An Information Bottleneck Asset Pricing Model

Che Sun

REVIEW 4 major objections 5 minor 31 references

An Information Bottleneck Asset Pricing Model

T0 review · 4 major / 5 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 Adding an information bottleneck constraint to an autoencoder factor model filters noise and raises out-of-sample R² from 8.9% to 13.9% at 12 factors.

desk verdict The IB idea is sensible, but the paper's core objective is mathematically undefined — beta is deterministic in Eq. (2) and the KL term in Eq. (17) is infinite. read the letter →

arxiv 2507.23218 v1 pith:7MZHTWET submitted 2025-07-31 cs.CE cs.AI

Che Sun This is my paper · ORCID

classification cs.CEcs.AI

keywords informationbottleneckassetpricingautoencoderfactormodelsmutualoverfittingout-of-sampleR-squaredneuralnetworks

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 the main obstacle to deep factor models in asset pricing is not model capacity but noise: stock returns have a low signal-to-noise ratio, so neural networks fitted to 94 firm characteristics tend to memorize noise and lose out-of-sample accuracy as the number of latent factors grows. Its proposed fix is an information bottleneck regularizer that compresses the intermediate factor-loading representation while keeping information relevant to realized returns. Using a US equity panel split into training, validation, and test periods, it reports that an autoencoder with one hidden $\beta$ layer and the IB constraint (CA1+IB) reaches out-of-sample $R^2$ of 13.9% at $K=12$ and 13.5% at $K=24$, versus 8.9% and 7.6% for the same autoencoder without the constraint. If the results hold, the information bottleneck penalty would be a cheap, architecture-agnostic fix for overfitting in high-dimensional factor models.

What carries the argument

The load-bearing mechanism is the information bottleneck, implemented as a variational lower bound on the IB objective. The paper replaces the intractable mutual information terms $I(z_{t-1};\beta_{t-1})$ and $I(\beta_{t-1};r_t)$ by tractable variational quantities $\mathbb{E}[-\log q(r_t \mid g_\theta(z_{t-1}))]$ and $\lambda\,\mathrm{KL}[p(\beta_{t-1}\mid z_{t-1}), h(\beta_{t-1})]$, where $q$ is a variational decoder and $h$ is a prior chosen as the standard Gaussian. The first term rewards representations $\beta_{t-1}$ that still predict returns; the second penalizes representations that carry too much information about the input characteristics. Adding both to the autoencoder loss is what forces the network to forget noise while keeping pricing signal, and it is the piece that distinguishes this model from an autoencoder with only LASSO-style sparsity.

What would settle it

Block-bootstrap the monthly out-of-sample $R^2$ difference between CA1+IB and CA1 at $K=24$ over the 1987-2021 test period; if the 95% confidence interval contains zero, or if the point estimate is driven mainly by the dot-com bust and COVID episodes, the paper's noise-filtering mechanism is not established.

Watch

Extended reading notes

Core claim

The paper's central claim is that adding an information bottleneck constraint to a deep factor model prevents the overfitting that otherwise appears when the number of factors grows. The model keeps the autoencoder return equation $r_t = \beta_{t-1} f_t + u_t$, with loadings $\beta_{t-1}=g_\theta(z_{t-1})$ generated by a neural network from lagged characteristics, but augments the reconstruction loss with a variational information bottleneck term. The final objective is $\min_{\theta,b,W}\sum_t \|r_t-\beta_{t-1}f_t\|^2 + \mathbb{E}[-\log q(r_t \mid g_\theta(z_{t-1}))] + \lambda\,\mathrm{KL}[p(\beta_{t-1}\mid z_{t-1}), h(\beta_{t-1})]$, with $h$ taken to be a standard Gaussian. The reported consequence is that at $K=12$ and $K=24$ the constrained autoencoder beats the unconstrained one out of sample, while at small $K$ the constraint slightly hurts, because scarce factor capacity makes extra compression counterproductive. The authors read the mutual-information and cumulative-return plots as confirming that the bottleneck retains return-relevant information as network depth grows.

Load-bearing premise

The empirical gains come from one train-validation-test split with no resampling or error bars, so the gap at K=24 could be an artifact of the particular 1987-2021 test window rather than a general property of the information bottleneck.

