pith. sign in

arxiv: 2607.00475 · v1 · pith:HZY4OXMGnew · submitted 2026-07-01 · 💱 q-fin.ST · q-fin.PM· q-fin.TR

End-to-End Parametric Portfolio Policies for Cross-Asset Futures Timing: When Do AI Models Beat Simple Rules?

Austin Pollok , Kevin Robik This is my paper

Pith reviewed 2026-07-02 02:09 UTC · model grok-4.3

classification 💱 q-fin.ST q-fin.PMq-fin.TR
keywords end-to-end policyportfolio timingfutures marketsLSTMtransformerSharpe ratiotransaction costscross-asset allocation
0
0 comments X

The pith

End-to-end AI policies that map market states to weights outperform simple rules on cross-asset futures portfolios, especially when transaction costs are considered.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The authors investigate an alternative to the standard forecast-then-optimize approach by training models to directly produce portfolio weights from market states. They apply this to timing across sixteen liquid CME futures contracts using a loss based on the Sharpe ratio. The resulting policies rank higher than equal weighting, risk parity, and momentum strategies on the combined portfolio and in several asset classes. LSTM and transformer models show similar gross performance, but the transformer trades much less and maintains superiority after costs.

Core claim

Training end-to-end policies on sixteen CME futures with a differentiable Sharpe ratio loss produces models that rank above equal weighting, risk parity, and time-series momentum on the pooled cross-asset portfolio. While LSTM and transformer architectures perform comparably without costs, the transformer creates a stronger policy by generating far lower turnover, allowing it to match or exceed equal weighting through moderate transaction cost levels.

What carries the argument

End-to-end parametric portfolio policy, implemented via LSTM or transformer networks, that directly outputs weights from market states and is trained by maximizing a differentiable Sharpe ratio.

If this is right

  • The learned policies demonstrate an advantage in timing-based tilts that contribute to risk and return in diversified portfolios.
  • Transformer architectures are more suitable for cost-sensitive applications due to significantly lower trading activity compared to LSTMs.
  • Performance gains are observed in the overall portfolio and multiple sub-asset classes, though not consistently across all.
  • Direct optimization of portfolio metrics can bypass the need for intermediate return forecasts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • These policies might extend to equity or fixed income timing if similar state representations are used.
  • Testing on post-sample periods after the training data could confirm stability across market regimes.
  • Alternative loss functions beyond Sharpe ratio may yield policies with different risk profiles.
  • The reduced turnover suggests scalability to larger position sizes without proportional cost increases.

Load-bearing premise

The out-of-sample advantage of the end-to-end policies over rules is robust to variations in training window, feature construction, and hyperparameter selection.

What would settle it

Training the LSTM and transformer on a shifted out-of-sample period or with different input features and checking if the ranking over rules still holds.

Figures

Figures reproduced from arXiv: 2607.00475 by Austin Pollok, Kevin Robik.

Figure 1
Figure 1. Figure 1: Sample period 2001–2024. (a) Rolling one-year Sharpe [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Volatility-matched cumulative returns (growth of $1). Each series is scaled to [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Rolling one-year (252-day) Sharpe ratio on the cross￾asset portfolio, net of a 10 bp transaction cost. The low￾turnover transformer is less eroded by trading costs than the LSTM. The cross-asset results make the cost effect clear. Before costs, the transformer has a slightly higher Sharpe than the LSTM, 0.55 versus 0.50. After applying the 2 bp transaction￾cost rate used for liquid futures as in [9, 7], th… view at source ↗
read the original abstract

Timing-based tilts across asset classes can drive much of the risk and return of a diversified cross-asset portfolio. The standard approach forecasts returns and then optimizes weights. We instead study an end-to-end AI-based policy that maps market states directly to portfolio weights, and we then ask when this one-step modeling approach outperforms simple rules-based strategies. We train these policies on the sixteen most liquid CME futures, where an edge is unlikely to be due to illiquidity, using a differentiable Sharpe ratio loss function, and we benchmark them against equal weighting, risk parity, and time-series momentum. The learned policies rank above the rules on the pooled cross-asset portfolio and in several sub-asset classes, but not uniformly. In gross terms, an LSTM and a transformer-based architecture perform comparably out-of-sample, but diverge when we consider transaction costs. The transformer generates the stronger learned policy, trades far less than the LSTM, and matches or exceeds equal weighting through moderate cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes training LSTM and transformer models as end-to-end parametric policies that map market-state features directly to weights across 16 liquid CME futures contracts, using a differentiable Sharpe-ratio objective. These policies are benchmarked against equal-weight, risk-parity, and time-series momentum rules; the abstract reports that the learned policies rank above the rules on the pooled cross-asset portfolio and in several sub-asset classes, with the transformer outperforming the LSTM once transaction costs are included because of substantially lower turnover.

