Pith. sign in

REVIEW 4 major objections 3 minor 85 references

Dynamic data generation and dynamic portfolio selection: an application of a score-based diffusion model

T0 review · 4 major / 3 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 An adaptive score-based diffusion sampler generates time-series distributions close to the unknown data model in the adapted Wasserstein metric, and the approximation error carries over to dynamic mean-variance portfolio selection.

desk verdict A solid stability theorem and a sensible adaptive sampling pipeline are paired with a central AW2 bound that does not actually vanish, so the paper needs substantive revision rather than desk rejection. read the letter →

arxiv 2507.09916 v4 pith:GJGXGWHK submitted 2025-07-14 q-fin.PM stat.ML

Ahmad Aghapour , Erhan Bayraktar , Fengyi Yuan This is my paper
classification q-fin.PMstat.ML MSC 91G1093E2060J60
keywords score-baseddiffusionmodelstime-seriesgenerationadaptivesamplingconditionalscorematchingadaptedWassersteindistancedynamicmean-varianceportfolioselectionpolicygradientgenerativeenvironment
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 extend score-based diffusion models, which normally learn a static data distribution, to dynamic time-series data where the order of information matters. Its central claim is that an adaptive sampling scheme trained by conditional score matching produces a generative model $\mathbb{Q}$ whose output distribution is close to the unknown original model $\mathbb{P}$ in the adapted Wasserstein metric $\mathcal{A}\mathcal{W}_2$, and that the same scheme supports conditional sampling given past observations. If true, this turns a diffusion model into a drop-in environment for dynamic decision problems: one can train a portfolio policy on many simulated price paths even when only a limited amount of real data is available. The paper also proves that the dynamic mean-variance portfolio value is Lipschitz in $\mathcal{A}\mathcal{W}_2$, so the generative approximation error translates into a performance error for portfolio policies, and it demonstrates the workflow on synthetic and real market data.

What carries the argument

The engine is the adaptive sampling scheme of Algorithm 1, built on the Ornstein-Uhlenbeck forward process $dX_\tau = -X_\tau\,d\tau + \sqrt{2}\,dB_\tau$. Instead of generating the whole path at once from the joint law on $\mathbb{R}^{dT}$, the algorithm generates sequentially: it first samples $y_1$ from noise through the reverse SDE using the score $s^1_\theta$, then, for each $t$, conditions on the already generated path $y_{1:t}$ and runs the reverse SDE with the conditional score approximation $s^{t+1}_\theta(\tau, y_{1:t}, \cdot)$. The error control rests on three pieces: the conditional denoising score-matching identity of Proposition 2.1, which turns conditional score error into an objective computable from joint samples; the dissipativity condition on the score network, $2x\cdot s_\theta \le -(1+\delta)|x|^2 + M_{\mathrm{disp}}$, which gives uniform-in-time tail estimates for the reverse process; and the bicausal-coupling definition of the adapted Wasserstein metric, under which the error accumulates additively over time. The implemented version replaces the conditioning path $x_{1:t}$ with a fixed-dimensional RNN feature $h_t$, so the score network sees $(\tau, h_t, x)$ instead of $(\tau, x_{1:t}, x)$.

What would settle it

A direct check of the algebra: with $M_{\mathrm{disp}} \sim \log(1/\varepsilon_{\mathrm{score}})$ as in Proposition 2.2, the term $e^{-cT+cM_{\mathrm{disp}}}$ in $\alpha(T)$ behaves like a positive power of $1/\varepsilon_{\mathrm{score}}$ times $e^{-cT}$, so for any fixed $T$ it diverges as $\varepsilon_{\mathrm{score}}\to 0$; the bound can therefore only go to zero in the joint limit $T\to\infty$, $\varepsilon_{\mathrm{score}}\to 0$. An empirical version of the same check: train the score network at several decreasing loss targets, record the achieved value of $M_{\mathrm{disp}}$ from the trained weights, and test whether the measured $\mathcal{A}\mathcal{W}_2$ distance between real and generated paths continues to decrease; if it plateaus while $M_{\mathrm{disp}}$ grows, the convergence statement of Remark 2 is not realised by the implemented model.

Watch

Extended reading notes

Core claim

At the core is Theorem 2.3, which bounds the adapted Wasserstein distance between the true law $\mathbb{P}$ and the law $\mathbb{Q}_T$ produced by the adaptive sampling algorithm. For time horizon $T$ and score-matching error $\varepsilon_{\mathrm{score}}$, the bound is $$\mathcal{A}\mathcal{W}$_2^{2}$(\mathbb{P},\mathbb{Q}_T) \le C\bigl($T^{{5T/2}}$\varepsilon_{\mathrm{score}}^{1/$2^{{T-1}}$} + $T^{{5(T-1)/2}}$\$\alpha$(T)^{1/$2^{{T-1}}$}\bigr), \quad \$\alpha$(T)=$T^{2}$$e^{{-T}}$+$e^{{-cT+cM_{\mathrm{disp}}$}},$$ and Remark 2 argues that this can be made arbitrarily small by sending the diffusion time $T$ to infinity while taking $\varepsilon_{\mathrm{score}}$ small at a coupled rate. The paper further derives Corollary 2.4, an expected conditional Wasserstein bound, which is the statement that makes one-step-ahead sampling from observed histories possible. In Section 3 it proves that the mean-variance value functions under $\mathbb{P}$ and $\mathbb{Q}$ differ by at most $C\,\mathcal{A}\mathcal{W}_2(\mathbb{P},\mathbb{Q})$, and that a policy that is $\varepsilon$-optimal under $\mathbb{Q}$ is within $\varepsilon + C\,\mathcal{A}\mathcal{W}_2(\mathbb{P},\mathbb{Q})$ of optimal under $\mathbb{P}$.