Editorial extensions

If this is right

At $K=12$ and $K=24$, CA1+IB improves out-of-sample $R^2$ from 8.9% to 13.9% and from 7.6% to 13.5%, directly addressing the degradation plain CA1 shows as factor count rises.
IPCA+IB also improves over IPCA at $K=12$ and $K=24$, suggesting the benefit is not specific to the autoencoder architecture.
Tangency portfolio Sharpe ratios for CA1+IB are highest at $K=12$ (3.90) and $K=24$ (3.94), implying the extracted factors are more mean-variance efficient out of sample.
At small factor counts ($K=1,3,6$), IB does not help and can lower $R^2$, so the mechanism only helps when the model has enough capacity to overfit.
If correct, the IB regularizer can be added to existing neural factor models as a drop-in objective term, preserving their nonlinear flexibility while removing the noise component.

Reading between the lines

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

The reported gain rests on a single chronological split with no resampling; a natural check is to re-estimate over rolling windows or bootstrap the test months, since the K=24 margin may be period-specific rather than a general property of the IB penalty.
Because the variational bound replaces $I(\beta;z)$ with a KL penalty to a Gaussian prior, the amount of compression depends on the chosen prior; tighter or data-dependent priors could either strengthen or weaken the apparent noise filtering.
The same compression principle should transfer to other noisy financial prediction tasks such as volatility, credit, or portfolio-weight modeling where deep nets overfit, though the paper does not test those settings.

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.

Desk Editor's Note

read the letter

The paper marries an information bottleneck penalty to the GKX autoencoder and reports that it lifts out-of-sample R2 from 8.9% to 13.9% at K=12 and from 7.6% to 13.5% at K=24. The empirical pattern is the interesting part: the penalty helps exactly where the unrestricted model overfits, and the paper honestly notes that IB hurts at small K. That is worth a look. It is also honest in not claiming a new theory; it is an engineering extension.

The blocking problem is mathematical. Eq. (2) defines beta_{t-1} = g_theta(z_{t-1}), deterministic. Eq. (17) then adds lambda KL[p(beta|z) || N(0,I)]. With a deterministic mapping, p(beta|z) is a point mass and the KL divergence is infinite for every parameter vector. The variational derivation in Eqs. (12)-(16) treats p(beta|z) as a genuine conditional density and integrates over dbeta, but no stochastic encoder — no mu, sigma, reparameterization — is ever introduced. As written, the objective is undefined, and the 13.9% number cannot be traced to a well-posed loss. That is more severe than missing error bars. If the authors actually used a Gaussian encoder, they omitted a core ingredient; if they did not, the stated loss is vacuous.

There is also a double-counting issue: with Gaussian q(r|beta), the term -log q is proportional to the same squared loss already in the first term. That is fixable, but it means the objective's weighting is not what it appears.

The empirics have additional weaknesses: one chronological split, no resampling, no standard errors, and IPCA+IB solved with SGD against an analytical IPCA baseline is not an apples-to-apples comparison. No code or data is provided. The MI visualization in Figure 1 has no IB-free control, so the 'filters noise' claim is asserted rather than demonstrated.

Overall: the idea is plausible and the paper points to a real overfitting problem, but the central equation is not well-defined as written. I would send this to a referee only with a clear request to fix the stochastic encoder and provide code; without that, the main empirical claim is unverifiable. I would not cite it in its current form.

Referee Report

4 major / 5 minor

Summary. The paper proposes an information bottleneck (IB) asset pricing model that adds IB-style regularization to neural factor models, specifically the autoencoder asset pricing model of Gu, Kelly, and Xiu and an IPCA-type baseline. The stated objective, Eq. (17), combines the original reconstruction loss with a variational lower bound on the mutual information between returns and loadings and a KL penalty that encourages the loadings to stay close to a standard Gaussian prior. The empirical section reports monthly out-of-sample R-squared and tangency portfolio Sharpe ratios for US equities over 1987-2021, with the headline result that CA1+IB achieves 13.9% and 13.5% R-squared at K=12 and K=24 versus 8.9% and 7.6% for plain CA1 (Table 1). The paper interprets these gains as evidence that IB constraints filter noise and prevent overfitting in high-dimensional neural factor models.

