Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Multi-objective Portfolio Optimization Via Gradient Descent

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

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

Pith's one-line read Gradient descent with penalty-encoded constraints solves multi-objective portfolio problems — including UCITS rules and tracking-error caps — that convex solvers cannot even express, and matches exact solvers where they exist.

desk verdict Useful, honest gradient-descent portfolio benchmark; constraint claims need out-of-sample evidence. read the letter →

arxiv 2507.16717 v1 pith:6YRMWVWV submitted 2025-07-22 cs.CE cs.LG

Christian Oliva , Pedro R. Ventura , Luis F. Lago-Fernández This is my paper
classification cs.CEcs.LG MSC 91G1090C29
keywords automaticdifferentiationmulti-objectiveportfoliooptimizationsparsemaxUCITSconstraintstrackingerrorCVaRgradientdescentconstraintpenalties
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 claims that an ordinary gradient-descent optimizer — the same iterative loop used to train neural networks — can serve as a general-purpose solver for portfolio construction. Weights are produced from free 'pre-weights' by a sparsemax mapping (with softmax as the dense alternative), which enforces non-negativity, sum-to-one, and exact zeros for unselected assets, while every objective and every constraint is folded into one differentiable loss as a weighted term. In the two cases where exact solvers exist, the gradient solution is nearly identical to them: maximizing the Sharpe ratio reproduces the convex solver's weights to the sixth decimal, and minimizing CVaR matches the dedicated solver's objective to the fifth decimal. In the four harder cases — UCITS concentration rules, tracking-error caps, minimum active weights, asset-count ranges, and sector masks — the method returns compliant portfolios while keeping risk-return metrics close to the unconstrained optimum, at the price of tuning penalty multipliers per problem. If the claim holds, practitioners get one extensible tool for problems that currently require bespoke reformulations or evolutionary heuristics.

What carries the argument

The load-bearing pieces are three. (1) The weight map: free pre-weights $z$ are passed through the sparsemax projection $\mathrm{sparsemax}(z) = [z - \tau(z)]_+$, which maps onto the probability simplex and returns exact zeros, so the portfolio always satisfies $w_i \ge 0$ and $\sum_i w_i = 1$, with sparsity obtained without ad-hoc truncation. (2) The multi-objective loss of Eq. (11): $L(z) = -\text{objective} + \sum_i \lambda_i C_i$, where the objective can be Sharpe ratio, CVaR, or volatility, and each constraint $C_i$ is written as a differentiable penalty — ReLU($\cdot$ - threshold) for caps such as the UCITS 10% rule or the tracking-error limit, and masked sums such as $C_{5\text{-}40\%} = \mathrm{ReLU}\big(\sum_i w_i \hat\sigma(w_i - 0.05) - 0.4\big)$ for the UCITS concentration rule. (3) The straight-through mask trick: boolean masks are computed as $\mathrm{round}(\mathrm{sigmoid}(x - \alpha))$, non-differentiable in forward evaluation, but backpropagation substitutes the smooth sigmoid derivative $\sigma(x)(1 - \sigma(x))$, keeping count- and membership-based constraints (minimum active weight, number of active assets in a range, asset subsets) trainable. Automatic differentiation then supplies every gradient, so a new objective or constraint is implemented as one additional term in the loss.