Load-bearing premise

The whole guarantee rests on the assumption that one network can be both a good score approximator and a stable, mean-reverting function whose instability constant stays small as the approximation error shrinks; if accuracy and stability cannot be achieved together, the stated error bound stops shrinking to zero.

Editorial extensions

If this is right

  • If the bound of Theorem 2.3 is correct, a generative model with small $\mathcal{A}\mathcal{W}_2(\mathbb{P},\mathbb{Q})$ can serve as a training environment for dynamic policies: near-optimality transfers from $\mathbb{Q}$ to $\mathbb{P}$ up to an additive error of order $C\,\mathcal{A}\mathcal{W}_2(\mathbb{P},\mathbb{Q})$.
  • The adaptive sampler gives conditional path generators: given an observed history $x_{1:t}$, the same algorithm samples the next step from the conditional law, which is exactly what a rebalancing policy needs for scenario generation.
  • The one-step case $T=1$ recovers a Wasserstein bound for ordinary static score-based generation without log-concavity assumptions on the data distribution, at the cost of a noise term that worsens as $\varepsilon_{\mathrm{score}}$ shrinks.
  • The duality between dynamic mean-variance and quadratic hedging (equation 3.3) is the bridge that converts $\mathcal{A}\mathcal{W}_2$ closeness into value-function closeness and provides the training signal for the policy-gradient algorithm.
  • On real monthly industry portfolio data, the generative-model TD3 strategy beats the S&P 500, equal weight, and historical Markowitz on Sharpe and Sortino ratios, and remains stable across the risk-aversion levels tested, while the generative Markowitz variant is competitive at high risk aversion but degrades at low risk aversion.

Reading between the lines

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

  • I infer that the practical working guarantee is Corollary 3.10 rather than the full convergence theorem: even a moderately accurate $\mathbb{Q}$ gives a usable scenario generator, because the policy-gradient agent trains on $\mathbb{Q}$ and its quality under $\mathbb{P}$ is controlled by the added term $C\,\mathcal{A}\mathcal{W}_2(\mathbb{P},\mathbb{Q})$.
  • I infer that the convergence statement in Remark 2 is a joint limit in $T$ and $\varepsilon_{\mathrm{score}}$; for fixed $T$, the term involving $\alpha(T)$ does not vanish as the score error alone goes to zero, so 'arbitrarily small' should be read as a statement about the coupled double limit.
  • I infer that the RNN-encoder version used in the experiments is not literally covered by Theorem 2.3, which assumes the full conditioning path $x_{1:t}$ is available; measuring $\mathcal{A}\mathcal{W}_2$ between the full-path sampler and the RNN-based sampler would quantify the gap introduced by feature compression.
  • A testable extension: use the adaptive conditional sampler as a scenario generator for other dynamic objectives, such as utility maximization or CVaR-constrained control, and check whether the same stability argument, which currently relies on the quadratic-hedging duality, can be replaced by a direct Lipschitz bound on the relevant value function.
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 / 3 minor

Summary. The paper develops an adaptive score-based diffusion framework for generating time-series paths, with the aim of using the learned model as a surrogate environment for dynamic mean-variance portfolio selection. The main theoretical result (Theorem 2.3) claims an adapted Wasserstein bound between the true model P and the generated model Q, with a right-hand side that, according to Remark 2(1), can be made arbitrarily small by increasing the diffusion time T and decreasing the score-matching error ε_score. The paper then proves a stability result for the mean-variance value function under the adapted Wasserstein metric (Theorem 3.8, Corollary 3.9), and combines these results with a policy-gradient algorithm (Algorithm 4). The empirical sections report synthetic and real-data experiments, including a generative-model-based TD3 strategy that outperforms several benchmarks. The central theoretical claim is the convergence guarantee of Theorem 2.3; the paper's applicability to portfolio selection rests on that guarantee being valid.