Significance. If the proposed objective were well defined and the empirical gains robust, the paper would offer a practical regularizer for a known weakness of deep factor models, namely overfitting to the low signal-to-noise ratio of financial data. The application of information bottleneck ideas to asset pricing is natural, and the variational lower-bound derivation in Eqs. (12)-(15) follows a recognizable framework. However, the manuscript in its current form does not specify a stochastic encoder, so the KL term in the final objective is not well defined; the objective also appears to double-count the reconstruction loss. The empirical evidence rests on a single chronological split without error bars, and the IPCA+IB baseline changes the optimizer relative to the IPCA baseline. These issues are load-bearing: until they are resolved, the reported R-squared improvements cannot be attributed to the mechanism described in the paper. The paper does not ship code or a data appendix, which further limits reproducibility.

major comments (4)

[§3.3, Eqs. (2) and (17)] The KL term in Eq. (17) is not a well-defined objective for the model as specified in Eq. (2). Because β_{t-1} = gθ(z_{t-1}) is a deterministic function, p(β_{t-1}|z_{t-1}) is a degenerate distribution, and the KL divergence KL[p(β|z)||N(0,I)] is infinite for every parameter value. The variational derivation in Eqs. (12)-(16) requires a stochastic encoder p(β|z) (e.g., a Gaussian with a separate variance), but no such encoder, variance schedule, or reparameterization is described anywhere in Section 3. As a result, the out-of-sample results in Table 1 cannot be traced to the objective stated in Eq. (17); the paper must either introduce a stochastic encoder and optimize over its parameters or explain what the implemented KL term actually was.
[§3.3, Eqs. (15) and (17)] Eq. (17) double-counts the reconstruction error. If q(r|β) is a Gaussian centered on β f_t, then -log q(r_t | gθ(z_{t-1})) equals ||r_t - β_{t-1} f_t||^2 / (2σ²) plus a constant, so the second term in Eq. (17) is proportional to the first term. The variational bound in Eq. (15) uses ∫ p(β|z) log q(r|β) dβ as the evidence lower-bound term for I(β;r), but Eq. (17) appends it to the original reconstruction loss rather than replacing it. The stated objective is therefore not the lower bound derived above, and the relative weight of the likelihood and KL terms is not what the derivation implies. The paper should clarify the relation between q and the first term, for example by removing one of the two reconstruction terms.
[§4.3, Table 1 and §4.1] The empirical central claim rests on a single chronological split with no resampling, bootstrap, or significance tests. The training/validation/testing windows are 1957-1974, 1975-1986, and 1987-2021, and the test period contains well-known regime shifts (dot-com boom, financial crisis, COVID). Without error bars or a robustness check across alternative splits, the 5.9-point R² gain at K=24 (13.5% versus 7.6%) cannot be distinguished from a period-specific realization. In addition, the paper states that 'IPCA+IB' is solved by gradient descent rather than the analytical IPCA solution used for the IPCA baseline, so the IPCA versus IPCA+IB comparison confounds the IB constraint with a change of optimizer.
[§4.4.1, Figure 1] Figure 1 does not support the claim that the IB constraint filters noise. The figure shows mutual information I(β,r) as a function of network depth, but there is no IB-free control, no measure of I(z,β), and no description of how the mutual information estimates were computed. The observation that deeper layers have higher MI is consistent with overfitting to noise as much as with the proposed mechanism; it does not verify that the information bottleneck is active.

minor comments (5)

[§3.3, Eq. (15)] The first integral in Eq. (15) writes p(x) where p(z) is intended, and the text contains 'Morkov chain' and 'Rate telling Theory' typos.
[§2.2] The reference [SS05] is cited for tree-based methods, but Shorish and Spear (2005) is an agency-theoretic asset pricing paper, not a tree-based machine learning reference; a correct citation to a tree ensemble method is needed.
[§3.4] The paper cites [STIM18] for batch normalization, but Santurkar et al. (2018) is a study of batch norm properties; the original Ioffe and Szegedy (2015) reference is missing.
[§4.1] The data processing details are referred to 'supplementary materials,' but no supplementary materials accompany the manuscript; either include them or cite a public data appendix.
[§4.4.1, Figure 1] The caption and text should specify how mutual information is estimated from finite monthly samples and what estimator is used, since naive plug-in estimates of MI in high dimensions are known to be unstable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the information bottleneck term is an added regularizer and the reported R-squared values are genuine out-of-sample evaluations.