What would settle it

Run Cases 3 and 5 again on a different universe or time window (for example, S&P 500 returns from 2021-2022 or the STOXX 600 index) using the paper's reported $\lambda$ values and learning schedules: if the resulting portfolio breaches the UCITS 5-40% rule, the 0.4% tracking-error cap, or the 20-30 active-asset range, the claim that penalty terms enforce these constraints is falsified for the general setting. A sharper version scans $\lambda_{5\text{-}40\%}$ over $\{0.1, 1.0, 10.0, 100.0\}$ on the new data and checks whether the compliance window shifts by orders of magnitude or disappears.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that constraint handling in portfolio optimization needs no specialised machinery: one penalty term per constraint, built from ReLU-threshold violations and rounded-sigmoid masks whose gradients are proxied by the smooth sigmoid derivative, is enough to steer gradient descent to a feasible portfolio. The authors demonstrate this across six scenarios on S&P 500 components over 2019-12-31 to 2020-12-31. Max-Sharpe agrees with the CVXPY convex solution to the sixth decimal, with an average Euclidean distance of $3.6179\times10^{-5}$ across 100 random simulations; min-CVaR converges to a CVaR value within $8.7775\times10^{-10}$ (MSE over simulations) of the SKFOLIO solution; the UCITS constraints (no asset above 10%, assets above 5% summing to at most 40%) are satisfied with the sum at 39.98%; the 0.4% tracking-error cap is met to within $10^{-6}$; the fully constrained Case 5 keeps 23 active assets each above 1% while lifting the Sharpe ratio to 0.0516 against the index's 0.0168; and mask-level weight caps are met with deviations below 0.01%. The paper's stated purpose is not to beat existing optimizers but to provide a flexible, extensible benchmark: adding an objective or constraint means adding one differentiable term to the loss.

Load-bearing premise

The load-bearing premise is that soft penalty terms with hand-chosen multipliers $\lambda$ can genuinely enforce hard constraints: the paper's own grid search shows UCITS compliance only when $\lambda_{5\text{-}40\%}$ lies in a narrow window (roughly 1.0 to 10.0), and every case is validated in-sample on a single year of S&P 500 data, so if penalty tuning proves brittle across datasets or constraint sets, the central flexibility claim loses its footing.

Editorial extensions

If this is right

  • Any differentiable objective or constraint can be dropped into the same optimizer: UCITS rules, tracking-error caps, cardinality ranges, and sector masks become additional $\lambda$-weighted terms in the loss, so no new solver or mathematical reformulation is required per problem.
  • On the two problems with exact solution methods, the gradient solution is effectively the exact solution (weights to about $10^{-5}$, CVaR objective to about $10^{-10}$), so the method can serve as a drop-in alternative where a convex library is available and as the only option where it is not.
  • Non-convex, cardinality-constrained selection problems that are NP-hard for exact methods become ordinary gradient-descent runs; Case 5 satisfies all five constraint types simultaneously while improving the Sharpe ratio over the index.
  • The price of this flexibility is hyperparameter tuning: the Case 3 grid search satisfies UCITS rules only when $\lambda_{5\text{-}40\%}$ is between 1.0 and 10.0, with CVaR rising by 0.00178 relative to the unconstrained optimum.

Reading between the lines

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

  • A stress test not run in the paper: re-execute Cases 3-5 on a different index or out-of-sample window with the reported $\lambda$ values; if the compliance window for $\lambda_{5\text{-}40\%}$ shifts or disappears across datasets, the flexibility claim is much weaker than it appears.
  • The rounded-sigmoid mask uses a straight-through gradient — the forward pass is a hard 0/1 step while backprop uses the smooth sigmoid density — so gradient signals for count-based constraints are biased, and one would expect active-asset counts to hover or oscillate near range boundaries during training; this is checkable directly from the paper's training curves.
  • CVaR estimated from empirical returns is piecewise-linear in the weights, so the sparsemax-gradient result can be benchmarked against the exact linear-programming CVaR formulation on the same data; the residual gap would quantify what the sparsity prior and local minima cost, given the paper's own concession that Case 2 settles into a suboptimal local minimum.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a gradient-descent-based benchmark framework for single- and multi-objective portfolio optimization. Portfolio weights are obtained by mapping unconstrained pre-weights through softmax or sparsemax, and constraints such as UCITS limits, tracking-error caps, minimum active weights, active-asset counts, and subset weight budgets are incorporated as penalty terms in a weighted loss function. The method is evaluated in six scenarios on S&P 500 data covering 2019-12-31 to 2020-12-31. In the two unconstrained single-objective cases the results are compared with CVXPY (max Sharpe) and SKFOLIO (min CVaR); in the remaining constrained and multi-objective cases the paper reports portfolio-level metrics and constraint-compliance checks. The central claim is that the framework achieves competitive performance while offering greater flexibility for modeling multiple objectives and constraints.

