Pith. sign in

REVIEW 4 major objections 6 minor 91 references

MIGT: Memory Instance Gated Transformer Framework for Financial Portfolio Management

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

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

Pith's one-line read MIGT, a gated-attention trading network, beats 15 strategies on Dow data.

desk verdict A cleanly specified transformer/PPO portfolio architecture whose empirical case rests on single-run results that do not support the advertised margins. read the letter →

arxiv 2502.07280 v1 pith:MJWGMT7Z submitted 2025-02-11 cs.LG cs.AI

Fengchen Gu , Angelos Stefanidis , Ángel García-Fernández , Jionglong Su , Huakang Li This is my paper · ORCID
classification cs.LGcs.AI
keywords portfoliomanagementdeepreinforcementlearningtransformerinstancenormalizationgatedattentionmemorytrajectoryPPODowJonesIndustrialAverage
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a deep-reinforcement-learning policy network for portfolio management, called MIGT, outperforms fifteen established trading strategies on backtests of the Dow Jones Industrial Average 30. The network wraps a transformer-style attention module with instance normalization and a lightweight gating layer, and adds a memory trajectory that carries previous hidden states into the attention computation. In the paper's experiments, MIGT's cumulative return is at least 9.75 percentage points higher than every comparison strategy, and its Sharpe, Sortino, and $\Omega$ ratios are also the highest. The result matters because it suggests that stabilizing training through per-sample normalization and gated residuals, rather than adding more data or more complex risk models, may be the bottleneck for profitable DRL trading agents.

What carries the argument

The load-bearing object is the Gated Instance Attention module, a modified transformer block in which multi-head scaled dot-product attention is preceded by instance normalization, and a Lite Gate Unit (LGU) serves as a gated fan-in layer that mixes the residual value with the gated hidden state via an update gate. A memory trajectory caches a fixed length of old hidden states and feeds them into the attention layer as extra context, letting the network see beyond a fixed window. This module is inserted between a fully connected input layer and a position-wise MLP that feeds a logits MLP, and the whole policy network is trained with PPO. Instance normalization is doing the outlier-suppression work, the gate is doing the training-stabilization work, and the multi-head attention plus memory is doing the long-range dependency capture; the paper attributes the performance gains to these three mechanisms acting together.

What would settle it

Re-run the same DJIA experiments with, say, ten random seeds per strategy (at least for MIGT and the strongest competitor TMP), and check whether MIGT's cumulative-return distribution still sits above the competitor's by the reported margin; if the 9.75-point gap falls within one standard deviation of either strategy, the claimed advantage is not established. A simpler check: test MIGT's policy on data after 2021 or on a different index, where the paper offers no evidence the advantage transfers.

Watch

Extended reading notes

Core claim

The paper's central discovery claim is that its proposed MIGT architecture achieves the best performance in all three DJIA backtest years (2019, 2020, and 2021), and on the combined 2019-2021 test period, beating DRL frameworks such as EIIE, FinRL, ES, TMP, and SARL plus ten classical mean-reversion and trend-following strategies. The reported cumulative-return margin over the runner-up is at least 9.75 percentage points in the single-year tests, and at least 126.21 percentage points on the three-year combined test; MIGT also posts the highest Sharpe, Sortino, and $\Omega$ ratios in every experiment, with the $\Omega$ margin the narrowest. The paper further claims that each component of the Gated Instance Attention module earns its keep: removing instance normalization costs an average of 5.79% in cumulative return, removing the gating layer costs 8.67%, and removing the transformer variant cuts the three-year average annual return by 22.76%; injecting 10% outlier pseudo-data makes the unnormalized variant lose 47.60% of its return while MIGT loses only 17.01%.

Load-bearing premise

The load-bearing premise is that one PPO training run per configuration represents that strategy's true performance, since the paper reports no random seeds, standard deviations, or confidence intervals for the backtest results.

Editorial extensions

If this is right

  • If the reported margins hold up under repeated training runs, MIGT would be the strongest published DRL portfolio manager on DJIA data among the fifteen strategies tested, which would push the field's default policy network from recurrent or simple MLP heads toward gated attention with memory.
  • The ablation numbers imply that instance normalization is what protects the strategy when outliers contaminate the data: with 10% outlier pseudo-data added, MIGT loses 17.01% of its return while the unnormalized version loses 47.60%.
  • Removing the transformer variant reduces the three-year average annual return by 22.76%, which the paper interprets as evidence that long-range attention over the time dimension is doing real work for profitability rather than just adding parameters.
  • The LGU gating layer yields faster early convergence in the reported training-curve plots, meaning that the gate's role is mostly to make PPO training reach a higher-return policy in fewer steps.
  • Across the three years, MIGT's cumulative-return edge grows when the backtest is extended to the full 2019-2021 period, reaching a margin of at least 126.21 percentage points over the next-best strategy, which the paper reads as evidence of durable rather than one-year luck.