full rationale

The paper's central claim is that adding an information-bottleneck regularizer to an autoencoder asset-pricing model improves out-of-sample performance. The objective in Eq. (17) augments the baseline Eq. (4) with a variational mutual-information penalty, and the reported 13.9% and 13.5% R-squared values at K=12 and K=24 are computed on a test period (1987-2021) that is not used for estimation or hyperparameter tuning. The derivation chain from Eqs. (10)-(17) is a standard variational bound argument, and no fitted parameter is renamed as a prediction. The only notable issue is an internal mathematical inconsistency, not circularity: Eq. (2) defines beta as a deterministic function g_theta(z), while Eq. (17) includes KL[p(beta|z) || h(beta)] as though p(beta|z) were a proper stochastic encoder; this makes the KL term ill-defined as written. That is a correctness and reproducibility concern, but it does not make the empirical claim equivalent to the model's inputs. There are no self-citations by the author, and the baseline comparison to CA1 and IPCA gives the result independent content.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The model relies on standard factor-model assumptions, a variational approximation, and a hand-chosen Gaussian prior; the only free parameters are hyperparameters (lambda and architecture). No new physical or economic entities are introduced.

free parameters (2)

lambda (IB/KL penalty coefficient) = selected by validation search (value not reported)
The weight on the KL divergence in Eq. (17); the paper does not report the chosen value or a sensitivity analysis.
network architecture (e.g., 32-neuron hidden layer) = 32 neurons for CA1; 10 layers for Figure 1
The capacity of the loading network is chosen by hand and drives the reported R-squared/Sharpe improvements.

assumptions (4)

domain assumption Returns follow a linear latent factor model with dynamic loadings (Eq. 1).
Borrowed from GKX21; the entire model is built on this factor structure.
standard math The variational approximation q(r|beta) and h(beta) are valid and the Markov chain z -> beta -> r holds.
Invoked in Eqs. (12)-(15); standard IB/VAE approximation, but the step is asserted without derivation.
ad hoc to paper h(beta) is chosen as a standard Gaussian prior.
Stated in Section 3.3; any prior independent of z would do, and the choice affects the regularization strength.
domain assumption Mutual information in Figure 1 is estimated with an unstated estimator.
The paper reports MI for different layers but does not describe how it is computed (histogram, kernel, neural estimator); the plot's claim depends on this choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Information Bottleneck Asset Pricing Model." pith.science (2026). https://pith.science/paper/7MZHTWET

@misc{pith2026250723218,
  author       = {Pith},
  title        = {Pith review of: An Information Bottleneck Asset Pricing Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7MZHTWET}},
  note         = {Machine review of arXiv:2507.23218}
}

read the original abstract

Deep neural networks (DNNs) have garnered significant attention in financial asset pricing, due to their strong capacity for modeling complex nonlinear relationships within financial data. However, sophisticated models are prone to over-fitting to the noise information in financial data, resulting in inferior performance. To address this issue, we propose an information bottleneck asset pricing model that compresses data with low signal-to-noise ratios to eliminate redundant information and retain the critical information for asset pricing. Our model imposes constraints of mutual information during the nonlinear mapping process. Specifically, we progressively reduce the mutual information between the input data and the compressed representation while increasing the mutual information between the compressed representation and the output prediction. The design ensures that irrelevant information, which is essentially the noise in the data, is forgotten during the modeling of financial nonlinear relationships without affecting the final asset pricing. By leveraging the constraints of the Information bottleneck, our model not only harnesses the nonlinear modeling capabilities of deep networks to capture the intricate relationships within financial data but also ensures that noise information is filtered out during the information compression process.

Figures

Figures reproduced from arXiv: 2507.23218 by the authors.

