncomponents = 4000 captures 89.92% variance. ... These findings confirm the robustness of our sparse-feature clusters under forward-looking evaluation."

    Although the rolling threshold θ*_y is chosen from a five-year lookback only, every yearly sparse-feature vector used for the MST is projected with PCA eigenvectors estimated on the full 1996–2020 sample, including the evaluation year being scored. Thus the 'strict out-of-sample' and 'forward-looking' labels in Section 3.3.1 and Appendix D apply only to the cutoff, not to the feature representation: the representation entering the out-of-sample clusters can carry information from the very year whose correlation is reported. The OOS correlation is therefore not a clean out-of-sample estimate of the full pipeline.

full rationale

The core derivation is not circular by definition: the SAE features come from an external model (Llama 3.1 8B with an EleutherAI SAE), the similarity metric MC is an external financial quantity (correlation of logged monthly returns), and the fixed SIC/BISC baseline clusters are not constructed from the paper's features. There is no self-citation chain or uniqueness theorem forcing the choice of SAE features. However, two fitted inputs contaminate the headline evaluation. First, the MST cutoff θ for GCD is explicitly chosen to maximize intra-cluster return correlation, and the same metric is then reported as evidence of superiority without a separate held-out evaluation for that variant. Second, the rolling out-of-sample variant GCDR, which is presented as the robustness confirmation, still relies on PCA eigenvectors fit globally across 1996–2020, so the 'out-of-sample' years are not out-of-sample for the representation. These issues make the primary MC comparison partially selected rather than predicted. The paper does retain independent content: the pairs-trading backtest (2014–2020) uses a separate out-of-sample period and co-integration logic, the comparison against fixed SIC/BISC codes is externally grounded, and the interpretability fuzzing results are benchmarked against an external Gemma 2 9B pipeline. For these reasons, the circularity is moderate rather than total, and the central claim does not reduce entirely to its inputs.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim rests on: (i) the return-correlation proxy, (ii) global PCA fitting without per-year retraining, (iii) feature summing as a valid aggregation, and (iv) zero-transaction-cost trading as evidence. The θ and embedder hyperparameters are fitted to the evaluation metric, which belongs in the free-parameter tally.

free parameters (7)
  • MST cutoff threshold θ (GCD) = -3.13
    Chosen to maximize mean intra-cluster correlation on two temporal cross-validation folds covering 25% and 50% of the dataset; gates cluster granularity and drives the headline MC(GCD)=0.359 (Section 3.3.1, Appendix C).
  • Per-year MST cutoff θ*_y (GCDR) = varies per year, grid -4.5 to -1.0
    Selected per year by maximizing MC over a five-year rolling lookback; used for the OOS evaluation MC(GCDR) (Section 3.3.1).
  • UMAP n_components (baselines) = 7 (BERT), 7 (SBERT), 6 (PaLM-gecko)
    Tuned by Optuna with TPE to maximize MC(Gk) for embedding clusters (Appendix B, Table 3).
  • UMAP n_neighbors and min_dist (baselines) = 119/0.109 (BERT), 79/0.012 (SBERT), 40/0.120 (PaLM)
    Same Optuna MC-maximizing search (Appendix B, Table 3).
  • Spectral n_clusters (baselines) = 10 (BERT), 49 (SBERT), 27 (PaLM)
    Tuned via Optuna; k fixed to 5 following Vamvourellis et al. (Appendix B, Table 4).
  • PCA n_components for sparse features = 4000
    Chosen to capture 89.92% variance globally across 1996-2020; affects downstream cosine distances (Section 3.3.1).
  • SAE layer and TopK = layer 30 of 32, TopK=128
    Layer chosen at 90% depth for abstract features; TopK comes from the off-the-shelf EleutherAI SAE, not tuned to the task (Section 3.2).
assumptions (6)
  • domain assumption Intra-cluster Pearson correlation of logged monthly returns is a valid proxy for fundamental company similarity.
    Defines the main evaluation metric MC in Section 3.3; if return co-movement is driven by transient sector rotation or liquidity, the headline result does not establish fundamental similarity.
  • domain assumption Globally fit PCA (1996-2020) does not materially leak future information into yearly clusters.
    Stated in Section 3.3.1; the OOS claim of GCDR assumes this transformation is innocuous for yearly cluster formation.
  • ad hoc to paper Summing sparse feature activations across tokens yields a meaningful magnitude-scaled frequency representation.
    Proposed in Section 3.2 as an alternative to max-pooling; the paper motivates it with financial word-count analogies but does not validate it against other aggregations for the downstream task.
  • domain assumption Cosine distance on PCA-transformed summed features is a reliable similarity metric for clustering.
    Used to define MST edge weights in Section 3.3.1; the discontinuity of TopK sparsity is cited as the reason PCA is applied first.
  • domain assumption Pairs trading with zero transaction costs and a ±1σ/±2σ spread rule isolates fundamental similarity.
    Trading details in Appendix E; the absence of transaction costs and borrow costs makes Sharpe ratios optimistic (acknowledged in Section 4.4).
  • domain assumption The EleutherAI SAE on Llama 3.1 faithfully decomposes activations into interpretable features relevant to financial semantics.
    The method treats this off-the-shelf SAE as a reliable feature extractor (Sections 2.1, 3.2), inherited from prior interpretability work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Company Similarity with Sparse Autoencoders." pith.science (2026). https://pith.science/paper/QDDACYBK