Reading between the lines

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

  • We infer that the same module stack—instance normalization ahead of attention, a gated residual, and cached memory—could improve PPO in other non-stationary sequential decision problems, such as robot sim-to-real transfer or energy demand response, although the paper only tests stock portfolios.
  • The paper's long-only constraint bounds the claim: with short-selling allowed, the action space changes and the ranking of strategies might shift in bear markets, so the reported advantage is specific to long-only DJIA portfolios.
  • We infer that the pseudo-data outlier protocol (injecting synthetic spikes at known rates and measuring return decay) could serve as a standard stress test for DRL trading policies, independent of the specific architecture.
  • Until multiple seeds and variance statistics are reported, the architectural recipe is the more durable takeaway than the precise return margins; the exact percentages in Tables 2 and 3 should be read as point estimates from single runs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes MIGT, a PPO-based portfolio management framework whose policy network combines a Transformer variant with instance normalization, a Lite Gate Unit (LGU) gating layer, and a memory trajectory cache. The authors formulate the portfolio task as an MDP, define the reward as the transaction-cost-adjusted change in portfolio value, and evaluate MIGT against fifteen classical and DRL baselines on DJIA data over test years 2019, 2020, and 2021, reporting cumulative returns, Sharpe, Sortino, and Omega ratios. The central claim is that MIGT outperforms all baselines by at least 9.75% in cumulative returns and at least 2.36% in risk-return ratios. The paper also includes ablation studies on the normalization, gating, and transformer components, and a pseudo-data experiment intended to demonstrate robustness to outliers. The central claim rests on a single training run per configuration with no reported variance, and the paper does not provide code or data.

Significance. If the reported performance were supported by proper statistical evidence, the MIGT framework would be a meaningful empirical contribution to DRL-based portfolio management: it combines several plausible mechanisms (instance normalization, gated residual paths, memory-augmented attention) and evaluates against a broad set of baselines. The formal parts of the paper—the portfolio-value algebra in Section 2 and the PPO update in Eq. (8)—are standard and clearly stated. However, the empirical central claim is currently not verifiable: only one trajectory per configuration is reported, no seeds or confidence intervals are given, and the code/data are not released. The paper also uses an ad-hoc pseudo-data injection scheme whose relation to real market outliers is not established. These issues make the headline 'at least 9.75% improvement' unsupported, so the significance of the contribution cannot be assessed from the manuscript as it stands.