Significance. If Theorem 2.3 were correct, the paper would make a useful contribution: it extends score-based diffusion to conditional time-series generation with adapted Wasserstein guarantees, proves a stability result for dynamic mean-variance problems, and gives an end-to-end generative reinforcement-learning pipeline with publicly available code. The conditional denoising score-matching equivalence (Proposition 2.1) and the dynamic-programming-based stability analysis in Section 3 are careful and self-contained, and the manuscript is unusually transparent in stating its assumptions. However, the central convergence guarantee is not delivered: the bound in Theorem 2.3 cannot be made arbitrarily small because of the dissipativity constant M_disp in α(T). Since the downstream portfolio guarantees in Corollaries 3.9 and 3.10 inherit that bound, the theoretical support for the proposed method collapses. The paper therefore cannot be accepted in its present form, and the main theorem is not a locally fixable presentation issue but a structural gap in the argument.

major comments (4)
  1. [Section 2, Theorem 2.3 and Remark 2(1)] The claimed convergence guarantee is not valid. Proposition 2.2 states that M_disp is of order log(1/ε_score), so in (2.12), α(T) = T^2 e^{-T} + e^{-cT + cM_disp} contains e^{-cT} ε_score^{-cC} for some C > 0. After taking the 2^{T-1}-th root in (2.14), the second term behaves as ε_score^{-cC/2^{T-1}} as ε_score → 0, so for every fixed T the right-hand side of (2.14) has a positive minimum and cannot be driven to zero. The alternative proposed in Remark 2(1), namely choosing ε_score = T^{-5·2^{T-3}+1} e^{-cT/2}, makes the first term behave like T^{5T/2 - 5/4 + 1/2^{T-1}} e^{-cT/2^T}, whose logarithm tends to +∞ as T → ∞. This is not a minor typo: the same e^{-cR+cM_disp} term appears in the proof of Theorem 2.3 in Appendix B immediately after (B.6), so the failure is structural.
  2. [Section 2, Theorem 2.3 and Remark 2(1)] The theorem and its Remark 2(1) use T simultaneously as the length of the time series and as the diffusion horizon. Letting T → ∞ therefore changes the ambient space from R^{dT} to R^{d(T+1)}; it is not a fixed-problem limit. In the portfolio application T = 12 is fixed, and for fixed T the bound cannot be optimized to smallness because of the ε_score^{-cC/2^{T-1}} factor. Consequently Corollary 3.9 and Corollary 3.10, which convert AW_2(P,Q) into value-function suboptimality, cannot be invoked to justify near-optimality of the generative RL pipeline.
  3. [Section 4.1, RNN encoder assumption] The theoretical conditional-score bound (2.3) is stated for the full path x_{1:t} as conditioning input, but the implemented Algorithm 3 and training objective (4.1) replace x_{1:t} by the RNN encoding h_t = R_θ(s_t, h_{t-1}). The paper assumes that s_{1:t} is recoverable from h_t (Section 4.1), i.e., that the encoder is injective on the data support. This is not established and is generally false for a trained recurrent network. Without injectivity, the conditional distributions of the generated process are not controlled by (2.3), so the application of Theorem 2.3 to the implemented model is unjustified.
  4. [Appendix A vs. Assumption 1] Proposition 2.2 is proved only for τ ∈ [τ_0, T] for some τ_0 > 0, whereas Assumption 1 and Theorem 2.3 require uniform score error on τ ∈ (0,T]. The text states that a standard early-stopping technique gives the same convergence guarantee, but no argument or estimate for the early-stopped process is provided. This creates an additional gap between the network construction and the main theorem, beyond the M_disp dependence discussed above.
minor comments (3)
  1. [Section 1.1, Eq. (1.1)] The definition of the variance-preserving SDE is incomplete: β(τ) is written as "β(τ) = β_min + (β_max − β_min), τ is a prescribed drift schedule," which is missing the normalization by the terminal time and is not a valid function definition.
  2. [Section 4.3.2] The test-period description states that the test dataset runs "from April 2009 to March 2008," which is an impossible interval; presumably the intended end is March 2010 or a similar later date. In the same subsection, "Dateset" and "splitted" are typos.
  3. [Section 1.4] Using the same symbol T for the time-series length and the diffusion time horizon is confusing and actively harmful in the statement of Theorem 2.3, where "T → ∞" is ambiguous. Distinct notation (e.g., N or H for the series length) would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Theorem 2.3's AW2 bound is derived from explicit score-error and dissipativity assumptions, and the only overlapping-author citation is not load-bearing.

full rationale

The theoretical derivation is self-contained rather than circular. Theorem 2.3 bounds AW2(P,QT) in terms of a score-matching error epsilon_score and a dissipativity constant M_disp that appear as explicit assumptions (Assumptions 1 and 3); the proof proceeds through standard total-variation bounds for the reverse SDE, a Wasserstein-to-TV comparison (Villani Theorem 6.15), and uniform moment/tail estimates (Lemma B.1, Corollary B.2). Neither epsilon_score nor M_disp is fitted to the target AW2 value, and the bound is not defined in terms of the quantity it claims to bound. Proposition 2.2 and Lemma A.2 construct a ReLU network satisfying the dissipativity condition with M_disp ~ log(1/epsilon_score) using an external universal-approximation lemma from Chen et al. (2022); this is an independent approximation argument, not a restatement of the AW2 result. The portfolio stability results (Theorem 3.8, Corollary 3.9) are proven from the quadratic-hedging DPP and the duality relation (3.3), without invoking the generative model's own output. The only self-citation by an overlapping author is Bayraktar and Han (2025+) in the definition of bicausal couplings; the load-bearing structural fact used in the proof is cited to Backhoff et al. (2017), so this self-citation is not load-bearing. The skeptical concern that the alpha(T) term in (2.14) may not vanish as epsilon_score -> 0 is a mathematical correctness issue about the stated convergence guarantee, not an instance of the derivation reducing to its inputs by construction. Accordingly, no circular step is identified.

