Pith. sign in

REVIEW 4 major objections 6 minor 38 references

Estimating Covariance for Global Minimum Variance Portfolio: A Decision-Focused Learning Approach

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that covariance estimators for minimum-variance portfolios should be trained on the portfolio's own regret loss, and reports lower out-of-sample volatility than standard estimators.

desk verdict A potentially interesting DFL-for-GMVP result whose theory is internally broken and whose baselines need scrutiny. read the letter →

arxiv 2508.10776 v1 pith:IV3XHMLT submitted 2025-08-14 q-fin.PM cs.AI

Juchan Kim , Inwoo Tae , Yongjae Lee This is my paper · ORCID
classification q-fin.PMcs.AI MSC 91G10
keywords globalminimumvarianceportfoliocovarianceestimationdecision-focusedlearningprediction-focusedregretlossshrinkageDLinearvolatility
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

This paper argues that a covariance estimator for the global minimum-variance portfolio (GMVP) should be judged on the portfolio risk it produces, not on how accurately its matrix predicts the true covariance. It builds a DLinear network that predicts a covariance matrix, plugs the prediction into the closed-form GMVP solution, and trains the network end-to-end on a regret loss: the gap between the realized variance of the chosen weights and the oracle portfolio that used the true covariance. Across three equity universes over 2021-2023, the decision-focused estimator reports annualized out-of-sample volatility of roughly 0.115-0.138, versus 0.121-0.158 for shrinkage, historical, and prediction-focused baselines. The paper additionally claims a theoretical characterization: the decision-loss gradient's principal components lie in subspaces spanned by the GMVP weights and their risk-adjusted counterparts. If the results hold, covariance estimation should be calibrated with the downstream portfolio problem in the loop.

What carries the argument

The machinery is the analytic GMVP map with its exact gradient, combined with a backbone that predicts a positive-semidefinite covariance. The model (DLinear) outputs a lower-triangular matrix $L$, reconstructs $\hat{\Sigma}=LL^\top$, and applies truncated spectral reconstruction (keeping eigenpairs with $\lambda_i \ge \varepsilon\lambda_{\max}$) for inversion stability. The decision loss is the regret between the volatility of the chosen weights and the oracle weights; backpropagation flows through the closed-form solution, and the paper derives the principal-component structure of the resulting gradient in terms of $w$, $\hat{\Sigma}^{-1}w$, and their Kronecker products. These components a

What would settle it

Take any DFL-trained covariance from Section 3.2 and test the assumption $\ker((\hat{\Sigma}^{-1})^2) = \mathrm{span}(\{w, \hat{\Sigma}^{-1}w\})$: because the truncated spectral reconstruction keeps eigenvalues $\lambda_i \ge \varepsilon\lambda_{\max}$, $\hat{\Sigma}$ is positive definite, $(\hat{\Sigma}^{-1})^2$ is invertible, and its kernel is $\{0\}$, not the two-dimensional span. This directly contradicts Proposition 1's condition. For the empirical claim, run the same comparison on a test window outside March 2021-December 2023 (e.g., 2015-2019) and check whether DFL still has lower reali

Watch

Extended reading notes

Core claim

The central discovery is that decision-focused learning (DFL) applied to the unconstrained GMVP consistently beats prediction-focused estimation and classical shrinkage in realized volatility. The training objective is the regret loss $L(\theta) = w^*(\hat{\Sigma})^\top \Sigma_{\text{true}} w^*(\hat{\Sigma}) - w^*(\Sigma_{\text{true}})^\top \Sigma_{\text{true}} w^*(\Sigma_{\text{true}})$, and the gradient is decomposed as $dL/d\theta = (dL/dw)(\partial w/\partial \hat{\Sigma})(\partial \hat{\Sigma}/\partial\theta)$, with the weight map given by $w^*(\hat{\Sigma}) = \hat{\Sigma}^{-1}\mathbf{1}/(\mathbf{1}^\top \hat{\Sigma}^{-1}\mathbf{1})$. Propositions 1-3 state that singular vectors of the

Load-bearing premise

The theoretical result assumes certain null spaces of the estimated precision matrix are non-empty and exactly spanned by the portfolio weights and their risk-adjusted counterparts; the paper concedes this may fail in practice, and its own way of making the covariance invertible makes the null spaces empty. Separately, the empirical win assumes a single 2021-2023 test window, without transaction costs, represents how the method would behave generally.

