Pith. sign in

REVIEW 4 major objections 5 minor 17 references

Quantum-Enhanced Reinforcement Learning with LSTM Forecasting Signals for Optimizing Fintech Trading Decisions

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read In backtests on weekly S&P 500 data from January 2022 to April 2024, a quantum-enhanced A3C agent fed with LSTM forecasts posted a 30.1% cumulative return and 4.01 Sharpe ratio, beating classical A3C and random trading.

desk verdict A cleanly written application paper whose headline result (4.01 Sharpe) is computed entirely in-sample, so it cannot be read as evidence of a real trading edge. read the letter →

arxiv 2507.12835 v1 pith:5TCUDKON submitted 2025-07-17 cs.CE

Yen-Ku Liu , Yun-Huei Pan , Pei-Fan Lu , Yun-Cheng Tsai , Samuel Yen-Chi Chen This is my paper
classification cs.CE MSC 81P6868T0791G80
keywords quantumreinforcementlearningA3CLSTMforecastingvariationalcircuitfinancialtradingS&P500backtestingmacroeconomicsignals
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 a hybrid quantum-classical reinforcement learning agent can beat its classical counterpart in financial trading when both are fed forecasts of next week's market direction. In backtests on weekly S&P 500 data from January 2022 to April 2024, the quantum A3C agent with LSTM predictions achieved roughly a 30.1% cumulative return, a Sharpe ratio of about 4.01, and a maximum drawdown of about 3.78%, the best of the five strategies tested. The authors argue that replacing the classical feedforward encoder in the A3C policy and value networks with a shallow variational quantum circuit gives the agent a richer representation of noisy financial features, and that this lets the LSTM's directional signal add more value than it does for classical A3C. If true, the result points to a practical role for shallow quantum circuits in financial reinforcement learning.

What carries the argument

The load-bearing component is a variational quantum circuit (VQC) placed inside the A3C policy and value networks. The circuit uses eight qubits at depth two, angle-encoded inputs, trainable Ry and Rz rotations, and entangling CNOT gates; a learnable classical projection first maps the state to a latent vector, the VQC transforms it, and classical linear layers produce action logits and state values. This replaces the standard feedforward encoders, providing a nonlinear, high-dimensional representation of market features such as volatility, interest rates, and credit spreads. The second component is the LSTM forecast: a model trained to predict the following week's S&P 500 direction, whose output is appended to the state so the agent can anticipate rather than merely react.

What would settle it

Train the LSTM and all agents on data from January 2022 through December 2023 and then evaluate each trained policy on the unseen January-April 2024 window; if quantum A3C with LSTM signals no longer beats classical A3C on return, Sharpe, and drawdown, the paper's central claim is refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a hybrid quantum-classical A3C agent, augmented with one-week-ahead LSTM forecasts of the S&P 500, outperforms classical A3C and random trading on every headline risk-adjusted metric in the tested window. The quantum-agent-with-forecasts configuration produced a 30.13% cumulative return, a 4.01 Sharpe ratio, a -3.78% maximum drawdown, and a 1.68 Calmar ratio, while classical A3C with the same forecasts returned 11.5% with a 1.38 Sharpe and a -6.79% drawdown. The paper also reports that the LSTM forecast reached 65.38% directional accuracy, and that the quantum agent increased its trading frequency from 24 to 54 cycles when forecasts were added, whereas the classical agent reduced its trading from 22 to 10 cycles. The authors interpret this as evidence that the quantum agent can exploit short-term fluctuations the classical agent cannot.

Load-bearing premise

The paper's reported metrics are computed on the same January 2022 to April 2024 data used to train both the LSTM and the reinforcement learning agents, and no temporal holdout or train/test split is described, so the superior results could be the product of in-sample fitting rather than a real trading edge.

Editorial extensions

If this is right

  • Quantum-enhanced A3C with LSTM forecasts may offer materially better risk-adjusted returns than classical A3C in volatile markets, making hybrid quantum-classical policies a candidate for real financial decision systems.
  • Forecast signals are not uniformly useful: they made the classical agent more conservative and the quantum agent more active, implying that model architecture determines whether predictive data helps or hurts.
  • Shallow variational circuits of eight qubits and depth two appear sufficient to change learning dynamics, which supports experimentation on near-term quantum hardware.
  • A directional accuracy of 65.38% from the LSTM was enough for the quantum agent to build an edge, suggesting that the marginal value of forecast precision is limited once the policy can represent the signal.
  • The reported superiority holds only for a single asset and a single backtest window, so the immediate corollary is a need for out-of-sample and multi-asset validation.