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

The central claims rest on several technical assumptions about the data and the network. Most are explicitly stated, but the RNN-encoder approximation (Section 4.1) is an unquantified assumption that bridges the theory and the implementation. No free parameters are fitted to data to make the theory work; the listed parameters are algorithmic hyperparameters.

free parameters (6)
  • beta schedule beta_min, beta_max = 0.01, 10.0
    Hyperparameters of the variance-preserving SDE used in real-data experiments (Section 4.3.2); not fitted to data, but chosen by hand and affect generation quality.
  • number of diffusion timesteps N = 1000
    Discretization steps for the VP-SDE in real-data experiments, used in Algorithm 3 (Section 4.3.2).
  • risk aversion gamma = 3 (also 0.5, 5)
    User-set objective parameter in the mean-variance criterion; performance of the strategies varies with it.
  • predictor/corrector step counts N_pre, N_cor = not reported
    Sampling hyperparameters in Algorithm 3 that control discretization error; no values given in the text.
  • scenario pool / replay buffer size L = not reported
    Capacity of the scenario pool and replay buffer in Algorithm 4; not specified in the text.
  • exponential prior rate for multiplier c = Exp(10)
    Distribution used to sample the dual multiplier c during training in Algorithm 4.
assumptions (7)
  • domain assumption Score-matching error bounds (Assumption 1): the trained network sθ satisfies (2.2)-(2.3) with error ε_score.
    The theory assumes the network has learned the conditional scores to within ε_score; in practice this is only approximately achieved by optimizing (4.1).
  • domain assumption Data distribution assumptions (Assumption 2): conditional score functions are Lipschitz and the data has exponential moments.
    Required for the total-variation and tail estimates in Appendix B; realistic for bounded return data but unverified on the real dataset.
  • ad hoc to paper Network dissipativity and Lipschitz (Assumption 3).
    The paper proves in Proposition 2.2 that such a network exists, but the construction in Appendix A requires support truncation and may conflict with attaining small ε_score while keeping M_disp small.
  • domain assumption Market second-moment bound (Assumption 4).
    EP[|S_{t+1}|^2|F_t] uniformly bounded; reasonable for price processes with bounded moments.
  • domain assumption Non-degeneracy condition (Assumption 5, eq (3.4)).
    Guarantees closedness of attainable wealth space and existence of hedging optimizers; for constrained portfolios this is an extra assumption.
  • ad hoc to paper RNN encoder injectivity (Section 4.1).
    The implementation assumes s_{1:t} is recoverable from h_t=R(s_t,h_{t-1}), so that the score network conditioned on h_t is equivalent to conditioning on the full path; no evidence is given that a trained LSTM satisfies this.
  • standard math Duality relation (3.3) from Labbe and Heunis (2007).
    Borrowed from the cited literature; used to convert the mean-variance problem into a quadratic hedging problem.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic data generation and dynamic portfolio selection: an application of a score-based diffusion model." pith.science (2026). https://pith.science/paper/GJGXGWHK

@misc{pith2026250709916,
  author       = {Pith},
  title        = {Pith review of: Dynamic data generation and dynamic portfolio selection: an application of a score-based diffusion model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GJGXGWHK}},
  note         = {Machine review of arXiv:2507.09916}
}
abstract

We study dynamic data generation and its application to model-free dynamic portfolio selection. Existing score-based diffusion models are typically designed to learn a static data distribution, whereas dynamic decision problems require generated trajectories that preserve the sequential information structure of the underlying process and support conditional sampling. To address this gap, we develop an adaptive score-based diffusion framework for dynamic data. Given samples from an unknown data-generating model $\mathbb P$, the framework learns a generative model $\mathbb Q$ through conditional score matching and generates trajectories sequentially by updating the conditioning information over time. We establish quantitative error bounds between $\mathbb P$ and $\mathbb Q$ under the adapted Wasserstein metric $\mathcal A\mathcal W_2$, which is tailored to nonanticipative dynamic problems, and show that the same adaptive sampling scheme provides conditional path generators. We then apply this dynamic data generation framework to dynamic mean-variance portfolio selection with limited historical price data. We prove stability of the dynamic mean-variance problem with respect to $\mathcal A\mathcal W_2$, thereby translating the generative approximation error into performance control for portfolio policies. Building on these results, we implement a policy-gradient algorithm in the learned generative environment, where adaptively sampled paths serve as training scenarios. A synthetic ARMA experiment shows that the proposed adaptive sampling scheme generates distributions close to the true data-generating process. On real market data, the proposed approach outperforms several benchmarks, including the Markowitz portfolio, the equal-weight portfolio, and the S\&P 500.

Figures

Figures reproduced from arXiv: 2507.09916 by the authors.