Editorial extensions

If this is right

  • Covariance estimators for GMVP should be selected and trained on the regret loss; the paper reports that the decision-focused estimator's out-of-sample annualized volatility is about 0.115-0.138 versus 0.121-0.158 for classical baselines across three equity universes.
  • MSE-trained covariance predictors can collapse toward an equal-weight allocation because they underfit off-diagonal terms, so prediction accuracy is not a proxy for decision quality.
  • DFL-trained precision matrices show stable two-block structure, so the estimated weights are more consistent over time than those from the OAS baseline.
  • Longer rebalancing horizons raise realized volatility monotonically, which informs how often the portfolio should be rebalanced.

Reading between the lines

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

  • The stated kernel condition behind Propositions 1-3 is likely violated by the paper's own positive-definite covariance reconstruction, as the authors' Remark concedes; that would leave the empirical comparison as the actual support for the contribution (my reading, not the paper's).
  • Because the regret loss ignores transaction costs and turnover, the volatility gain may shrink net of costs; adding a turnover penalty to the decision loss is an immediate testable extension.
  • The 2021-2023 window features rising rates and elevated volatility; re-running the comparison on calm, low-volatility regimes would tell whether the decision-focused advantage is regime-specific.
  • Since only variance is optimized, the method says nothing about returns or Sharpe ratios; combining the decision loss with a return target is the natural next step and the authors list it as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes decision-focused learning (DFL) for estimating the covariance matrix used in the unconstrained global minimum-variance portfolio (GMVP). A DLinear network predicts a covariance matrix; GMVP weights are obtained from the closed-form solution; the training loss is the regret loss against the oracle covariance. The paper claims two contributions: (i) a theoretical characterization of the singular vectors / principal components of the decision-gradient (Propositions 1–3), and (ii) an empirical demonstration that DFL portfolios have lower out-of-sample volatility than classical shrinkage estimators and prediction-focused learning across three asset universes and several horizons (Table 1). The empirical section also analyzes the structure of the learned precision matrix, weight stability, and volatility attribution.

Significance. If the empirical result holds, the paper would make a useful practical point: covariance estimators for GMVP can be trained end-to-end on portfolio risk rather than on matrix error. The experimental reporting is mostly honest: five seeds, hyperparameters are listed, early stopping is used, and the comparison includes several standard baselines. The attribution analysis (block structure, low-volatility selection) is a valuable addition. However, the theoretical contribution is not established as stated: the closed-form gradient in Section 3.3 is incorrect, and the kernel assumptions in Proposition 1 are incompatible with the paper's own positive-definite estimator. The empirical comparison is also weakened by unequal information sets between DFL and classical baselines and by the short test window for long horizons. These issues affect the central claims, so the paper needs substantial revision.