Reading between the lines

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

  • The absence of any train/test split in the paper means the headline numbers should be read as in-sample results; re-evaluating with a temporal holdout would reveal whether the quantum advantage survives out-of-sample data.
  • The trading environment rewards only sell actions with portfolio profit, so high trading frequency can mechanically harvest the reward signal; a continuous portfolio-return reward would test whether the quantum agent's edge is real or an artifact of reward design.
  • Comparing quantum circuit depths (1, 2, 4, 6) while holding the LSTM fixed would pin down how much representational capacity the quantum layer provides, and would connect this result to the broader question of when quantum feature maps help.
  • The same architecture could be tested on multi-asset portfolios or on regime-switching benchmarks to see whether the forecast-quantum synergy generalizes beyond the S&P 500.
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 / 5 minor

Summary. This manuscript presents a hybrid quantum-classical A3C agent with LSTM-based one-week-ahead forecasts for S&P 500 trading. It reports training curves, trading frequencies, and performance metrics for Classical A3C, Quantum A3C, with and without LSTM signals, and a random policy on weekly data from January 2022 to April 2024. The headline conclusion is that Quantum A3C with LSTM signals achieves the highest cumulative return (approximately 30.1%), the highest Sharpe ratio (approximately 4.01), and the lowest drawdown (approximately 3.78%). The evaluation protocol is the paper's central weakness: all metrics are computed from a single in-sample execution on the training window, with no temporal holdout, no repeated seeds, and no buy-and-hold baseline, so the reported advantage cannot currently be distinguished from overfitting.

Significance. If the results were reproduced out-of-sample, the paper would offer a useful demonstration of quantum reinforcement learning in a financial sequential decision-making task, together with a concrete modular environment (SP500TradingEnv plus TorchVQC) that others could build on. The manuscript is also honest about its narrow market and period. However, the absence of a held-out evaluation, statistical repetition, and a buy-and-hold baseline means that the current evidence does not support the stated conclusion of 'superior performance and stability' for quantum models. The contribution is therefore at the level of a promising architecture and environment, not an established empirical result.

major comments (4)
  1. [II-C(c), III-C, Table II, IV] The headline results are computed on the same window used for training. Section II-A defines the dataset as the period from January 2022 to April 2024, and Section II-C(c) states that after training 'the global policy G(θ) is executed once in SP500TradingEnv to generate the final asset history.' No train/test split, walk-forward scheme, or held-out period is described anywhere. Table II and Fig. 4 are therefore in-sample evaluations of the LSTM-augmented policy and of the A3C policies trained in that environment. An agent that memorizes the training path can look excellent on these metrics, so the 30.13% return, 4.01 Sharpe, and 3.78% drawdown cannot be attributed to trading skill. The limitation paragraph in Section IV mentions single-asset and limited-period scope but does not acknowledge the missing temporal holdout, which is a more serious threat to the claim.
  2. [III-B, III-C, Figs. 2, 4] All numerical comparisons appear to come from one stochastic run per configuration. A3C involves random initialization and stochastic policy sampling, and the VQC parameters are also randomly initialized, yet Table I and Table II report single point estimates with no standard deviations, confidence intervals, or number of seeds. With a single run, the observed gaps, e.g., 14.9% versus 15.7% return between Classical A3C and Quantum A3C without predictions, may be within run-to-run noise. The paper should report means and dispersions over multiple seeds for every metric in Tables I and II, and the trajectories in Fig. 4 should be representative rather than selected.
  3. [II-B, Eq. (1), Table II] The comparison omits the most natural financial baseline, buy-and-hold of the S&P 500, and it does not specify the trade cost rate c that appears in the reward function in Eq. (1). The value of c is never stated, and the computation of 'Sharpe Ratio' (annualization, risk-free rate) is not described. Without a buy-and-hold benchmark and an explicit cost model, even a properly out-of-sample result would not show that the learned policies add value over holding the index. Random trading is not a sufficient baseline for the claims made.
  4. [II-A, III-A, Fig. 3] The LSTM is presented as providing 'predictions of the following week's S&P 500 direction,' but no training/validation split for the LSTM is described. Section III-A reports LSTM RMSE 2.02, Pearson correlation 0.595, and directional accuracy 65.38% on what appears to be the same January 2022-April 2024 period plotted in Fig. 3. If the LSTM was fit on that interval, these are in-sample fits, not forecasts, and the RL agents are then trained on signals that are already adapted to the evaluation period. The paper must state how the LSTM was trained and validated (e.g., rolling-origin evaluation) before the 'predictive signal' contribution can be assessed.
minor comments (5)
  1. [II-B, Eq. (1)] The symbol '⊮position=1' is nonstandard and should be written as the indicator function 1_{position=1}; the treatment of an open position at episode end is also unspecified.
  2. [II-C] Key hyperparameters are missing, including the number of A3C workers N, the learning rate α, MAXEP, UPDATE_GLOBAL_ITER, and the LSTM architecture and training epochs; without these or a code release, the experiments are not reproducible.
  3. [References] References [10] and [15] are identical (Havlček et al., 'Supervised learning with quantum-enhanced feature spaces') and should be consolidated.
  4. [Table I and Table II] Classical A3C with LSTM is described as 'extremely conservative' with 10 trades in Table I, but Table II reports 63% time in market; the relationship between trade count and time in market should be clarified to avoid an apparent contradiction.
  5. [Fig. 2] The left panel of Fig. 2 is described as showing training curves over 3,000 episodes, but the axes are unlabeled; adding axis labels and a legend for all five strategies would make the convergence claim verifiable.