Figure 1
Figure 1. Model-free dynamic MV portfolio selection: theory [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Model-free dynamic MV portfolio selection: implementation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Testing results after each epoch of training. [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Agent performance in one testing scenario [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Output portfolio and wealth trajectory the RNN encoder Rθ. They are both trained in an adaptive manner via the proposed Algorithm 2. For the GenMarkowitz strategy, while the feature variables h t are updated online, it treats the portfolio selection as a one-period pro…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 68 canonical work pages

  1. [1]

    Nature 630(8016):493--500

    Abramson J, Adler J, Dunger J, Evans R, Green T, Pritzel A, Ronneberger O, Willmore L, Ballard AJ, Bambrick J, Bodenstein SW, Evans DA, Hung CC, O'Neill M, Reiman D, Tunyasuvunakool K, Wu Z, Z emgulyt \.e A, Arvaniti E, Beattie C, Bertolli O, Bridgland A, Cherepanov A, Congreve M, Cowen-Rivers AI, Cowie A, Figurnov M, Fuchs FB, Gladman H, Jain R, Khan YA,...

    2024

  2. [2]

    Review of Financial Studies 15(4):1137--1187

    Ang A, Bekaert G (2002) International asset allocation with regime shifts. Review of Financial Studies 15(4):1137--1187

    2002

  3. [3]

    SIAM Journal on Optimization 27(4):2528--2562

    Backhoff J, Beiglb\" o ck M, Lin Y, Zalashko A (2017) Causal transport in discrete time and applications. SIAM Journal on Optimization 27(4):2528--2562

    2017

  4. [4]

    Finance and Stochastics 24(3):601--632

    Backhoff-Veraguas J, Bartl D, Beiglb \"o ck M, Eder M (2020) Adapted Wasserstein distances and stability in mathematical finance . Finance and Stochastics 24(3):601--632

    2020

  5. [5]

    Management Science 64(3):1136--1154

    Ban GY, El Karoui N, Lim AEB (2018) Machine learning and portfolio optimization. Management Science 64(3):1136--1154

    2018

  6. [6]

    Review of Financial Studies 23(8):2970--3016

    Basak S, Chabakauri G (2010) Dynamic mean--variance asset allocation. Review of Financial Studies 23(8):2970--3016

    work page 2010

  7. [7]

    Review of Financial Studies 9(1):69--107

    Bates DS (1996) Jumps and stochastic volatility: Exchange‐rate processes implicit in deutsche mark options. Review of Financial Studies 9(1):69--107

    work page 1996

  8. [8]

    to appear in Applied Mathematics and Optimization ://arxiv.org/abs/2306.12658, arXiv: 2306.12658

    Bayraktar E, Han B (2025+) Fitted value iteration methods for bicausal optimal transport. to appear in Applied Mathematics and Optimization ://arxiv.org/abs/2306.12658, arXiv: 2306.12658

    arXiv 2025

Show all 85 references
  1. [9]

    Mathematical Finance 24(1):1--24

    Björk T, Murgoci A, Zhou XY (2014) Mean–variance portfolio optimization with state-dependent risk aversion. Mathematical Finance 24(1):1--24

    2014

  2. [10]

    Management Science 68(9):6382--6410

    Blanchet J, Chen L, Zhou XY (2022) Distributionally robust mean-variance portfolio selection with Wasserstein distances . Management Science 68(9):6382--6410

    2022

  3. [11]

    ://arxiv.org/abs/2407.21492, arXiv:2407.21492

    Blanchet J, Larsson M, Park J, Wiesel J (2024) Bounding adapted Wasserstein metrics . ://arxiv.org/abs/2407.21492, arXiv:2407.21492

    2024 arXiv

  4. [12]

    Transactions on Machine Learning Research ISSN 2835-8856

    Bortoli VD (2022) Convergence of denoising diffusion models under the manifold hypothesis. Transactions on Machine Learning Research ISSN 2835-8856

    2022

  5. [13]

    The Review of Financial Studies 18(3):831--873

    Brandt MW, Goyal A, Santa-Clara P, Stroud JR (2005) A simulation approach to dynamic portfolio choice with an application to learning about return predictability. The Review of Financial Studies 18(3):831--873

    2005

  6. [14]

    The Quarterly Journal of Economics 114(2):433--495

    Campbell JY, Viceira LM (1999) Consumption and portfolio decisions when expected returns are time varying. The Quarterly Journal of Economics 114(2):433--495

    1999

  7. [15]

    International Conference on Machine Learning, 4672--4712 (PMLR)

    Chen M, Huang K, Zhao T, Wang M (2023 a ) Score approximation, estimation and distribution recovery of diffusion models on low-dimensional data. International Conference on Machine Learning, 4672--4712 (PMLR)

    2023

  8. [16]

    ://arxiv.org/abs/2002.03938, arXiv: 2002.03938

    Chen M, Liao W, Zha H, Zhao T (2022) Distribution approximation and statistical estimation guarantees of generative adversarial networks. ://arxiv.org/abs/2002.03938, arXiv: 2002.03938

    2022 arXiv

  9. [17]