Significance. If the central claim were fully established, the paper would provide a useful, extensible engineering tool for practitioners, with the appealing property that new objectives and constraints can be added by editing the loss function. The explicit open-source TensorFlow implementations in the appendices and the external comparisons in Cases 1 and 2 are genuine strengths, as is the 100-run simulation check in those cases. However, the significance is currently conditional: the penalty-based constraint mechanism is the load-bearing component of the flexibility claim, and the evidence for its reliability is limited to one in-sample year with hand-selected hyperparameters. The paper is best read as a demonstration of a framework rather than as a validated methodology, and the reported results do not yet support broad claims about robustness or real-world readiness.

major comments (4)
  1. [Section 4.3 / Figure 7] The grid search in Case 3 shows that the simplified UCITS constraint is satisfied only when lambda_5-40% is within the range roughly 1.0 to 10.0, with the portfolio infeasible for smaller values including 0.1. Because all six cases are evaluated on a single in-sample year and no rolling-window or out-of-sample test is reported, there is no evidence that these penalty hyperparameters transfer to other periods or asset universes. This is load-bearing for the central 'enhanced flexibility' claim: constraint satisfaction is achieved by tuned penalties, not by a projection or repair step, so the paper needs to demonstrate that the penalty mechanism is robust rather than dataset-specific.
  2. [Section 3.4.7 / Eq. (27) and Section 4.6] The mask constraint is defined with an absolute value, Cmask = sum_j |mmax,j - sum_i w_i M_ij|, which penalizes both under-allocation and over-allocation relative to the stated upper bound. The text describes an upper-bound constraint ('does not exceed a specified threshold'), but the implementation actually enforces equality of each mask's total weight to its target. Figure 14 confirms this by reporting that 'all masks achieve their targets.' Case 6 therefore tests a different constraint from the one claimed, and the paper does not demonstrate the advertised flexibility for upper-bound subset constraints.
  3. [Section 3.5.6 / Eq. (32)] The loss for Case 6 is written as L(z, lambda_mask) = sigma(R) - lambda_mask * Cmask. Since the optimizer minimizes L and Cmask is a penalty term described in Eq. (11) with a plus sign, the minus sign in Eq. (32) is inconsistent: as printed, the loss would encourage maximizing the constraint violation. Unless this is a typographical error, the formulation of Case 6 is internally incoherent and needs correction before the results can be interpreted.
  4. [Section 3.5.5 / Eq. (31) and Table 8] In Case 5, lambda_TE is set to 0.004, exactly equal to TEmax. Because the tracking-error penalty is C_TE = ReLU(TE - TEmax), the gradient contribution of this penalty is proportional to lambda_TE, so the enforcement pressure near the boundary is very weak (the penalty coefficient is only 0.004 times the excess). No sensitivity analysis is reported for any of the seven lambda hyperparameters in Eq. (31), and the single reported value TE = 0.003996 does not establish robustness. Since this case is the paper's main example of joint multi-objective, multi-constraint optimization, the fragility of the penalty mechanism here is a serious concern.
minor comments (4)
  1. [Section 4.2] The text states that CVaR 'in its standard form is known to lead to a generally non-convex optimization problem.' This is not correct: CVaR is convex in the portfolio weights (Rockafellar and Uryasev, 2000). The non-convexity in the GD setup likely comes from the sparsemax parameterization or added constraints, not from the CVaR objective itself. The statement should be corrected.
  2. [Section 3.5.5 / Eq. (31)] The symbol lambda is overloaded: it denotes the risk-aversion coefficient in Eq. (10), the vector of constraint multipliers in Eq. (31), and also appears as lambda_1 and lambda_2 for the two objectives. Please use distinct notation for the vector of constraint weights.
  3. [Appendix B / Table B.12] The function names mask_lower_than and mask_greater_than appear reversed relative to their behavior: mask_lower_than returns 1 when x > threshold, and mask_greater_than returns 1 when x < threshold. This is confusing and should be renamed or documented clearly.
  4. [Figure 4] The label 'Acummulated returns' contains a typo; it should be 'Accumulated returns.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper's claims are anchored by external solvers (CVXPY, SKFOLIO); constrained-case constraint checks are consistency checks on the optimizer's own penalized loss, not fitted predictions or self-citation chains.