Circularity Check

2 steps flagged · score 6.0 of 10

The reported 30.1% return and 4.01 Sharpe are computed by training the LSTM and the A3C agent on the same Jan 2022–Apr 2024 window and then replaying the learned policy in the same environment; the 'prediction' is an in-sample fitted value, not an independent forecast.

  1. fitted input called prediction [Section II-A (Dataset Description); Section III-A (Training Performance), Fig. 3]
    "Predictive versions augment the dataset with LSTM forecasts of the following week's S&P 500 direction, enhancing the agent's environmental perception. ... As shown in Fig. 3, the LSTM achieves an RMSE of 2.02, Pearson correlation of 0.595, and directional accuracy of 65.38%."

    The LSTM forecast is fit on the January 2022 to April 2024 window, and its accuracy is reported on the same window; no temporal train/test split or walk-forward holdout is described in Sections II-A or III-A. The 'following week's S&P 500 direction' is therefore a fitted value within the evaluation period, not an independent out-of-sample forecast. Any improvement in the agent attributed to these predictive signals is measured on the same data used to fit the LSTM, so the claimed predictive benefit cannot be distinguished from fitting noise.

  2. fitted input called prediction [Section II-B (Environment Formalization); Section II-C c (Evaluation & Plotting); Section III-C (Trading Outcomes)]
    "Episodes terminate at the end of the dataset. ... Post-training, the global policy G(θ) is executed once in SP500TradingEnv to generate the final asset history. ... Quantum A3C with Predict Data yields the highest return (≈ 30.13%), Sharpe ratio (≈ 4.01), and lowest drawdown (≈ 3.78%)."

    The A3C policy learns on episodes that terminate at the end of the dataset, and the headline metrics are produced by executing the trained policy once in the same SP500TradingEnv. The reported cumulative return, Sharpe ratio, and drawdown are therefore training-window performance statistics rather than out-of-sample predictions. Because the agent was optimized on the same market path used for evaluation, the claimed quantum-plus-LSTM superiority is statistically forced by in-sample fitting and cannot be separated from memorization.