Significance. If the reported ranking advantages survive proper out-of-sample validation, the work would provide concrete evidence that direct policy learning can improve upon standard rule-based timing strategies in futures markets and would highlight the practical importance of turnover when selecting among neural architectures. The differentiable-Sharpe training approach is a methodological strength that enables truly end-to-end optimization.

major comments (2)
  1. [Abstract] Abstract: the central empirical claim—that the LSTM and transformer policies produce higher pooled and sub-asset rankings than equal-weight, risk-parity, and TSMOM, both gross and net of costs—cannot be evaluated because the abstract (and the supplied text) supplies no information on the sample period, walk-forward or cross-validation scheme, number of random seeds, or any statistical significance tests for the reported rankings.
  2. [Abstract] The claim that the transformer’s lower turnover produces a net-of-cost advantage that “matches or exceeds equal weighting” is load-bearing for the paper’s conclusion about when AI models beat rules; without reported ablation on feature sets, hyperparameter sensitivity, or multiple training windows, it is impossible to rule out that the advantage is an artifact of the particular CME futures sample or market-state construction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful review and for identifying points where additional detail will strengthen the manuscript. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central empirical claim—that the LSTM and transformer policies produce higher pooled and sub-asset rankings than equal-weight, risk-parity, and TSMOM, both gross and net of costs—cannot be evaluated because the abstract (and the supplied text) supplies no information on the sample period, walk-forward or cross-validation scheme, number of random seeds, or any statistical significance tests for the reported rankings.

    Authors: We agree that the abstract and main text must supply these details for the empirical claims to be evaluable. We will revise the abstract to summarize the sample period, the walk-forward validation procedure, the number of random seeds, and the statistical tests performed on the rankings. We will also ensure the methodology section states these elements explicitly and prominently. revision: yes

  2. Referee: [Abstract] The claim that the transformer’s lower turnover produces a net-of-cost advantage that “matches or exceeds equal weighting” is load-bearing for the paper’s conclusion about when AI models beat rules; without reported ablation on feature sets, hyperparameter sensitivity, or multiple training windows, it is impossible to rule out that the advantage is an artifact of the particular CME futures sample or market-state construction.

    Authors: We agree that additional robustness checks are warranted to support the load-bearing claim. The manuscript already reports turnover and net-of-cost metrics across asset classes, but we will add an appendix containing ablations on feature sets, hyperparameter grids, and results from alternative training windows. These additions will allow readers to assess whether the transformer advantage is robust or sample-specific. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical ranking claims rest on independent data comparisons

full rationale

The paper is an empirical machine-learning study that trains LSTM and transformer policies on CME futures returns using a differentiable Sharpe loss and reports out-of-sample rankings versus equal-weight, risk-parity, and TSMOM rules. No equations, derivations, or self-citations appear in the provided text that would reduce any performance claim to a fitted input or prior author result by construction. The central claims are falsifiable via hold-out data and are not self-definitional or load-bearing on self-citation chains. This is the normal non-circular outcome for a pure empirical benchmark paper.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract-only review yields minimal ledger entries; the central claim rests on the unstated assumption that the differentiable Sharpe loss and chosen architectures produce generalizable policies.

free parameters (1)
  • neural network hyperparameters
    LSTM and transformer architectures contain numerous fitted parameters whose values are not reported.
axioms (1)
  • domain assumption Differentiable Sharpe ratio is a suitable training objective for portfolio policies
    Used to train the end-to-end models as stated in the abstract.

pith-pipeline@v0.9.1-grok · 5707 in / 1401 out tokens · 28283 ms · 2026-07-02T02:09:33.264855+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

  1. [1]

    Deep reinforce- ment learning for trading,

    Z. Zhang, S. Zohren, and S. Roberts, “Deep reinforce- ment learning for trading,” J. Financ. Data Sci., vol. 2, no. 2, pp. 25–40, Spring 2020

    work page 2020

  2. [2]