full rationale

The paper is an engineering/benchmark contribution rather than a derivation chain, and I find no load-bearing step that reduces to its own inputs by construction. The objectives (Sharpe ratio, CVaR, volatility) and the constraints (tracking-error cap, UCITS limits, minimum weights, active-count range, mask budgets) are defined independently of the optimizer's outputs in Eqs. (15)-(27); no parameter is fitted to a subset of data and then reported as a prediction of a closely related quantity. Cases 1 and 2 are validated against external, independent solvers CVXPY and SKFOLIO, and the reported agreement in weights and metrics (Tables 1, 2 and 4) provides real external support for the gradient-descent machinery. In the constrained cases, the paper verifies that the penalty terms included in the loss (Eqs. 11, 29-32) have been driven to (near) zero, e.g., 'we confirm that the UCITS constraints have been satisfied' after optimizing Eq. (29). That is a consistency check on the optimizer's own objective rather than an independent prediction, but it is not circular in the derivation-chain sense: the penalties are soft, and the optimizer must actually find a feasible point, which the grid-search result in §4.3 shows is not automatic ('Compliance is achieved only when λ5-40% takes a value within the range between 1.0 and 10.0'). The lambda grid search is a hyperparameter sensitivity analysis, not a fitted-input-called-prediction step. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in through prior work by the same authors; the only external citation used for a methodological component is Martins and Astudillo (2016) for sparsemax, which is standard and independently published. The method's weakness — that penalty tuning is required and out-of-sample behavior is not demonstrated — is a correctness/generalization risk, not a circularity. Therefore the appropriate score is 0.

Assumptions & free parameters 12 free parameters · 6 assumptions · 0 invented entities

The framework's central claim rests on many user-chosen hyperparameters (the lambdas, learning rates, and epochs) and on domain assumptions about the reliability of penalty-based constraint enforcement and in-sample optimization. No new physical entities are introduced; the straight-through gradient is a technique, not an entity.

free parameters (12)
  • lambda_CVaR = 100 (Case 5)
    Weights the CVaR objective in the multi-objective loss; chosen by hand, affects trade-off.
  • lambda_Sharpe = 10 (Case 5)
    Weights the Sharpe objective; chosen by hand.
  • lambda_TE = 3 (Case 4), 0.004 (Case 5)
    Penalty weight for tracking error constraint; not fitted but user-selected.
  • lambda_10pct = 1.0 (Case 3), 10 (Case 5)
    Penalty weight for UCITS 10% constraint.
  • lambda_5_40 = 1.0 (Case 3), 10 (Case 5)
    Penalty weight for UCITS 5-40% constraint; Case 3 required grid search to find values that make portfolio compliant.
  • lambda_min = 10 (Case 5)
    Penalty weight for minimum active weight constraint.
  • lambda_range = 0.00001 (Case 5)
    Penalty weight for active asset count range.
  • lambda_mask = 0.1 (Case 6)
    Penalty weight for mask subset constraints.
  • learning_rate = 0.01 to 0.00005 per case
    Optimizer step size, hand-tuned per case.
  • epochs = 500 to 8000 per case
    Number of optimization steps, hand-tuned.
  • risk_free_rate = 0.0
    Assumed zero for Sharpe ratio; standard simplification.
  • CVaR_confidence_level = 0.05
    Confidence level for VaR/CVaR; standard but chosen by user.
assumptions (6)
  • domain assumption Gradient descent with automatic differentiation converges to a useful stationary point for the non-convex loss functions.
    The paper provides no convergence guarantees; in Case 2 it explicitly acknowledges convergence to suboptimal local minima.
  • domain assumption Penalty terms with sufficiently large multipliers can enforce hard constraints.
    Constraint satisfaction is not hard-coded; it emerges only if the lambda hyperparameters are high enough, as shown by the Case 3 grid search.
  • domain assumption In-sample optimization over a single historical window (2020) is indicative of practical portfolio performance.
    No out-of-sample or walk-forward validation is performed; all metrics are computed on the same data used to optimize.
  • domain assumption The custom gradient for the rounded sigmoid mask (straight-through estimator) provides reliable gradients.
    The paper uses the sigmoid derivative as a proxy gradient for the round function without theoretical or empirical justification beyond the reported runs.
  • standard math Sparsemax projects to the probability simplex correctly.
    Taken from Martins and Astudillo (2016); assumed correct.
  • standard math TensorFlow's automatic differentiation computes gradients of the composed loss correctly.
    Trusted software dependency; not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-objective Portfolio Optimization Via Gradient Descent." pith.science (2026). https://pith.science/paper/6YRMWVWV