full rationale

The paper's central quantitative claim — QRL with LSTM forecasts achieves the highest return, Sharpe ratio, and lowest drawdown — rests entirely on Table II and Fig. 4, which are generated by training both the LSTM and the A3C agent on the January 2022–April 2024 dataset and then executing the learned policy once in the same environment. The paper's own descriptions ('Episodes terminate at the end of the dataset'; 'Post-training, the global policy G(θ) is executed once in SP500TradingEnv to generate the final asset history') show that there is no temporal holdout, and Fig. 3 reports LSTM accuracy on the same period. Consequently, the 'following week's S&P 500 direction' signal is an in-sample fitted variable, and the reported superiority of the quantum hybrid is not an out-of-sample prediction. This fits pattern 2 (fitted input called prediction) rather than definitional circularity: the result is not derived from itself, but the headline evidence reduces to training-window performance, so the comparative advantage is statistically forced rather than independently demonstrated. The self-citations to Chen et al. [16]–[18] are background architecture credits and are not load-bearing for the new experimental numbers, so they do not raise the score further.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the dataset plus several unverified modeling assumptions. The main burden is the in-sample evaluation and the unspecified cost and hyperparameters.

free parameters (4)
  • Discount factor gamma = 0.9
    Set by hand for episodic evaluation; it affects the value of future rewards and therefore policy learning.
  • Trade cost rate c = unspecified
    Appears in the reward function but no value is given; the outcomes depend on this cost.
  • LSTM forecast model parameters = RMSE 2.02, directional accuracy 65.38% on the same dataset
    The LSTM is fit to the S&P 500 and macro data, then its predictions are used as state features; the fitted model is part of the evaluated pipeline.
  • Quantum circuit hyperparameters = 8 qubits, depth 2
    Chosen by hand; no ablation shows sensitivity to these choices.
assumptions (3)
  • domain assumption The historical S&P 500 and macro time series define a valid MDP transition function for the trading task.
    Section II-B treats P as implicitly defined by historical data; this assumes the observed sequence is a faithful and stable environment for RL.
  • domain assumption The LSTM predictions are causally valid and do not leak future information into the state.
    Section II-A claims each point uses only current values at time t, but the LSTM target is the following week's direction and no separate training window is described.
  • domain assumption The reward function with a fixed cost c captures realistic trading, ignoring slippage, spreads, and market impact.
    Section II-B defines reward as price difference minus cost; c is never specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum-Enhanced Reinforcement Learning with LSTM Forecasting Signals for Optimizing Fintech Trading Decisions." pith.science (2026). https://pith.science/paper/5TCUDKON

@misc{pith2026250712835,
  author       = {Pith},
  title        = {Pith review of: Quantum-Enhanced Reinforcement Learning with LSTM Forecasting Signals for Optimizing Fintech Trading Decisions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5TCUDKON}},
  note         = {Machine review of arXiv:2507.12835}
}
read the original abstract

Financial trading environments are characterized by high volatility, numerous macroeconomic signals, and dynamically shifting market regimes, where traditional reinforcement learning methods often fail to deliver breakthrough performance. In this study, we design a reinforcement learning framework tailored for financial systems by integrating quantum circuits. We compare (1) the performance of classical A3C versus quantum A3C algorithms, and (2) the impact of incorporating LSTM-based predictions of the following week's economic trends on learning outcomes. The experimental framework adopts a custom Gymnasium-compatible trading environment, simulating discrete trading actions and evaluating rewards based on portfolio feedback. Experimental results show that quantum models - especially when combined with predictive signals - demonstrate superior performance and stability under noisy financial conditions, even with shallow quantum circuit depth.

Figures

Figures reproduced from arXiv: 2507.12835 by the authors.

Figure 1
Figure 1. Experimental pipeline integrating LSTM-based prediction with Classical A3C, Quantum A3C, and Random trading [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Left: Training curves across different reinforcement learning strategies, showing raw and average episodic rewards over [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Predicted vs Actual Weekly Return (%) for S&P 500 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Cumulative returns over the trading period for each [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 10 canonical work pages

  1. [1]

    Empirical properties of asset returns: stylized facts and statistical issues,

    R. Cont, “Empirical properties of asset returns: stylized facts and statistical issues,” Quantitative finance, vol. 1, no. 2, p. 223, 2001

    work page 2001

  2. [2]

    A Novel Experts Advice Aggregation Framework Using Deep Reinforcement Learning for Portfolio Management

    M. Fazli, M. Lashkari, H. Taherkhani, and J. Habibi, “A novel experts advice aggregation framework using deep reinforcement learning for portfolio management,” 2022. [Online]. Available: https://arxiv.org/abs/2212.14477

    work page Pith review arXiv 2022

  3. [3]