major comments (4)
  1. [Sec. 3.3] The closed-form gradient is incorrect. Direct differentiation of w* = Σ̂^{-1}1/(1^TΣ̂^{-1}1) gives ∂w_i/∂Σ̂_{kl} = -(Σ̂^{-1})_{ik} w_l + D w_i w_k w_l, where D = 1^TΣ̂^{-1}1. Therefore the first term should be -Σ̂^{-1} ⊗ w^T, not -w^T ⊗ (Σ̂^{-1})^T. The paper's expression has the row index attached to the wrong factor: (w^T ⊗ (Σ̂^{-1})^T)_{i,(k,l)} = w_k (Σ̂^{-1})_{l,i}, which is not ∂w_i/∂Σ̂_{kl}. Since Propositions 1–3 analyze this J, their singular-vector conclusions are not derived for the actual problem. If the implementation uses the printed closed form, the reported updates are not gradients of the regret loss; if it uses autodiff through the solve, the theoretical derivation is disconnected. Please correct the formula and re-derive, or state clearly that training differentiates through the solve and remove the closed-form claim.
  2. [Sec. 3.2 / Prop. 1] The kernel assumptions in Proposition 1 cannot hold for the estimator the paper constructs. Section 3.2 retains eigenpairs with λ_i ≥ ε λ_max, making Σ̂ positive definite for ε>0; hence (Σ̂^{-1})^2 is invertible and its kernel is {0}, not span({w, Σ̂^{-1}w}). Similarly, the second kernel assumption is incompatible with a positive-definite Σ̂. The proof's key step ker(S1) = span({w, Σ̂^{-1}w}) is therefore false. The paper's own Remark concedes that the condition 'may fail in practice.' This makes Propositions 1–3 vacuous as stated. A non-vacuous analysis is needed, for example based on perturbation theory for invertible matrices, or the theoretical contribution should be removed.
  3. [Table 1 / Sec. 5.1] The comparison between DFL and the classical estimators is not on equal footing. The caption says 'All models use δin = 21 days for estimation,' but the DLinear models are trained on 2010–2018 daily data, whereas Historical, LW-D, LW-CC, and OAS use only the last 21 returns at each rebalance. A 21-day sample covariance is a weak baseline for GMVP; conventional practice often uses longer lookbacks (e.g., 60, 126, or 252 days). Thus the 'consistently superior' claim against classical estimators is not established by Table 1. A fairer comparison would give the classical estimators the same training information (e.g., an expanding window with the same initial training data).
  4. [Sec. 5.1 / Table 1] The test set spans March 2021 to December 2023, about 2.8 years. For δout = 126 and 252, the number of non-overlapping buy-and-hold rebalances is only about 5 and 2, respectively, so the annualized-volatility estimates in those columns have very large time-series sampling error. No standard errors or significance tests are reported for the volatility differences; the reported standard deviations are over five seeds and do not capture time-series sampling uncertainty. The claim of consistent superiority across horizons in Table 1 is therefore overstated, especially for long horizons.
minor comments (6)
  1. [Prop. 2 proof] Typo: 'equaiton' should be 'equation'.
  2. [Sec. 5.2] Typo: 'wherea' should be 'whereas'.
  3. [Algorithm 1] The subroutine DetectBlocks is called but never defined. Please specify what 'blocks' means and how they are detected.
  4. [Table 3] The 'average precision' metric is not defined. Please state the formula, how ranks are aggregated over the test period, and how ties are handled.
  5. [Sec. 3.2] The truncated spectral reconstruction uses a threshold ε, but the value of ε is not reported or included in the hyperparameter tables. Please specify how it is chosen and what values are used.
  6. [Sec. 4] The phrase 'Σ̂-invariant singular vectors' is used without a formal definition. Please define precisely what invariance means here.

Circularity Check

0 steps flagged · score 2.0 of 10

No circularity in the empirical or theoretical chain; one non-load-bearing self-citation, with correctness risks that are not circularity.

full rationale

The paper's headline claim—DFL yields lower out-of-sample GMVP volatility—is measured on a holdout period with buy-and-hold rebalancing against independent baselines (historical, LW, OAS, PFL); the regret-loss training target uses the training-period true covariance, so the test numbers are not fits relabeled as predictions. The only self-citation ([25], the authors' own MVO-DFL paper) is used to state the gap that DFL has not been applied to GMVP covariance estimation; it is not load-bearing and no uniqueness theorem is imported from it. The theoretical Section 4 is not circular: Propositions 1–3 derive singular-vector structure from explicit kernel/invertibility assumptions rather than from the target conclusion. However, those assumptions are contradicted by the paper's own construction: Section 3.2's truncated spectral reconstruction makes Σ̂ positive definite, so the nontrivial-kernel conditions in Proposition 1 cannot hold; the paper's Remark concedes this. In addition, the closed-form Jacobian in Section 3.3 appears to have transposed Kronecker factors. These are correctness/rigor concerns, not circular steps. The BBC block analysis is a post-hoc descriptive exercise, not a prediction defined by its own output. Overall, no step in the derivation chain reduces to its inputs by construction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The empirical claim rests mainly on domain assumptions (stationarity over a single test window, variance-only objective) and standard calculus; the theoretical claim rests on an ad hoc kernel condition that cannot be satisfied by the positive-definite covariance matrices the pipeline constructs. There are no invented physical entities; the claimed invariant singular vectors fail to exist under the paper's own assumptions.