**Figure 2.** The curves of cumulative return in the last 5 years of testing subset. The number of hidden [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages

[1]

Backpropagation and stochastic gradient descent method

Shun-ichi Amari. Backpropagation and stochastic gradient descent method. Neurocomputing , 5(4-5):185--196, 1993

work page 1993
[2]

Perspectives on the future of asset pricing

Markus Brunnermeier, Emmanuel Farhi, Ralph SJ Koijen, Arvind Krishnamurthy, Sydney C Ludvigson, Hanno Lustig, Stefan Nagel, and Monika Piazzesi. Perspectives on the future of asset pricing. The review of financial studies , 34(4):2126--2160, 2021

work page 2021
[3]

Asset pricing at the millennium

John Y Campbell. Asset pricing at the millennium. The Journal of Finance , 55(4):1515--1567, 2000

work page 2000
[4]

Information bottleneck for gaussian variables

Gal Chechik, Amir Globerson, Naftali Tishby, and Yair Weiss. Information bottleneck for gaussian variables. Advances in Neural Information Processing Systems , 16, 2003

work page 2003
[5]

Asset pricing: Revised edition

John H Cochrane. Asset pricing: Revised edition . Princeton university press, 2009

work page 2009
[6]

Deep learning in asset pricing

Luyang Chen, Markus Pelger, and Jason Zhu. Deep learning in asset pricing. Management Science , 70(2):714--750, 2024

work page 2024
[7]

Factorvae: A probabilistic dynamic factor model based on variational autoencoder for predicting cross-sectional stock returns

Yitong Duan, Lei Wang, Qizhong Zhang, and Jian Li. Factorvae: A probabilistic dynamic factor model based on variational autoencoder for predicting cross-sectional stock returns. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 36, pages 4468--4476, 2022

work page 2022
[8]

Common risk factors in the returns on stocks and bonds

Eugene F Fama and Kenneth R French. Common risk factors in the returns on stocks and bonds. Journal of financial economics , 33(1):3--56, 1993

work page 1993

Show all 31 references

[9]

Comparing cross-section and time-series factor models

Eugene F Fama and Kenneth R French. Comparing cross-section and time-series factor models. The Review of Financial Studies , 33(5):1891--1926, 2020

1926
[10]

Dissecting characteristics nonparametrically

Joachim Freyberger, Andreas Neuhierl, and Michael Weber. Dissecting characteristics nonparametrically. The Review of Financial Studies , 33(5):2326--2377, 2020

2020
[11]

Empirical asset pricing via machine learning

Shihao Gu, Bryan Kelly, and Dacheng Xiu. Empirical asset pricing via machine learning. The Review of Financial Studies , 33(5):2223--2273, 2020

2020
[12]

Autoencoder asset pricing models

Shihao Gu, Bryan Kelly, and Dacheng Xiu. Autoencoder asset pricing models. Journal of Econometrics , 222(1):429--450, 2021

2021
[13]

Thousands of alpha tests

Stefano Giglio, Yuan Liao, and Dacheng Xiu. Thousands of alpha tests. The Review of Financial Studies , 34(7):3456--3496, 2021

2021
[14]

Asset pricing with omitted factors

Stefano Giglio and Dacheng Xiu. Asset pricing with omitted factors. Journal of Political Economy , 129(7):1947--1990, 2021

1947
[15]

Assessing specification errors in stochastic discount factor models

Lars Peter Hansen and Ravi Jagannathan. Assessing specification errors in stochastic discount factor models. The Journal of Finance , 52(2):557--590, 1997

1997
[16]

Lucky factors

Campbell R Harvey and Yan Liu. Lucky factors. Journal of Financial Economics , 141(2):413--435, 2021

2021
[17]

A survey on information bottleneck

Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

2024
[18]

(re-) imag (in) ing price trends

Jingwen Jiang, Bryan Kelly, and Dacheng Xiu. (re-) imag (in) ing price trends. The Journal of Finance , 78(6):3193--3249, 2023

2023
[19]

Adam: A method for stochastic optimization

Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 , 2014

2014 arXiv
[20]

Shrinking the cross-section

Serhiy Kozak, Stefan Nagel, and Shrihari Santosh. Shrinking the cross-section. Journal of Financial Economics , 135(2):271--292, 2020

2020
[21]

Elastic information bottleneck

Yuyan Ni, Yanyan Lan, Ao Liu, and Zhiming Ma. Elastic information bottleneck. Mathematics , 10(18):3352, 2022