    The Eleventh International Conference on Learning Representations

    Chen S, Chewi S, Li J, Li Y, Salim A, Zhang A (2023 b ) Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions. The Eleventh International Conference on Learning Representations

    2023

  10. [18]

    Available at SSRN: 3554486 ://ssrn.com/abstract=3554486

    Cong LW, Tang K, Wang J, Zhang Y (2021) AlphaPortfolio: Direct construction through deep reinforcement learning and interpretable AI . Available at SSRN: 3554486 ://ssrn.com/abstract=3554486

    2021

  11. [19]

    ://arxiv.org/abs/2203.01664, arXiv: 2203.01664

    Cont R, Cucuringu M, Xu R, Zhang C (2025) Tail-GAN: learning to simulate tail risk scenarios . ://arxiv.org/abs/2203.01664, arXiv: 2203.01664

    2025 arXiv

  12. [20]

    Mathematical Finance 1(1):1--29

    Cover TM (1991) Universal portfolios. Mathematical Finance 1(1):1--29

    1991

  13. [21]

    Mathematical Finance 29(3):773--803

    Cuchiero C, Schachermayer W, Wong TKL (2019) Cover's universal portfolio, stochastic portfolio theory and the numéraire portfolio. Mathematical Finance 29(3):773--803

    2019

  14. [22]

    Advances in Applied Probability 44(4):1084 -- 1112

    Czichowsky C, Schweizer M (2012) Convex duality in mean-variance hedging under convex trading constraints . Advances in Applied Probability 44(4):1084 -- 1112

    2012

  15. [23]

    DeMiguel V, Garlappi L, Uppal R (2009) Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? The Review of Financial Studies 22(5):1915--1953

    2009

  16. [24]

    Efron B, Tibshirani RJ (1994) An introduction to the bootstrap (Chapman and Hall/CRC)

    1994

  17. [25]

    Proceedings of the 35th International Conference on Machine Learning (ICML) , 1582--1591 (PMLR)

    Fujimoto S, van Hoof H, Meger D (2018) Addressing function approximation error in actor-critic methods. Proceedings of the 35th International Conference on Machine Learning (ICML) , 1582--1591 (PMLR)

    2018

  18. [26]

    Journal of Machine Learning Research 26(43):1--54

    Gao X, Nguyen HM, Zhu L (2025) Wasserstein convergence guarantees for a general class of score-based generative models. Journal of Machine Learning Research 26(43):1--54

    2025

  19. [27]

    Applications of Mathematics (Springer)

    Glasserman P (2004) Monte Carlo Methods in Financial Engineering. Applications of Mathematics (Springer)

    2004

  20. [28]

    The Review of Financial Studies 33(5):2223--2273

    Gu S, Kelly B, Xiu D (2020) Empirical asset pricing via machine learning. The Review of Financial Studies 33(5):2223--2273

    2020

  21. [29]

    European Journal of Operational Research 192(2):500--511

    Guastaroba G, Mansini R, Speranza MG (2009) On the effectiveness of scenario generation techniques in single-period portfolio optimization. European Journal of Operational Research 192(2):500--511

    2009

  22. [30]

    Mathematical Finance 33(3):437--503

    Hambly B, Xu R, Yang H (2023) Recent advances in reinforcement learning in finance. Mathematical Finance 33(3):437--503

    2023

  23. [31]

    Harvard Business Review 42(1):95--106

    Hertz DB (1964) Risk analysis in capital investment. Harvard Business Review 42(1):95--106

    1964

  24. [32]

    Review of Financial Studies 6(2):327--343

    Heston SL (1993) A closed-form solution for options with stochastic volatility with applications to bond and currency options. Review of Financial Studies 6(2):327--343

    1993

  25. [33]

    Advances in Neural Information Processing Systems

    Ho J, Jain A, Abbeel P (2020) Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems

    2020

  26. [34]

    Neural computation 9(8):1735--1780

    Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural computation 9(8):1735--1780

    1997

  27. [35]

    Proceedings of the 39th International Conference on Machine Learning, 8867--8887

    Hoogeboom E, Satorras VG, Vignac C, Welling M (2022) Equivariant diffusion for molecule generation in 3 D . Proceedings of the 39th International Conference on Machine Learning, 8867--8887

    2022

  28. [36]

    The Twelfth International Conference on Learning Representations

    Huang H, Chen M, Qiao X (2024 a ) Generative learning for financial time series with irregular and scale-invariant patterns. The Twelfth International Conference on Learning Representations

    2024

  29. [37]

    ://arxiv.org/abs/2412.16175, arXiv: 2412.16175

    Huang Y, Jia Y, Zhou XY (2024 b ) Mean--variance portfolio selection by continuous-time reinforcement learning: Algorithms, regret analysis, and empirical study. ://arxiv.org/abs/2412.16175, arXiv: 2412.16175

    2024

  30. [38]

    Journal of Machine Learning Research 6:695--709

    Hyv\"arinen A (2005) Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6:695--709

    2005

  31. [39]

    ://arxiv.org/abs/1706.10059, arXiv: 1706.10059

    Jiang Z, Xu D, Liang J (2017) A deep reinforcement learning framework for the financial portfolio management problem. ://arxiv.org/abs/1706.10059, arXiv: 1706.10059