@misc{pith2026241202605,
  author       = {Pith},
  title        = {Pith review of: Interpretable Company Similarity with Sparse Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QDDACYBK}},
  note         = {Machine review of arXiv:2412.02605}
}
read the original abstract

Determining company similarity is a vital task in finance, underpinning risk management, hedging, and portfolio diversification. Practitioners often rely on sector and industry classifications such as SIC and GICS codes to gauge similarity, the former being used by the U.S. Securities and Exchange Commission (SEC), and the latter widely used by the investment community. Since these classifications lack granularity and need regular updating, using clusters of embeddings of company descriptions has been proposed as a potential alternative, but the lack of interpretability in token embeddings poses a significant barrier to adoption in high-stakes contexts. Sparse Autoencoders (SAEs) have shown promise in enhancing the interpretability of Large Language Models (LLMs) by decomposing Large Language Model (LLM) activations into interpretable features. Moreover, SAEs capture an LLM's internal representation of a company description, as opposed to semantic similarity alone, as is the case with embeddings. We apply SAEs to company descriptions, and obtain meaningful clusters of equities. We benchmark SAE features against SIC-codes, Industry codes, and Embeddings. Our results demonstrate that SAE features surpass sector classifications and embeddings in capturing fundamental company characteristics. This is evidenced by their superior performance in correlating logged monthly returns - a proxy for similarity - and generating higher Sharpe ratios in co-integration trading strategies, which underscores deeper fundamental similarities among companies. Finally, we verify the interpretability of our clusters, and demonstrate that sparse features form simple and interpretable explanations for our clusters.

Figures

Figures reproduced from arXiv: 2412.02605 by the authors.

Figure 1
Figure 1. Distribution of summed feature activations. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall Mean Correlation (MC(Gk)) of GCD (Normalized Cosine Distance Cluster Group) vs PaLM vs SIC Benchmarks between 1996-2020. Note that we use PaLM and SIC-codes for comparison, as they have the highest MC(Gk) among the embedding-based and traditional benchmark groups, respectively. 4 Results 4.1 Clustering results For each clustering method group Gk, we evalu￾ate their MC(Gk), and Sharpe Ratios (see [PITH_FULL_… view at source ↗
Figure 3
Figure 3. Interpretability Score of Features by Percent [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Interpretability Score of Features by Percent [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Optuna Study – Histogram of Sparse Fea￾tures’ MST cutoff thresholds. Maximizing Threshold = -3.130 [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: Distribution of the proportion of important [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 6
Figure 6. Figure 6: Walk-forward tuning results for the sparse [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 28 canonical work pages

  1. [1]

    Bailey and Marcos L \'o pez de Prado

    David H. Bailey and Marcos L \'o pez de Prado. 2012. https://doi.org/10.2139/ssrn.1821643 The sharpe ratio efficient frontier . Journal of Risk, 15(2):3--44. Available at SSRN: https://ssrn.com/abstract=1821643

    work page doi:10.2139/ssrn.1821643 2012

  2. [2]