free parameters (3)
  • Spectral truncation threshold epsilon = not reported
    Section 3.2: eigenvalues below epsilon times lambda_max are discarded before inversion; this controls conditioning of Sigma-hat and hence GMVP weights, but its value is never given.
  • Learning rate and batch size (per delta_in, delta_out pair) = Table 2, e.g., (1e-4, 32)
    Grid-searched hyperparameters selected on validation regret; the grid ranges and the selection criterion are not reported.
  • Architecture and stopping choices (hidden dim 128, kernel size rule, 50 epochs, patience 7) = h_dim=128, kernel=max{5,min{delta_in/3,50}}
    Chosen by hand; they affect whether DFL's advantage comes from the loss function or from model capacity relative to the collapsed PFL baseline.
assumptions (5)
  • standard math GMVP closed form w*(Sigma) = Sigma^-1 1 / (1^T Sigma^-1 1)
    Section 3.1; standard result from the Lagrangian of min w^T Sigma w subject to 1^T w = 1.
  • domain assumption Realized covariance of the next delta_out days is the ground-truth risk measure (Sigma_true)
    Section 3.3; the regret loss assumes the ex-post covariance is the correct evaluation of decision quality.
  • domain assumption Training-period structure (2010-2018) persists into the single test window (2021-2023)
    Sections 5.1-5.2; the entire 'consistently superior' claim rests on one held-out period with no significance testing.
  • ad hoc to paper Proposition 1 kernel conditions, e.g., ker((Sigma-hat^-1)^2) = span({w, Sigma-hat^-1 w})
    Section 4; unsatisfiable as stated because the pipeline's truncated spectral reconstruction makes Sigma-hat positive definite, so (Sigma-hat^-1)^2 is invertible with trivial kernel; the Remark concedes the condition may fail.
  • domain assumption Variance is the decision objective; returns, Sharpe ratio, and costs are out of scope
    Section 6; the paper explicitly excludes return and risk-adjusted metrics and transaction costs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Estimating Covariance for Global Minimum Variance Portfolio: A Decision-Focused Learning Approach." pith.science (2026). https://pith.science/paper/IV3XHMLT

@misc{pith2026250810776,
  author       = {Pith},
  title        = {Pith review of: Estimating Covariance for Global Minimum Variance Portfolio: A Decision-Focused Learning Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IV3XHMLT}},
  note         = {Machine review of arXiv:2508.10776}
}
read the original abstract

Portfolio optimization constitutes a cornerstone of risk management by quantifying the risk-return trade-off. Since it inherently depends on accurate parameter estimation under conditions of future uncertainty, the selection of appropriate input parameters is critical for effective portfolio construction. However, most conventional statistical estimators and machine learning algorithms determine these parameters by minimizing mean-squared error (MSE), a criterion that can yield suboptimal investment decisions. In this paper, we adopt decision-focused learning (DFL) - an approach that directly optimizes decision quality rather than prediction error such as MSE - to derive the global minimum-variance portfolio (GMVP). Specifically, we theoretically derive the gradient of decision loss using the analytic solution of GMVP and its properties regarding the principal components of itself. Through extensive empirical evaluation, we show that prediction-focused estimation methods may fail to produce optimal allocations in practice, whereas DFL-based methods consistently deliver superior decision performance. Furthermore, we provide a comprehensive analysis of DFL's mechanism in GMVP construction, focusing on its volatility reduction capability, decision-driving features, and estimation characteristics.

Figures

Figures reproduced from arXiv: 2508.10776 by the authors.

Figure 1
Figure 1. DFL framework for GMVP construction. Historical returns of N assets are processed by DLinear to predict covariance [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Rolling window approach for time series validation. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Annualized volatility of each model’s portfolio with [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Annualized volatility of DFL when 𝛿𝑖𝑛 and 𝛿𝑜𝑢𝑡 have different values in the Industry dataset. We set 𝛿𝑖𝑛, 𝛿𝑜𝑢𝑡 ∈ {5, 21, 63, 126, 252} [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 4
Figure 4. Figure 4: Average portfolio weights in the Industry portfolio [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Precision matrix derived by DFL and OAS for single test period in the Industry dataset (left), Average of the permuted [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Average true covariance matrix over the test period in the Industry dataset, with assets ordered by descending [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Portfolio weights by the models in whole test period [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Average portfolio weights in the S&P and Dow Jones dataset during the test period with [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Precision matrix derived by DFL and OAS for single test period (left), average precision matrices reordered by BBC [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Average true covariance matrix over the test period in the Industry dataset, with assets ordered by descending [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Portfolio weights by different models in the test period for S&P 500 and Dow Jones datasets. Stock indices are sorted [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages

  1. [1]

    Zico Kolter

    Akshay Agrawal, Brandon Amos, Shane Barratt, Stephen Boyd, Steven Diamond, and J. Zico Kolter. 2019. Differentiable Convex Optimization Layers. In Advances in Neural Information Processing Systems , Vol. 32

    work page 2019

  2. [2]