    2017 arXiv

  32. [40]

    4th Edition

    J\,P Morgan (1996) Riskmetrics—technical document. 4th Edition

    1996

  33. [41]

    Finance and Stochastics 21(3):753--787

    Karatzas I, Ruf J (2017) Trading strategies generated by Lyapunov functions . Finance and Stochastics 21(3):753--787

    2017

  34. [42]

    Khasminskii RZ (2012) Stochastic Stability of Differential Equations, volume 66 of Stochastic Modelling and Applied Probability (Berlin: Springer), 2nd edition

    2012

  35. [43]

    Engineering Applications of Artificial Intelligence 125:106739

    Kim J, Lee M (2023) Portfolio optimization using predictive auxiliary classifier generative adversarial networks. Engineering Applications of Artificial Intelligence 125:106739

    2023

  36. [44]

    Advances in Neural Information Processing Systems

    Kollovieh M, Ansari AF, Bohlke-Schneider M, Zschiegner J, Wang H, Wang YB (2023) Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting. Advances in Neural Information Processing Systems

    2023

  37. [45]

    Proceedings of the 5th ACM International Conference on AI in Finance, 1--9

    Kubiak S, Weyde T, Galkin O, Philps D, Gopal R (2024) Denoising diffusion probabilistic model for realistic financial correlation matrices. Proceedings of the 5th ACM International Conference on AI in Finance, 1--9

    2024

  38. [46]

    Advances in Neural Information Processing Systems

    Kwon D, Fan Y, Lee K (2022) Score-based generative modeling secretly minimizes the Wasserstein distance . Advances in Neural Information Processing Systems

    2022

  39. [47]

    Advances in Applied Probability 39(1):77--104

    Labbé C, Heunis AJ (2007) Convex duality in constrained mean-variance portfolio optimization. Advances in Applied Probability 39(1):77--104

    2007

  40. [48]

    Mathematical Finance 10(3):387--406

    Li D, Ng WL (2000) Optimal dynamic portfolio selection: Multiperiod mean--variance formulation. Mathematical Finance 10(3):387--406

    2000

  41. [49]

    Mathematical Finance 34(2):622--670

    Liao S, Ni H, Sabate-Vidales M, Szpruch L, Wiese M, Xiao B (2024) Sig-Wasserstein GANs for conditional time series generation . Mathematical Finance 34(2):622--670

    2024

  42. [50]

    International Conference on Learning Representations (ICLR)

    Lillicrap TP, Hunt JJ, Pritzel A, Heess N, Erez T, Tassa Y, Silver D, Wierstra D (2016) Continuous control with deep reinforcement learning. International Conference on Learning Representations (ICLR)

    2016

  43. [51]

    Mathematics of Operations Research 29(1):132--161

    Lim B Andrew E\ (2004) Quadratic hedging and mean-variance portfolio selection with random parameters in an incomplete market. Mathematics of Operations Research 29(1):132--161

    2004

  44. [52]

    Frontiers of Information Technology & Electronic Engineering 25(1):19--41

    Lin L, Li Z, Li R, Li X, Gao J (2024) Diffusion models for time-series applications: a survey. Frontiers of Information Technology & Electronic Engineering 25(1):19--41

    2024

  45. [53]

    The Journal of Finance 7(1):77--91

    Markowitz H (1952) Portfolio selection. The Journal of Finance 7(1):77--91

    1952

  46. [54]

    ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 8459--8463 (IEEE)

    Marti G (2020) Corrgan: Sampling realistic financial correlation matrices using generative adversarial networks. ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 8459--8463 (IEEE)

    2020

  47. [55]

    Melnik A, Ljubljanac M, Lu C, Yan Q, Ren W, Ritter H (2024) Video diffusion models: A survey ://arxiv.org/abs/2405.03150, arXiv: 2405.03150

    2024 arXiv

  48. [56]

    Journal of Economic Theory 3(4):373--413

    Merton RC (1971) Optimum consumption and portfolio rules in a continuous-time model. Journal of Economic Theory 3(4):373--413

    1971

  49. [57]

    Journal of Financial Economics 3:125--144

    Merton RC (1976) Option pricing when underlying stock returns are discontinuous. Journal of Financial Economics 3:125--144

    1976

  50. [58]

    Mescheder L, Geiger A, Nowozin S (2018) Which training methods for GANs do actually converge? International Conference on Machine Learning, 3481--3490 (PMLR)

    2018

  51. [59]

    Michaud RO (1998) Efficient Asset Management: A Practical Guide to Stock Portfolio Optimization and Asset Allocation (Harvard Business School Press)

    1998

  52. [60]

    Journal of Business 41(2):215--229

    Mossin J (1968) Optimal multiperiod portfolio policies. Journal of Business 41(2):215--229

    1968

  53. [61]

    Journal of Banking & Finance 125:106046

    Paolella MS, Polak P, Walker PS (2021) A non-elliptical orthogonal GARCH model for portfolio selection under transaction costs . Journal of Banking & Finance 125:106046

    2021

  54. [62]

    International Conference on Machine Learning, 8857--8868

    Rasul K, Seward C, Schuster I, Vollgraf R (2021) Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. International Conference on Machine Learning, 8857--8868

    2021

  55. [63]

    Journal of Risk 2(3):21--41

    Rockafellar RT, Uryasev S (2000) Optimization of conditional value-at-risk. Journal of Risk 2(3):21--41

    2000

  56. [64]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 10684--10695

    Rombach R, Blattmann A, Lorenz D, Esser P, Ommer B (2022) High-resolution image synthesis with latent diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 10684--10695

    2022

  57. [65]

    Ronneberger O, Fischer P, Brox T (2015) U-net: Convolutional networks for biomedical image segmentation. Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, 234--2...