@misc{pith2026250716717,
  author       = {Pith},
  title        = {Pith review of: Multi-objective Portfolio Optimization Via Gradient Descent},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6YRMWVWV}},
  note         = {Machine review of arXiv:2507.16717}
}
read the original abstract

Traditional approaches to portfolio optimization, often rooted in Modern Portfolio Theory and solved via quadratic programming or evolutionary algorithms, struggle with scalability or flexibility, especially in scenarios involving complex constraints, large datasets and/or multiple conflicting objectives. To address these challenges, we introduce a benchmark framework for multi-objective portfolio optimization (MPO) using gradient descent with automatic differentiation. Our method supports any optimization objective, such as minimizing risk measures (e.g., CVaR) or maximizing Sharpe ratio, along with realistic constraints, such as tracking error limits, UCITS regulations, or asset group restrictions. We have evaluated our framework across six experimental scenarios, from single-objective setups to complex multi-objective cases, and have compared its performance against standard solvers like CVXPY and SKFOLIO. Our results show that our method achieves competitive performance while offering enhanced flexibility for modeling multiple objectives and constraints. We aim to provide a practical and extensible tool for researchers and practitioners exploring advanced portfolio optimization problems in real-world conditions.

Figures

Figures reproduced from arXiv: 2507.16717 by the authors.

Figure 1
Figure 1. Round effect on the sigmoid activation function, following the equation 13. Note that the vertical orange line is exactly at α = 0.5, where the mask switches its value. This function, of course, presents a problem when applying gradient descent: the round func￾tion is not differentiable. However, this issue has an easy solution if, for gradient computation during training, we use the derivative of the original sigmo… view at source ↗
Figure 2
Figure 2. Weights comparison for selected assets (CVXPY vs GD) of S&P 500, whose weights are greater than 0. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Weights comparison for selected assets (SKFOLIO vs GD). [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Acummulated returns comparison (SKFOLIO vs GD) [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Minimizing CVaR with simplified UCITS constraints training metrics. The first figure ( [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Top figure: Weight assigned to every single asset of the universe along with 5% and 10% UCITS limits. Bottom [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: CVaR corresponding to each combination of regularization parameters [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Minimize CVaR with Tracking Error constraint training metrics [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Accumulated returns comparison between the resulting portfolio and the reference index S&P 500 when [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Maximize Sharpe ratio, minimize CVaR with Tracking Error, UCITS, minimum asset weight constraints, and [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Top figure: Selected assets from the whole S&P 500 universe. Bottom figure: Portfolio weights for the 23 [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Accumulated returns comparison between portfolio optimized using Gradient Descent and the reference index [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Minimize Risk with weight constraints in assets subsets training metrics [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: Asset weights selected by each Mask, showing the target objective for the optimizer and the value achieved. [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: Portfolio weights for 13 selected assets. [PITH_FULL_IMAGE:figures/full_fig_p024_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages

  1. [1]

    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 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  3. [3]

    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 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  4. [4]

    , author Agarwal, A

    author Abadi, M. , author Agarwal, A. , author Barham, P. , author Brevdo, E. , author Chen, Z. , author Citro, C. , author Corrado, G.S. , author Davis, A. , author Dean, J. , author Devin, M. , author Ghemawat, S. , author Goodfellow, I. , author Harp, A. , author Irving, G. , author Isard, M. , author Jia, Y. , author Jozefowicz, R. , author Kaiser, L....

    work page 2015

  5. [5]

    , author Verschueren, R

    author Agrawal, A. , author Verschueren, R. , author Diamond, S. , author Boyd, S. , year 2018 . title A rewriting system for convex optimization problems . journal Journal of Control and Decision volume 5 , pages 42--60

    work page 2018

  6. [6]

    , author Loraschi, A

    author Arnone, S. , author Loraschi, A. , author Tettamanzi, A. , year 1993 . title A genetic approach to portfolio selection . journal Neural Network World - International Journal on Neural and Mass-Parallel Computing and Information Systems volume 3 , pages 597--604

    work page 1993

  7. [7]

    , year 1990

    author Black, F. , year 1990 . title Asset allocation: Combining investor views with market equilibrium . journal Fixed Income Research/Goldman Sachs & Company

    work page 1990

  8. [8]

    , author Meade, N

    author Chang, T.J. , author Meade, N. , author Beasley, J.E. , author Sharaiha, Y.M. , year 2000 . title Heuristics for cardinality constrained portfolio optimisation . journal Comput. Oper. Res. volume 27 , pages 1271--1302

    work page 2000

Show all 46 references
  1. [9]

    , author Uryasev, S

    author Chekhlov, A. , author Uryasev, S. , author Zabarankin, M. , year 2004 . title Portfolio Optimization With Drawdown Constraints , in: editor Pardalos, P.M. , editor Migdalas, A. , editor Baourakis, G. (Eds.), booktitle Supply Chain And Finance . publisher World Scientifi...

    2004

  2. [10]

    , author Zhang, R.T

    author Chen, W. , author Zhang, R.T. , author Cai, Y.M. , author Xu, F. , year 2006 . title Particle swarm optimization for constrained portfolio selection problems , in: booktitle Proceedings of the 2006 International Conference on Machine Learning and Cybernetics , pp. pages...

    2006

  3. [11]

    , author Wang, T

    author Deng, L. , author Wang, T. , author Zhao, Y. , author Zheng, K. , year 2024 . title Million: A general multi-objective framework with controllable risk for portfolio management . https://arxiv.org/abs/2412.03038, http://arxiv.org/abs/2412.03038 arXiv:2412.03038

    2024 arXiv

  4. [12]

    , author Boyd, S

    author Diamond, S. , author Boyd, S. , year 2016 . title CVXPY : A P ython-embedded modeling language for convex optimization . journal Journal of Machine Learning Research volume 17 , pages 1--5

    2016

  5. [13]

    , author French, K.R

    author Fama, E.F. , author French, K.R. , year 1992 . title The cross-section of expected stock returns . journal The Journal of Finance volume 47 , pages 427--465 . http://www.jstor.org/stable/2329112

    1992

  6. [14]

    , author Gómez, S

    author Fernández, A. , author Gómez, S. , year 2007 . title Portfolio selection using neural networks . journal Computers & Operations Research volume 34 , pages 1177--1191

    2007

  7. [15]

    , author Fleming, P

    author Fonseca, C. , author Fleming, P. , year 1993 . title Genetic algorithms for multiobjective optimization: Formulation, discussion and generalization , in: booktitle Proceedings of the 5th International Conference on Genetic Algorithms , publisher Morgan Kaufmann Publishe...