    Benjamin Bischke, Patrick Helber, Damian Borth, and Andreas Dengel. 2019. https://www.dfki.de/fileadmin/user_upload/import/10526_2019_bischke_igarss_final.pdf Multi-task learning for disaster image classification . In 2019 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pages 224--227. IEEE

    work page 2019

  3. [3]

    Bonanno, G

    G. Bonanno, G. Caldarelli, F. Lillo, S. Miccicche, N. Vandewalle, and R. N. Mantegna. 2004. https://doi.org/10.1140/epjb/e2004-00129-6 Networks of equities in financial markets . The European Physical Journal B - Condensed Matter, 38(2):363–371

    doi:10.1140/epjb/e2004-00129-6 2004

  4. [4]

    Trenton Bricken, Jonathan Marcus, Siddharth Mishra-Sharma, Meg Tong, Ethan Perez, Mrinank Sharma, Kelley Rivoire, and Thomas Henighan. 2024. Using dictionary learning features as classifiers. Technical report, Anthropic

    work page 2024

  5. [5]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, and Adam Jermyn. 2023. https://transformer-circuits.pub/2023/monosemantic-features/index.html Towards monosemanticity: Decomposing language models with dictionary learning

    work page 2023

  6. [6]

    Valentin Buchner, Lele Cao, Jan-Christoph Kalo, and Vilhelm Von Ehrenheim. 2024. https://doi.org/10.18653/v1/2024.naacl-industry.10 Prompt tuned embedding classification for industry sector allocation . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume ...

    work page doi:10.18653/v1/2024.naacl-industry.10 2024

  7. [7]

    Shuangshuang Chen and Wei Guo. 2023. https://doi.org/10.3390/math11081777 Auto-encoders in deep learning—a review with new perspectives . Mathematics, 11(8)

    work page doi:10.3390/math11081777 2023

  8. [8]

    Wanghu Chen, Huijun Li, Jing Li, and Ali Arshad. 2020. https://doi.org/10.1109/BigData50022.2020.9378325 Autoencoder-based outlier detection for sparse, high dimensional data . In 2020 IEEE International Conference on Big Data (Big Data), pages 2735--2742

    arXiv 2020

Show all 54 references
  1. [9]

    Chowdhery. 2022. https://arxiv.org/abs/2204.02311 Palm: Scaling language modeling with pathways . Preprint, arXiv:2204.02311

    2022 arXiv

  2. [10]

    Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adri \`a Garriga-Alonso. 2023. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36:16318--16352

    2023

  3. [11]

    Danile Delphini, Stefano Battiston, Guido Caldarelli, and Massimo Raccaboni. 2019. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0217141 Systemic risk from investment similarities . PLOS ONE

    2019 doi

  4. [12]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://arxiv.org/abs/1810.04805 Bert: Pre-training of deep bidirectional transformers for language understanding . Preprint, arXiv:1810.04805

    2019 arXiv

  5. [13]

    Douglas S. Ehrman. 2012. https://doi.org/doi.org/10.1002/9781119201526.ch6 The Handbook of Pairs Trading . Wiley Trading

    2012 doi

  6. [14]

    EleutherAI. 2024. Sae-llama-3-8b-32x. Hugging Face. Model card: "This is a set of sparse autoencoders (SAEs) trained on the residual stream of Llama 3 8B using the RedPajama corpus. The SAEs are organized by layer, and can be loaded using the EleutherAI sae library." Retrieved...

    2024

  7. [15]

    Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. 2022. https://arxiv.org...

    2022 arXiv

  8. [16]

    Engle and C

    Robert F. Engle and C. W. J. Granger. 1987. http://www.jstor.org/stable/1913236 Co-integration and error correction: Representation, estimation, and testing . Econometrica, 55(2):251--276

    1987

  9. [17]

    Saeid Fallahpour, Hasan Hakimian, Khalil Taheri, and Ehsan Ramezanifar. 2016. https://doi.org/10.1007/s00500-016-2298-4 Pairs trading strategy optimization using the reinforcement learning method: a cointegration approach . Soft Computing, 20(12):5051--5066. Intraday US stocks...