    Enhancing time series momentum strategies using deep neural networks,

    B. Lim, S. Zohren, and S. Roberts, “Enhancing time series momentum strategies using deep neural networks,” J. Financ. Data Sci., vol. 1, no. 4, pp. 19–38, 2019

    work page 2019

  3. [3]

    Deep learning for portfolio optimization,

    Z. Zhang, S. Zohren, and S. Roberts, “Deep learning for portfolio optimization,” J. Financ. Data Sci., vol. 2, no. 4, pp. 8–20, 2020

    work page 2020

  4. [4]

    Trading with the momentum transformer: An intelligent and interpretable architecture,

    K. Wood, S. Giegerich, S. Roberts, and S. Zohren, “Trading with the momentum transformer: An intelligent and interpretable architecture,” arXiv:2112.08534, 2021

    work page arXiv 2021

  5. [5]

    Para- metric portfolio policies: Exploiting characteristics in the cross-section of equity returns,

    M. W. Brandt, P. Santa-Clara, and R. Valkanov, “Para- metric portfolio policies: Exploiting characteristics in the cross-section of equity returns,” Rev. Financ. Stud., vol. 22, no. 9, pp. 3411–3447, Sep. 2009

    work page 2009

  6. [6]

    Optimal versus naive diversification: How inefficient is the1/Nportfolio strategy?

    V . DeMiguel, L. Garlappi, and R. Uppal, “Optimal versus naive diversification: How inefficient is the1/Nportfolio strategy?” Rev. Financ. Stud., vol. 22, no. 5, pp. 1915– 1953, 2009

    work page 1915

  7. [7]

    summer-child

    C. Zhang, Z. Zhang, M. Cucuringu, and S. Zohren, “A universal end-to-end approach to portfolio optimization via deep learning,” arXiv:2111.09170, 2021

    work page arXiv 2021

  8. [8]

    W. B. Powell,Reinforcement Learning and Stochastic Optimization: A Unified Framework for Sequential De- cisions. Wiley, 2022

    work page 2022

  9. [9]

    Portfolio transformer for attention-based asset allocation,

    D. Kisiel and D. Gorse, “Portfolio transformer for attention-based asset allocation,” in Artificial Intelligence and Soft Computing. ICAISC 2022

    work page 2022

  10. [10]

    Deep parametric portfolio policies,

    F. Simon, S. Weibels, and T. Zimmermann, “Deep parametric portfolio policies,” SSRN Working Paper 4150292, 2022

    work page 2022

  11. [11]

    Machine learning and the implementable efficient frontier,

    T. I. Jensen, B. T. Kelly, S. Malamud, and L. H. Peder- sen, “Machine learning and the implementable efficient frontier,” Swiss Finance Inst. Res. Paper No. 22-63, Jun. 2024

    work page 2024

  12. [12]

    Portfolio selection,

    H. Markowitz, “Portfolio selection,” J. Finance, vol. 7, no. 1, pp. 77–91, 1952

    work page 1952

  13. [13]

    Efficient capital markets: A review of theory and empirical work,

    E. F. Fama, “Efficient capital markets: A review of theory and empirical work,” J. Finance, vol. 25, no. 2, pp. 383– 417, 1970

    work page 1970

  14. [14]

    Reinforcement learning for trading,

    J. Moody and M. Saffell, “Reinforcement learning for trading,” in NeurIPS, vol. 11, 1998

    work page 1998

  15. [15]

    Smart ‘predict, then optimize’,

    A. N. Elmachtoub and P. Grigas, “Smart ‘predict, then optimize’,” Manage. Sci., vol. 68, no. 1, pp. 9–26, 2022

    work page 2022

  16. [16]

    Task-based end- to-end model learning in stochastic optimization,

    P. Donti, B. Amos, and J. Z. Kolter, “Task-based end- to-end model learning in stochastic optimization,” in NeurIPS, vol. 30, 2017

    work page 2017

  17. [17]

    From predictive to prescrip- tive analytics,

    D. Bertsimas and N. Kallus, “From predictive to prescrip- tive analytics,” Manage. Sci., vol. 66, no. 3, pp. 1025– 1044, 2020

    work page 2020