    Zico Kolter

    Brandon Amos and J. Zico Kolter. 2017. OptNet: Differentiable Optimization as a Layer in Neural Networks. In Proceedings of the 34th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 70) . PMLR, 136–145

    work page 2017

  3. [3]

    Best and Robert R

    Michael J. Best and Robert R. Grauer. 1991. On the sensitivity of mean-variance- efficient portfolios to changes in asset means: some analytical and computational results. The Review of Financial Studies 4, 2 (April 1991), 315–342

    work page 1991

  4. [4]

    Taras Bodnar, Stepan Mazur, and Yarema Okhrin. 2017. Bayesian estimation of the global minimum variance portfolio. European Journal of Operational Research 256, 1 (Jan. 2017), 292–307

    work page 2017

  5. [5]

    Taras Bodnar, Nestor Parolya, and Wolfgang Schmid. 2018. Estimation of the global minimum variance portfolio in high dimensions. European Journal of Operational Research 266, 1 (April 2018), 371–390

    work page 2018

  6. [6]

    Christian Bongiorno and Damien Challet. 2023. Non-linear shrinkage of the price return covariance matrix is far from optimal for portfolio optimization. Finance Research Letters 52 (2023), 103383

    work page 2023

  7. [7]

    Andrew Butler and Roy H. Kwon. 2023. Integrating Prediction in Mean–Variance Portfolio Optimization. Quantitative Finance 23, 3 (March 2023), 429–452

    work page 2023

  8. [8]

    Zhi Chai, Kai-Kit Wong, Kin-Fai Tong, Yu Chen, and Yangyang Zhang. 2022. Port Selection for Fluid Antenna Systems. IEEE Communications Letters 26, 5 (May 2022), 1180–1184

    work page 2022

Show all 38 references
  1. [9]

    Eldar, and Alfred O

    Yilun Chen, Ami Wiesel, Yonina C. Eldar, and Alfred O. Hero III. 2010. Shrink- age Algorithms for MMSE Covariance Estimation. IEEE Transactions on Signal Processing 58, 10 (Oct. 2010), 5016–5029

    2010

  2. [10]

    Chopra and William T

    Vijay K. Chopra and William T. Ziemba. 1993. The effect of errors in means, variances and covariances on optimal portfolio choice. Journal of Portfolio Management 19, 2 (1993), 6–11

    1993

  3. [11]

    Tsai-Hsuan Chung, Vahid Rostami, Hamsa Bastani, and Osbert Bastani

  4. [12]

    Clarke, Harindra De Silva, and Steven Thorley

    Roger G. Clarke, Harindra De Silva, and Steven Thorley. 2006. Minimum-variance portfolios in the US equity market. Journal of Portfolio Management 33, 1 (2006), 10–24

    2006

  5. [13]

    Clarke, Harindra De Silva, and Steven Thorley

    Roger G. Clarke, Harindra De Silva, and Steven Thorley. 2011. Minimum-variance portfolio composition. Journal of Portfolio Management 37, 2 (2011), 31–45

    2011

  6. [14]

    Giorgio Costa and Garud N. Iyengar. 2023. Distributionally Robust End-to-End Portfolio Construction. Quantitative Finance 23, 10 (Oct. 2023), 1465–1482

    2023

  7. [15]

    Victor DeMiguel, Lorenzo Garlappi, and Raman Uppal. 2009. Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? The Review of Financial Studies 22, 5 (May 2009), 1915–1953

    2009

  8. [16]

    Predict, then Optimize

    Adam N. Elmachtoub and Paul Grigas. 2022. Smart “Predict, then Optimize”. Management Science 68, 1 (Jan. 2022), 9–26

    2022

  9. [17]

    Aaron Ferber, Emily Griffin, Bistra Dilkina, Burcu Keskin, and Meredith Gore

  10. [18]

    Gabriel Frahm and Christoph Memmel. 2010. Dominating estimators for minimum-variance portfolios. Journal of Econometrics 159, 2 (Dec. 2010), 289– 302

    2010

  11. [19]

    Leonard R. Haff. 1980. Empirical Bayes Estimation of the Multivariate Normal Covariance Matrix. The Annals of Statistics 8, 3 (May 1980), 586–597

    1980

  12. [20]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Opti- mization. arXiv preprint arXiv:1412.6980 (Dec. 2014). arXiv:1412.6980 [cs.LG]

    2014 arXiv

  13. [21]