    1993

  8. [16]

    , author Shoaf, J

    author Foster, J.A. , author Shoaf, J. , year 1996 . title A genetic algorithm solution to the efficient set problem: a technique for portfolio selection based on the markowitz model , in: booktitle 1996 Annual Meeting, Decision Sciences Institute

    1996

  9. [17]

    , author Ehrgott, M

    author Gandibleux, X. , author Ehrgott, M. , year 2005 . title 1984-2004 - 20 years of multiobjective metaheuristics. but what about the solution of combinatorial problems with multiple objectives? , in: booktitle International Conference on Evolutionary Multi-Criterion Optimization

    2005

  10. [18]

    , author Quintana, D

    author García, S. , author Quintana, D. , author Galván, I.M. , author Isasi, P. , year 2012 . title Time-stamped resampling for robust evolutionary portfolio optimization . journal Expert Systems with Applications volume 39 , pages 10722--10730

    2012

  11. [19]

    , author Nafpliotis, N

    author Horn, J.D. , author Nafpliotis, N. , author Goldberg, D.E. , year 1994 . title A niched pareto genetic algorithm for multiobjective optimization . journal Proceedings of the First IEEE Conference on Evolutionary Computation. IEEE World Congress on Computational Intellig...

    1994

  12. [20]

    , year 2023

    author Hugo Delatte, C.N. , year 2023 . title skfolio . https://github.com/skfolio/skfolio

    2023

  13. [21]

    , year 2001

    author Jorion, P. , year 2001 . title Value at risk . edition 2. ed. ed., publisher McGraw-Hill , address New York [u.a.]

    2001

  14. [22]

    , author Ertenlice, O

    author Kalayci, C.B. , author Ertenlice, O. , author Akbay, M.A. , year 2019 . title A comprehensive review of deterministic models and applications for mean-variance portfolio optimization . journal Expert Syst. Appl. volume 125 , pages 345–368

    2019

  15. [23]

    , author Corne, D.W

    author Knowles, J.D. , author Corne, D.W. , year 2000 . title Approximating the nondominated front using the pareto archived evolution strategy . journal Evol. Comput. volume 8 , pages 149–172

    2000

  16. [24]

    , year 1965

    author Lintner, J. , year 1965 . title The valuation of risk assets and the selection of risky investments in stock portfolios and capital budgets . journal The Review of Economics and Statistics volume 47 , pages 13--37

    1965

  17. [25]

    , year 1952

    author Markowitz, H. , year 1952 . title Portfolio selection . journal The Journal of Finance volume 7 , pages 77--91

    1952

  18. [26]

    , year 1959

    author Markowitz, H.M. , year 1959 . title Portfolio Selection: Efficient Diversification of Investments . publisher Yale University Press

    1959

  19. [27]

    , author Astudillo, R

    author Martins, A. , author Astudillo, R. , year 2016 . title From softmax to sparsemax: A sparse model of attention and multi-label classification , in: editor Balcan, M.F. , editor Weinberger, K.Q. (Eds.), booktitle Proceedings of The 33rd International Conference on Machine...

    2016

  20. [28]

    , author Liagkouras, K

    author Metaxiotis, K. , author Liagkouras, K. , year 2012 . title Multiobjective evolutionary algorithms for portfolio management: A comprehensive literature review . journal Expert Systems with Applications volume 39 , pages 11685--11698

    2012

  21. [29]

    , author Panda, G

    author Mishra, S. , author Panda, G. , author Majhi, B. , year 2016 . title Prediction based mean-variance model for constrained portfolio assets selection using multiobjective evolutionary algorithms . journal Swarm and Evolutionary Computation volume 28

    2016

  22. [30]

    , author Ruiz-Torrubiano, R

    author Moral-Escudero, R. , author Ruiz-Torrubiano, R. , author Su \'a rez, A. , year 2006 . title Selection of optimal investment portfolios with cardinality constraints , in: booktitle IEEE International Conference on Evolutionary Computation, CEC 2006, part of WCCI 2006, Va...