    Performance functions and reinforcement learning for trading systems and portfolios,

    J. Moody, L. Wu, Y . Liao, and M. Saffell, “Performance functions and reinforcement learning for trading systems and portfolios,” Journal of forecasting, vol. 17, no. 5-6, pp. 441–470, 1998

    work page 1998

  4. [4]

    Deep direct rein- forcement learning for financial signal representation and trading,

    Y . Deng, F. Bao, Y . Kong, Z. Ren, and Q. Dai, “Deep direct rein- forcement learning for financial signal representation and trading,” IEEE transactions on neural networks and learning systems , vol. 28, no. 3, pp. 653–664, 2016

    work page 2016

  5. [5]

    Asynchronous methods for deep rein- forcement learning,

    V . Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, “Asynchronous methods for deep rein- forcement learning,” in International conference on machine learning . PmLR, 2016, pp. 1928–1937

    2016

  6. [6]

    Deep reinforcement learning that matters,

    P. Henderson, R. Islam, P. Bachman, J. Pineau, D. Precup, and D. Meger, “Deep reinforcement learning that matters,” in Proceedings of the AAAI /uni00000013/uni00000018/uni00000013/uni00000013/uni00000014/uni00000013/uni00000013/uni00000013/uni00000014/uni00000018/uni00000013/uni00000013/uni00000015/uni00000013/uni00000013/uni00000013/uni00000015/uni000000...

    work page

  7. [7]

    Quantum machine learning,

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671, pp. 195–202, 2017

    2017

  8. [8]

    Supervised learning with quantum computers,

    M. Schuld and F. Petruccione, “Supervised learning with quantum computers,” Quantum science and technology , vol. 17, 2018

    work page 2018

Show all 17 references
  1. [9]

    Variational quantum algorithms,

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio et al. , “Variational quantum algorithms,” Nature Reviews Physics , vol. 3, no. 9, pp. 625– 644, 2021

    2021

  2. [11]

    Quantum feature optimization for enhanced clustering of blockchain transaction data,

    Y .-C. Tsai and S. Y .-C. Chen, “Quantum feature optimization for enhanced clustering of blockchain transaction data,” 2025. [Online]. Available: https://arxiv.org/abs/2505.16672

    2025 arXiv

  3. [12]

    Quantum-enhanced machine learning,

    V . Dunjko, J. M. Taylor, and H. J. Briegel, “Quantum-enhanced machine learning,” Physical review letters, vol. 117, no. 13, p. 130501, 2016

    2016

  4. [13]

    Quantum computing in the nisq era and beyond,

    J. Preskill, “Quantum computing in the nisq era and beyond,” Quantum, vol. 2, p. 79, 2018

    2018

  5. [14]

    Quantum machine learning in feature hilbert spaces,

    M. Schuld and N. Killoran, “Quantum machine learning in feature hilbert spaces,” Physical review letters, vol. 122, no. 4, p. 040504, 2019

    2019

  6. [15]

    Supervised learning with quantum- enhanced feature spaces,

    V . Havl ´ıˇcek, A. D. C ´orcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, “Supervised learning with quantum- enhanced feature spaces,” Nature, vol. 567, no. 7747, pp. 209–212, 2019

    2019

  7. [16]

    Asynchronous training of quantum reinforcement learn- ing,

    S. Y .-C. Chen, “Asynchronous training of quantum reinforcement learn- ing,” Procedia Computer Science , vol. 222, pp. 321–330, 2023

    2023

  8. [17]

    Efficient quantum recurrent reinforcement learning via quantum reservoir computing,

    ——, “Efficient quantum recurrent reinforcement learning via quantum reservoir computing,” in ICASSP 2024-2024 IEEE International Con- (a) Classical A3C without fore- casting (b) Classical A3C with forecast- ing (c) Quantum A3C without fore- casting (d) Quantum A3C with forecas...

    2024

  9. [18]

    Differentiable quantum architecture search in asynchronous quantum reinforcement learning,

    ——, “Differentiable quantum architecture search in asynchronous quantum reinforcement learning,” in Proceedings of the 2024 IEEE International Conference on Quantum Computing and Engineering (QCE). IEEE, 2024, pp. 1516–1524

    2024

Pith tools

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