    Klein and Vijay S

    Robert W. Klein and Vijay S. Bawa. 1976. The effect of estimation risk on optimal portfolio choice. Journal of Financial Economics 3, 3 (June 1976), 215–231

    1976

  14. [22]

    Olivier Ledoit and Michael Wolf. 2003. Improved estimation of the covariance matrix of stock returns with an application to portfolio selection. Journal of Empirical Finance 10, 5 (Dec. 2003), 603–621

    2003

  15. [23]

    Olivier Ledoit and Michael Wolf. 2004. A Well -Conditioned Estimator for Large-Dimensional Covariance Matrices. Journal of Multivariate Analysis 88, 2 (Feb. 2004), 365–411

    2004

  16. [24]

    Olivier Ledoit and Michael Wolf. 2012. Nonlinear Shrinkage Estimation of Large-Dimensional Covariance Matrices. The Annals of Statistics 40, 2 (2012), 1024–1060

    2012

  17. [25]

    Junhyeong Lee, Inwoo Tae, and Yongjae Lee. 2024. Anatomy of Machines for Markowitz: Decision-Focused Learning for Mean-Variance Portfolio Optimiza- tion. arXiv:2409.09684 [q-fin.PM]

    2024

  18. [26]

    Yongjae Lee, Jang Ho Kim, Woo Chang Kim, and Frank J Fabozzi. 2024. An Overview of Machine Learning for Portfolio Optimization. Journal of Portfolio Management 51, 2 (2024)

    2024

  19. [27]

    Jayanta Mandi, Víctor Bucarey, Maxime Mulamba Ke Tchomba, and Tias Guns

  20. [28]

    Jayanta Mandi, James Kotary, Senne Berden, Maxime Mulamba, Víctor Bucarey, Tias Guns, and Ferdinando Fioretto. 2024. Decision-Focused Learning: Founda- tions, State of the Art, Benchmark and Future Opportunities. Journal of Artificial Intelligence Research 80 (Jan. 2024), 1623–1701

    2024

  21. [29]

    Markowitz

    Harry M. Markowitz. 1952. Portfolio selection. Journal of Finance 7, 1 (March 1952), 71–91

    1952

  22. [30]

    In Proceedings of the 39th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol

    Decision-Focused Learning: Through the Lens of Learning to Rank. In Proceedings of the 39th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 162) . PMLR, 14935–14947

  23. [31]

    Torsten Mörstedt, Bernhard Lutz, and Dirk Neumann. 2024. Cross validation based transfer learning for cross-sectional non-linear shrinkage: A data-driven approach in portfolio optimization. European Journal of Operational Research 318, 2 (May 2024), 670–685

    2024

  24. [32]

    Alessio Petrozziello, Luigi Troiano, Angela Serra, Ivan Jordanov, Giuseppe Storti, Roberto Tagliaferri, and Michele La Rocca. 2022. Deep learning for volatility forecasting in asset management. Soft Computing 26, 17 (Sept. 2022), 8553–8574

    2022

  25. [33]

    Robert C. Merton. 1980. On estimating the expected return on the market: An exploratory investigation. Journal of Financial Economics 8, 4 (Dec. 1980), 323–361

    1980

  26. [34]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are Transformers Effective for Time Series Forecasting?. InProceedings of the 37th AAAI Conference on Artificial Intelligence, Vol. 37. AAAI Press, 11121–11128. Estimating Covariance for Global Minimum Variance Portfolio: ...

    2023

  27. [36]

    Sanket Shah, Kai Wang, Bryan Wilder, Andrew Perrault, and Milind Tambe

  28. [37]

    In Advances in Neural Information Processing Systems, Vol

    Decision-Focused Learning without Decision-Making: Learning Locally Optimized Decision Losses. In Advances in Neural Information Processing Systems, Vol. 35

  29. [2022]

    arXiv:2211.08507 [cs.LG]

    Decision-Aware Learning for Optimizing Health Supply Chains. arXiv:2211.08507 [cs.LG]

    arXiv

  30. [2023]

    In Integration of Constraint Programming, Artificial Intelligence, and Operations Research (Lecture Notes in Computer Science)

    Predicting Wildlife Trafficking Routes with Differentiable Shortest Paths. In Integration of Constraint Programming, Artificial Intelligence, and Operations Research (Lecture Notes in Computer Science) . Springer Nature Switzerland, 460– 476

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.