major comments (4)
  1. [Section 5.4, Tables 2–4] The central empirical claim is based on a single training run per configuration. PPO is stochastic in initialization, exploration, and minibatch order, yet no seeds, standard deviations, or confidence intervals are reported. The smallest cumulative-return gap to the second-best strategy is 9.75 percentage points (Dataset 2, MIGT 0.24440 vs. EIIE 0.14689); if run-to-run variance is of that magnitude, the claim of consistent superiority over all fifteen strategies collapses. The manuscript needs multiple independent runs per configuration with reported mean, standard deviation, and ideally per-seed trajectories before the result can be evaluated.
  2. [Abstract and Section 5.4] The advertised 'at least 9.75% improvement in cumulative returns and a minimum 2.36% increase in risk-return ratios' mixes incomparable quantities: the 9.75% is an absolute percentage-point gap in cumulative returns, while the 2.36% is a relative change in the Omega ratio. Since the two margins are not on the same scale and are taken from different datasets, the notion of a single 'minimum' improvement is not well-defined. The authors should state exactly which metrics and baselines define each advertised margin, or replace the composite claim with per-dataset, per-metric comparisons that include uncertainty.
  3. [Section 5.5.2 and Figure 12] The pseudo-data ablation injects synthetic outlier samples into the training set to test instance normalization, but no justification is given that the injected pseudo-data distribution resembles real market outliers or black-swan events. The observed differences in cumulative return between MIGT and MIGT w/o Norm under pseudo-data perturbations therefore do not establish robustness to realistic outliers. The authors should either use empirically motivated outlier injection (e.g., drawn from historical crisis periods) or temper the claim that instance normalization 'significantly mitigates the negative impact of outliers.'
  4. [Section 5.1 and Table 1] The experimental protocol uses the same three test years (2019–2021) that are also likely used for architecture and indicator selection, and no validation split is mentioned. If hyperparameters, feature sets, or component choices were tuned on these test years, the reported backtest performance is subject to selection bias. The authors should clarify whether any validation data outside the test periods were used for model selection, and if so, describe that procedure; otherwise the comparisons risk overfitting to the test period.
minor comments (6)
  1. [Declarations] The Data Availability statement says data 'will be made available on request,' but no code repository or model checkpoint is provided; this prevents independent reproduction of the single-run results.
  2. [Equation (15)] The Sortino ratio formula uses both 'r_q' and 'r' inconsistently; the notation should be unified to avoid ambiguity about the minimum acceptable return.
  3. [Figure 10] The axis labels in Figure 10 appear garbled and contain non-English characters; the figure should be regenerated with clear English labels for holdings and returns.
  4. [Section 5.5.1] The sentence 'The strategy applying Attention and Memory Trajectory increases the return by an average of 5.25% compared to the basic reinforcement learning framework' lacks a clear definition of the 'basic' baseline; please specify which ablation branch serves as the reference.
  5. [Abstract and Section 1] Several references are loosely cited (e.g., 'Bengio & Courville, 2013' for deep learning and 'Dai et al., 2020' for gating), and some bibliographic entries are incomplete; a copy-editing pass would improve the presentation.
  6. [Section 6] The limitations paragraph is a positive addition, but it should be expanded to acknowledge the lack of multiple seeds and the absence of a proper validation split, as these directly affect the reliability of the reported results.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the performance claim is an out-of-sample empirical backtest, not a derivation equivalent to its inputs; the same-group citations are motivation-level and non-load-bearing.

full rationale

The paper's central result is an empirical backtest, not a derived equivalence. Section 5.1 fixes a temporal split (train on three prior calendar years, test on the next), and Section 5.4 reports cumulative returns and risk-adjusted ratios on those held-out years. Nothing in Equations (1)-(16) fits a parameter to the test period and then reports that fitted value as a prediction; the reward in Equation (7) is the per-period portfolio value change, and the evaluation metric in Equation (13) compounds that same quantity, but this consistency is shared by all compared DRL agents rather than being a self-referential derivation of MIGT's superiority. The same-group citations (e.g., Gu, Jiang, & Su 2021; Gao et al. 2022; Sun et al. 2021; Zhang et al. 2022) appear in the motivation and related-work discussion and are not used to justify the reported 9.75% margin, so they are not load-bearing. The absence of multiple seeds, standard deviations, and confidence intervals around Tables 2-4, and the single-trajectory plots in Figures 7-9 and 13, are genuine statistical robustness weaknesses, but they are not circularity. The authors' own Section 6 limitations (modest Omega-ratio advantage and long-only constraint) are acknowledged and do not make the claim self-referential. The reuse of Transformer-XL-style memory and gated-recurrence ideas under new names is an attribution/originality concern rather than a circular derivation of the backtest results. Because the self-citations present are non-load-bearing, a score of 2 rather than 0 is appropriate.

Assumptions & free parameters 9 free parameters · 7 assumptions · 2 invented entities

The central claim rests on a set of hand-chosen architectural and training hyperparameters, several of which are unreported, plus assumptions about market behavior (no price impact, long-only, technical indicators as sufficient state). The only genuinely new component, LGU, is supported solely by in-paper ablations.