    2016 doi

  10. [18]

    Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2024. https://arxiv.org/abs/2406.04093 Scaling and evaluating sparse autoencoders . Preprint, arXiv:2406.04093

    2024 arXiv

  11. [19]

    Davide Ghilardi, Federico Belotti, Marco Molinari, and Jaehyuk Lim. 2024. https://aclanthology.org/2024.blackboxnlp-1.32 Accelerating sparse autoencoder training via layer-wise transfer learning in large language models . In Proceedings of the 7th BlackboxNLP Workshop: Analyzi...

    2024

  12. [20]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, and Zhiyu Ma. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . Preprint, arXi...

    2024 arXiv

  13. [21]

    Paolo Guasoni and Eberhard Mayerhofer. 2018. https://doi.org/10.1111/mafi.12172 The limits of leverage . Mathematical Finance, 29(1):249–284

    2018 doi

  14. [22]

    Sungju Hong and Soosung Hwang. 2023. https://doi.org/10.1080/1351847X.2022.206 In search of pairs using firm fundamentals: is pairs trading profitable? The European Journal of Finance, 29(5):508--526

    2023 doi

  15. [23]

    Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. 2024. https://openreview.net/forum?id=F76bwRSLeK Sparse autoencoders find highly interpretable features in language models . In The Twelfth International Conference on Learning Representations

    2024

  16. [24]

    Subhash Kantamneni, Joshua Engels, Senthooran Rajamanoharan, Max Tegmark, and Neel Nanda. 2025. https://arxiv.org/abs/2502.16681 Are sparse autoencoders useful? a case study in sparse probing . Preprint, arXiv:2502.16681

    2025 arXiv

  17. [25]

    Sidhu, and Chuan Yu

    Dean Katselas, Baljit K. Sidhu, and Chuan Yu. 2017. https://doi.org/10.1111/acfi.12285 Know your industry: the implications of using static gics classifications in financial research . Accounting and Finance

    2017 doi

  18. [26]

    Zemel, Antonio Torralba, Raquel Urtasun, and Sanja Fidler

    Ryan Kiros, Yukun Zhu, Ruslan Salakhutdinov, Richard S. Zemel, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. 2015. https://aclanthology.org/P15-1009 Skip-thought vectors . In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the ...

    2015

  19. [27]

    Michael Lan, Philip Torr, Austin Meek, Ashkan Khakzar, David Krueger, and Fazl Barez. 2024. https://ar5iv.org/html/2410.06981v1 Sparse autoencoders reveal universal feature spaces across large language models . arXiv preprint, 2410.06981v1. Work done during the ERA-Krueger AI ...

    2024 arXiv

  20. [28]

    Andrew Lo. 2003. https://doi.org/10.2469/faj.v58.n4.2453 The statistics of sharpe ratios . Financial Analysts Journal, 58

    2003 doi

  21. [29]

    André Longon. 2024. https://arxiv.org/abs/2407.05340 Interpreting the residual stream of resnet18 . arXiv preprint arXiv:2407.05340

    2024 arXiv

  22. [30]

    Mico Loretan and William B. English. 2000. https://www.bis.org/publ/r_qt0006e.pdf Evaluating changes in correlations during periods of high market volatility . BIS Quarterly Review

    2000

  23. [31]

    Tim Loughran, Bill McDonald, and Hayong Yun. 2009. https://doi.org/10.1007/s10551-008-9910-1 A wolf in sheep?s clothing: The use of ethics-related terms in 10-k reports . Journal of Business Ethics, 89(S1):39--49

    2009 doi

  24. [32]

    Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. 2024. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. arXiv preprint arXiv:2403.19647

    2024 arXiv

  25. [33]

    Leland McInnes, John Healy, and James Melville. 2020. https://arxiv.org/abs/1802.03426 Umap: Uniform manifold approximation and projection for dimension reduction . Preprint, arXiv:1802.03426

    2020 arXiv

  26. [34]

    George J. Miao. 2014. https://doi.org/10.5539/ijef.v6n3p96 High frequency and dynamic pairs trading based on statistical arbitrage using a two-stage correlation and cointegration approach . International Journal of Economics and Finance, 6(3)

    2014 doi

  27. [35]