    2015

  58. [66]

    International Journal of Forecasting 34(3):497--506

    Sahamkhadam N, Stephan A, Ostermark R (2018) Portfolio Optimization Based on GARCH–EVT–Copula Forecasting Models . International Journal of Forecasting 34(3):497--506

    2018

  59. [67]

    Industrial Management Review 6(2):13--39

    Samuelson PA (1965) Rational theory of warrant pricing. Industrial Management Review 6(2):13--39

    1965

  60. [68]

    Mathematics of Operations Research 20(1):1--32

    Schweizer M (1995) Variance-optimal hedging in discrete time. Mathematics of Operations Research 20(1):1--32

    1995

  61. [69]

    International Conference on Machine Learning, 31016--31029 (PMLR)

    Shen L, Kwok J (2023) Non-autoregressive conditional diffusion models for time series prediction. International Conference on Machine Learning, 31016--31029 (PMLR)

    2023

  62. [70]

    Proceedings of the 32nd International Conference on Machine Learning , 2256--2265

    Sohl-Dickstein J, Weiss E, Maheswaranathan N, Ganguli S (2015) Deep unsupervised learning using nonequilibrium thermodynamics. Proceedings of the 32nd International Conference on Machine Learning , 2256--2265

    2015

  63. [71]

    Advances in Neural Information Processing Systems

    Song Y, Durkan C, Murray I, Ermon S (2021 a ) Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems

    2021

  64. [72]

    Advances in Neural Information Processing Systems 32

    Song Y, Ermon S (2019) Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems 32

    2019

  65. [73]

    International Conference on Learning Representations

    Song Y, Sohl-Dickstein J, Kingma DP, Kumar A, Ermon S, Poole B (2021 b ) Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations

    2021

  66. [74]

    Yale Economic Essays 1(2):179--231

    Sprenkle CM (1961) Warrant prices as indicators of expectations and preferences. Yale Economic Essays 1(2):179--231

    1961

  67. [75]

    ://arxiv.org/abs/2401.13115, arXiv: 2401.13115

    Tang W, Zhao H (2024) Contractive diffusion probabilistic models. ://arxiv.org/abs/2401.13115, arXiv: 2401.13115

    2024 arXiv

  68. [76]

    ://arxiv.org/abs/2402.07487, arXiv: 2402.07487

    Tang W, Zhao H (2025) Score-based diffusion models via stochastic differential equations -- a technical tutorial. ://arxiv.org/abs/2402.07487, arXiv: 2402.07487

    2025 arXiv

  69. [77]

    Advances in Neural Information Processing Systems 34:24804--24816

    Tashiro Y, Song J, Song Y, Ermon S (2021) CSDI: Conditional score-based diffusion models for probabilistic time series imputation . Advances in Neural Information Processing Systems 34:24804--24816

    2021

  70. [78]

    Mathematics of Operations Research 49(2):752--781

    C ern\' y A, Czichowsky C, Kallsen J (2024) Numeraire-invariant quadratic hedging and mean–variance portfolio allocation. Mathematics of Operations Research 49(2):752--781

    2024

  71. [79]

    Grundlehren der mathematischen Wissenschaften (Springer Berlin Heidelberg)

    Villani C (2008) Optimal Transport: Old and New. Grundlehren der mathematischen Wissenschaften (Springer Berlin Heidelberg)

    2008

  72. [80]

    Quantitative Finance 20(9):1419--1440

    Wiese M, Knobloch R, Korn R, Kretschmer P (2020) Quant GANs: deep generation of financial time series . Quantitative Finance 20(9):1419--1440

    2020

  73. [81]

    Yan T, Zhang H, Zhou T, Zhan Y, Xia Y (2021) Scoregrad: Multivariate probabilistic time series forecasting with continuous energy-based generative models ://arxiv.org/abs/2106.10121, arXiv: 2106.10121

    2021 arXiv

  74. [82]

    Advances in Neural Information Processing Systems 32

    Yoon J, Jarrett D, Van der Schaar M (2019) Time-series generative adversarial networks. Advances in Neural Information Processing Systems 32

    2019

  75. [83]

    Applied Mathematics & Optimization 42:19--33

    Zhou XY, Li D (2000) Continuous-time mean–variance portfolio selection: A stochastic LQ framework . Applied Mathematics & Optimization 42:19--33

    2000

  76. [84]

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

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution isbn issn journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sen...

  77. [85]

    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.