free parameters (9)
  • Transaction cost rate c = 0.1%
    Assumed, not fitted; Sharpe and Sortino comparisons would shift with fee level.
  • Risk-free rate / minimum acceptable return r = 3% per year
    Set to the highest level over 2019-2021, affecting Sharpe and Sortino comparisons.
  • Attention dimension = 64
    Hand-chosen; no sensitivity analysis reported.
  • Number of attention heads = 4
    Hand-chosen; no sensitivity analysis reported.
  • Head dimension = 32
    Hand-chosen; equals attention dimension divided by heads.
  • PW-MLP hidden dimension = 32
    Hand-chosen; no sensitivity analysis reported.
  • Memory trajectory length
    Described as a predetermined length but the value is never stated; required for reproduction.
  • PPO hyperparameters (learning rate, clip epsilon, GAE, batch size, training steps)
    Not reported anywhere in the paper.
  • Pseudo-data percentages in Section 5.5.2 = 0%, 5%, 10%
    Chosen for the outlier-robustness demonstration; no justification for these proportions.
assumptions (7)
  • domain assumption The portfolio problem is a Markov decision process with state st equal to the listed price and indicator features.
    Section 3.2 defines st as feature vectors but provides no proof that these features are a sufficient statistic for future prices.
  • domain assumption Transactions do not affect market prices.
    Section 2.2 assumption, standard for this literature, but untestable at daily frequency for a large fund.
  • domain assumption Only long positions are allowed; no short selling.
    Section 2.2; acknowledged in Section 6 as a limitation.
  • domain assumption Technical indicators (BOLL, CCI, RSI, TR, DMI, MACD, MFI) carry predictive information for portfolio weights.
    Section 3.2 lists these indicators as state inputs without empirical evidence.
  • domain assumption PPO training with the given reward converges reliably to a good policy.
    Section 3.2 and 4.2 assume PPO's clipped objective works; no convergence guarantee, and Figure 13 shows only one run.
  • ad hoc to paper Single-seed results are representative of expected performance.
    Section 5.4 reports each table entry as a point value with no variance, implying determinism that PPO does not have.
  • ad hoc to paper Pseudo-data injection in Section 5.5.2 simulates real outliers.
    The outlier-robustness claim rests on synthetic perturbations whose realism is not validated.
invented entities (2)
  • Lite Gate Unit (LGU)
    purpose: Gating layer replacing residual connections to stabilize PPO training (Eqs. 10-12).
    A simplified GRU-style update gate; the only evidence is the paper's own ablation (Table 4, Figure 13). No external validation.
  • Memory trajectory cache
    purpose: Stores past hidden states as additional attention input to capture long-range dependencies.
    Reuses Transformer-XL's segment-level memory (Dai et al.); performance contribution is supported only by the paper's ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MIGT: Memory Instance Gated Transformer Framework for Financial Portfolio Management." pith.science (2026). https://pith.science/paper/MJWGMT7Z

@misc{pith2026250207280,
  author       = {Pith},
  title        = {Pith review of: MIGT: Memory Instance Gated Transformer Framework for Financial Portfolio Management},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJWGMT7Z}},
  note         = {Machine review of arXiv:2502.07280}
}
read the original abstract

Deep reinforcement learning (DRL) has been applied in financial portfolio management to improve returns in changing market conditions. However, unlike most fields where DRL is widely used, the stock market is more volatile and dynamic as it is affected by several factors such as global events and investor sentiment. Therefore, it remains a challenge to construct a DRL-based portfolio management framework with strong return capability, stable training, and generalization ability. This study introduces a new framework utilizing the Memory Instance Gated Transformer (MIGT) for effective portfolio management. By incorporating a novel Gated Instance Attention module, which combines a transformer variant, instance normalization, and a Lite Gate Unit, our approach aims to maximize investment returns while ensuring the learning process's stability and reducing outlier impacts. Tested on the Dow Jones Industrial Average 30, our framework's performance is evaluated against fifteen other strategies using key financial metrics like the cumulative return and risk-return ratios (Sharpe, Sortino, and Omega ratios). The results highlight MIGT's advantage, showcasing at least a 9.75% improvement in cumulative returns and a minimum 2.36% increase in risk-return ratios over competing strategies, marking a significant advancement in DRL for portfolio management.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

91 extracted references · 61 canonical work pages

  1. [1]

    write newline

    " write newline " cite write " FUNCTION editor.postfix editor num.names #1 > "( )" "( )" if FUNCTION editor.trans.postfix editor num.names #1 > "( )" "( )" if FUNCTION trans.postfix translator num.names #1 > "( )" "( )" if FUNCTION authors.editors.reflist.apa5 'field := 'dot := field num.names 'numnames := numnames 'format.num.names := format.num.names na...