    MSCI. 2020. https://www.msci.com/documents/1296102/11185224/GICS+Methodology+2020.pdf Gics methodology 2020

    2020

  28. [36]

    Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. 2023. https://openreview.net/forum?id=9XFSbDPmdW Progress measures for grokking via mechanistic interpretability . In The Eleventh International Conference on Learning Representations

    2023

  29. [37]

    Kiho Park, Yo Joong Choe, and Victor Veitch. 2024. https://arxiv.org/abs/2311.03658 The linear representation hypothesis and the geometry of large language models . Preprint, arXiv:2311.03658

    2024 arXiv

  30. [38]

    Gonçalo Paulo, Alex Mallen, Caden Juang, and Nora Belrose. 2024. https://arxiv.org/abs/2410.13928 Automatically interpreting millions of features in large language models . Preprint, arXiv:2410.13928

    2024 arXiv

  31. [39]

    Ole Peters. 2011. https://doi.org/10.1080/14697688.2010.513338 Optimal leverage from non-ergodicity . Quantitative Finance, 11(11):1593–1602

    2011

  32. [40]

    Manda Raghava and Santosh Bharadwaj. 2014. Pairs trading using cointegration in pairs of stocks. Master of finance research project, Saint Mary’s University, Halifax, Nova Scotia, September. Submitted for MFIN 6692 under the direction of Dr. J. Colin Dodds and approved by Dr. ...

    2014

  33. [41]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . CoRR, abs/1908.10084

    2019 arXiv

  34. [42]

    Daniel Freeman, Theodore R

    Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, Nicholas L Turner, Callum McDougall, Monte MacDiarmid, C. Daniel Freeman, Theodore R. Sumers, Edward Rees, Joshua ...

    2024

  35. [43]

    Occupational Safety and Health Administration

    U.S. Occupational Safety and Health Administration . 2001. https://www.osha.gov/data/sic-manual Standard industrial classification (sic) manual . Accessed: 2024-11-08

    2001

  36. [44]

    Securities and Exchange Commission

    U.S. Securities and Exchange Commission . 2023. https://www.sec.gov/files/form10-k.pdf Form 10-k: Annual report pursuant to section 13 or 15(d) of the securities exchange act of 1934 . Accessed: 2024-12-02

    2023

  37. [45]

    Securities and Exchange Commission

    U.S. Securities and Exchange Commission . n.d. Cik lookup. https://www.sec.gov/search-filings/cik-lookup. Accessed: 2025-03-20

    2025

  38. [46]

    Zekeriya Uykan. 2021. https://doi.org/10.1109/ICDABI53623.2021.9655880 On the effect of data centering on spectral clustering with cosine similarity . In 2021 International Conference on Data Analytics for Business and Industry (ICDABI), pages 326--331

    2021

  39. [47]

    Dimitrios Vamvourellis, Michael Toth, Shubham Bhagat, Dhairya Desai, Dhruv Mehta, and Sara Pasquali. 2023. https://arxiv.org/abs/2308.08031 Company similarity using large language models . arXiv preprint arXiv:2308.08031. [Online; accessed 2-Dec-2024]

    2023 arXiv

  40. [48]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. https://arxiv.org/abs/1706.03762 Attention is all you need . In Proceedings of the 31st International Conference on Neural Information Processing...

    2017 arXiv

  41. [49]

    Winton. 2018. Systematic methods for classifying equities. Technical report, Winton Capital Management Limited (“WCM”)

    2018

  42. [50]

    Yahoo Finance . 2024. https://finance.yahoo.com Yahoo finance

    2024

  43. [51]

    Elnaz Zafarani-Moattar, Mohammad Reza Kangavari, and Amir Masoud Rahmani. 2021. https://arxiv.org/abs/2111.08658 A comparative study on transfer learning and distance metrics in semantic clustering over the covid-19 tweets . arXiv preprint, arXiv:2111.08658

    2021 arXiv

  44. [52]

    Fred Zhang and Neel Nanda. 2024. https://openreview.net/forum?id=Hf17y6u9BC Towards best practices of activation patching in language models: Metrics and methods . In Proceedings of the Twelfth International Conference on Learning Representations

    2024

  45. [53]

    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...

  46. [54]

    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 11, 2026 · model on record in the stance chip above.