2022
[22]

On the information bottleneck theory of deep learning

Andrew M Saxe, Yamini Bansal, Joel Dapello, Madhu Advani, Artemy Kolchinsky, Brendan D Tracey, and David D Cox. On the information bottleneck theory of deep learning. Journal of Statistical Mechanics: Theory and Experiment , 2019(12):124020, 2019

2019
[23]

Capital asset prices: A theory of market equilibrium under conditions of risk

William F Sharpe. Capital asset prices: A theory of market equilibrium under conditions of risk. The journal of finance , 19(3):425--442, 1964

1964
[24]

Shaking the tree: an agency-theoretic model of asset pricing

Jamsheed Shorish and Stephen E Spear. Shaking the tree: an agency-theoretic model of asset pricing. Annals of Finance , 1:51--72, 2005

2005
[25]

The deterministic information bottleneck

DJ Strouse and David J Schwab. The deterministic information bottleneck. Neural computation , 29(6):1611--1630, 2017

2017
[26]

How does batch normalization help optimization? Advances in neural information processing systems , 31, 2018

Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? Advances in neural information processing systems , 31, 2018

2018
[27]

The information bottleneck method

Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057 , 2000

2000 arXiv
[28]

Deep learning and the information bottleneck principle

Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In 2015 ieee information theory workshop (itw) , pages 1--5. IEEE, 2015

2015
[29]

Pac-bayes information bottleneck

Zifeng Wang, Shao-Lun Huang, Ercan E Kuruoglu, Jimeng Sun, Xi Chen, and Yefeng Zheng. Pac-bayes information bottleneck. arXiv preprint arXiv:2109.14509 , 2021

2021 arXiv
[30]

Graph information bottleneck

Tailin Wu, Hongyu Ren, Pan Li, and Jure Leskovec. Graph information bottleneck. Advances in Neural Information Processing Systems , 33:20437--20448, 2020

2020
[31]

Multitask image clustering via deep information bottleneck

Xiaoqiang Yan, Yiqiao Mao, Mingyuan Li, Yangdong Ye, and Hui Yu. Multitask image clustering via deep information bottleneck. IEEE Transactions on Cybernetics , 54(3):1868--1881, 2023

2023

Pith tools

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

[1] [1]

Backpropagation and stochastic gradient descent method

Shun-ichi Amari. Backpropagation and stochastic gradient descent method. Neurocomputing , 5(4-5):185--196, 1993

work page 1993

[2] [2]

Perspectives on the future of asset pricing

Markus Brunnermeier, Emmanuel Farhi, Ralph SJ Koijen, Arvind Krishnamurthy, Sydney C Ludvigson, Hanno Lustig, Stefan Nagel, and Monika Piazzesi. Perspectives on the future of asset pricing. The review of financial studies , 34(4):2126--2160, 2021

work page 2021

[3] [3]

Asset pricing at the millennium

John Y Campbell. Asset pricing at the millennium. The Journal of Finance , 55(4):1515--1567, 2000

work page 2000

[4] [4]

Information bottleneck for gaussian variables

Gal Chechik, Amir Globerson, Naftali Tishby, and Yair Weiss. Information bottleneck for gaussian variables. Advances in Neural Information Processing Systems , 16, 2003

work page 2003

[5] [5]

Asset pricing: Revised edition

John H Cochrane. Asset pricing: Revised edition . Princeton university press, 2009

work page 2009

[6] [6]

Deep learning in asset pricing

Luyang Chen, Markus Pelger, and Jason Zhu. Deep learning in asset pricing. Management Science , 70(2):714--750, 2024

work page 2024

[7] [7]

Factorvae: A probabilistic dynamic factor model based on variational autoencoder for predicting cross-sectional stock returns

Yitong Duan, Lei Wang, Qizhong Zhang, and Jian Li. Factorvae: A probabilistic dynamic factor model based on variational autoencoder for predicting cross-sectional stock returns. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 36, pages 4468--4476, 2022

work page 2022

[8] [8]

Common risk factors in the returns on stocks and bonds

Eugene F Fama and Kenneth R French. Common risk factors in the returns on stocks and bonds. Journal of financial economics , 33(1):3--56, 1993

work page 1993