    2006 arXiv

  23. [31]

    , author Gross, S

    author Paszke, A. , author Gross, S. , author Massa, F. , author Lerer, A. , author Bradbury, J. , author Chanan, G. , author Killeen, T. , author Lin, Z. , author Gimelshein, N. , author Antiga, L. , author Desmaison, A. , author Kopf, A. , author Yang, E. , author DeVito, Z....

    2019

  24. [32]

    , author Uryasev, S

    author Rockafellar, R.T. , author Uryasev, S. , year 2000 . title Optimization of conditional value-at-risk . journal Journal of Risk volume 3 , pages 21--41

    2000

  25. [33]

    , year 2013

    author Roncalli, T. , year 2013 . title Introduction to risk parity and budgeting . journal SSRN Electronic Journal :10.2139/ssrn.2272973

    2013 doi

  26. [34]

    , year 2014

    author Roncalli, T. , year 2014 . title Introduction to risk parity and budgeting . https://arxiv.org/abs/1403.1889, http://arxiv.org/abs/1403.1889 arXiv:1403.1889

    2014 arXiv

  27. [35]

    , year 1985 a

    author Schaffer, J. , year 1985 a. title Multiple objective optimization with vector evaluated genetic algorithms. , pp. pages 93--100

    1985

  28. [36]

    , year 1985 b

    author Schaffer, J. , year 1985 b. title Some Experiments in Machine Learning Using Vector Evaluated Genetic Algorithms . Ph.D. thesis

    1985

  29. [37]

    , author Grefenstette, J.J

    author Schaffer, J.D. , author Grefenstette, J.J. , year 1985 . title Multi-objective learning via genetic algorithms , in: booktitle Proceedings of the 9th International Joint Conference on Artificial Intelligence - Volume 1 , publisher Morgan Kaufmann Publishers Inc. , addre...

    1985

  30. [38]

    , year 1964

    author Sharpe, W.F. , year 1964 . title Capital asset prices: A theory of market equilibrium under conditions of risk . journal The Journal of Finance volume 19 , pages 425--442

    1964

  31. [39]

    , author Deb, K

    author Srinivas, N. , author Deb, K. , year 1994 . title Muiltiobjective optimization using nondominated sorting in genetic algorithms . journal Evol. Comput. volume 2 , pages 221–248

    1994

  32. [40]

    , author Chan, L.C

    author Vedarajan, G. , author Chan, L.C. , author Goldberg, D. , year 1997 . title Investment portfolio optimization using genetic algorithms , in: editor Koza, J.R. (Ed.), booktitle Late Breaking Papers at the 1997 Genetic Programming Conference , publisher Stanford Bookstore...

    1997

  33. [41]

    , author Wang, S

    author Yu, L. , author Wang, S. , author Lai, K.K. , year 2008 . title Neural network-based mean–variance–skewness model for portfolio selection . journal Computers & Operations Research volume 35 , pages 34--46 . note Part Special Issue: Applications of OR in Finance

    2008

  34. [42]

    , author Deb, K

    author Zitzler, E. , author Deb, K. , author Thiele, L. , year 2000 . title Comparison of multiobjective evolutionary algorithms: Empirical results . journal Evol. Comput. volume 8 , pages 173–195

    2000

  35. [43]

    , author Laumanns, M

    author Zitzler, E. , author Laumanns, M. , author Thiele, L. , author Fonseca, C.M. , author da Fonseca, V.G. , year 2002 . title Why quality assessment of multiobjective optimizers is difficult , in: booktitle Proceedings of the 4th Annual Conference on Genetic and Evolutiona...

    2002

  36. [44]

    , author Thiele, L

    author Zitzler, E. , author Thiele, L. , year 1999 . title Multiobjective evolutionary algorithms: a comparative case study and the strength pareto approach . journal Trans. Evol. Comp volume 3 , pages 257–271

    1999

  37. [45]

    , " * write output.state after.block = add.period write

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION in...

  38. [46]

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