  2. [2]

    , Hartford, D

    Ale2020Dragons APACrefauthors Ale, B. , Hartford, D. Slater, D.H. APACrefauthors \ 2020 . Dragons, black swans and decisions. Dragons, black swans and decisions. Environmental research 183 109127, APACrefDOI doi:10.1016/j.envres.2020.109127 APACrefDOI

    arXiv 2020

  3. [3]

    \ Karasu, S

    Altan2022 APACrefauthors Altan, A. \ Karasu, S. APACrefauthors \ 2022 . Crude oil time series prediction model based on LSTM network with chaotic Henry gas solubility optimization Crude oil time series prediction model based on lstm network with chaotic henry gas solubility optimization . Energy 242 , APACrefDOI doi:10.1016/j.energy.2021.122964 APACrefDOI

    work page arXiv 2022

  4. [4]

    , Deisenroth, M.P

    Arulkumaran2017 APACrefauthors Arulkumaran, K. , Deisenroth, M.P. , Brundage, M. Bharath, A.A. APACrefauthors \ 2017 . Deep reinforcement learning: A brief survey Deep reinforcement learning: A brief survey \ ( 34)

    2017

  5. [5]

    APACrefauthors \ 2002

    Avramov2002 APACrefauthors Avramov, D. APACrefauthors \ 2002 . Stock return predictability and model uncertainty Stock return predictability and model uncertainty . Journal of Financial Economics 64 , APACrefDOI doi:10.1016/S0304-405X(02)00131-9 APACrefDOI

    work page doi:10.1016/s0304-405x(02)00131-9 2002

  6. [6]

    \ Puterman, M.L

    Baxter1995 APACrefauthors Baxter, L.A. \ Puterman, M.L. APACrefauthors \ 1995 . Markov Decision Processes: Discrete Stochastic Dynamic Programming Markov decision processes: Discrete stochastic dynamic programming . Technometrics 37 , APACrefDOI doi:10.2307/1269932 APACrefDOI

    work page doi:10.2307/1269932 1995

  7. [7]

    , Zoph, B

    2016Neural APACrefauthors Bello, I. , Zoph, B. , Vasudevan, V. Le, Q.V. APACrefauthors \ 2016 . Neural Optimizer Search with Reinforcement Learning Neural optimizer search with reinforcement learning

    2016

  8. [8]

    \ Courville, A.C

    2013Deep APACrefauthors Bengio, Y. \ Courville, A.C. APACrefauthors \ 2013 . Deep Learning of Representations Deep learning of representations . International Conference on Neural Information Processing. International conference on neural information processing

    2013

Show all 91 references
  1. [9]

    Bollerslev2020Good APACrefauthors Bollerslev, T. , Li, S. Zhao, B. APACrefauthors \ 2020 . Good Volatility, Bad Volatility, and the Cross Section of Stock Returns Good volatility, bad volatility, and the cross section of stock returns . Journal of Financial and Quantitative An...

    2020 doi

  2. [10]

    \ Sakemoto, R

    Byrne2021The APACrefauthors Byrne, J.P. \ Sakemoto, R. APACrefauthors \ 2021 . The conditional volatility premium on currency portfolios The conditional volatility premium on currency portfolios . Journal of International Financial Markets, Institutions and Money 74 101415, AP...

    2021

  3. [11]

    Cao2019Deep APACrefauthors Cao, C. , Li, D. Fair, I. APACrefauthors \ 2019 . Deep Learning-Based Decoding of Constrained Sequence Codes Deep learning-based decoding of constrained sequence codes . IEEE Journal on Selected Areas in Communications 37 2532-2543, APACrefDOI doi:10...

    2019

  4. [12]

    \ AYDIN, A.D

    Seyma2020 APACrefauthors CAVDAR, S.C. \ AYDIN, A.D. APACrefauthors \ 2020 . Hybrid Model Approach to the Complexity of Stock Trading Decisions in Turkey Hybrid model approach to the complexity of stock trading decisions in turkey . Journal of Asian Finance, Economics and Busin...

    2020 doi

  5. [13]

    Chang2019 APACrefauthors Chang, V. , Li, T. Zeng, Z. APACrefauthors \ 2019 . Towards an improved Adaboost algorithmic method for computational financial analysis Towards an improved adaboost algorithmic method for computational financial analysis . Journal of Parallel and Dist...

    2019 doi

  6. [14]

    , Elie, R

    Charpentier2020Reinforcement APACrefauthors Charpentier, A. , Elie, R. Remlinger, C. APACrefauthors \ 2020 . Reinforcement Learning in Economics and Finance Reinforcement learning in economics and finance . ArXiv abs/2003.10014 , APACrefDOI doi:10.1007/S10614-021-10119-4 APACrefDOI

    2020 arXiv

  7. [15]

    \ Shamma, J

    Chasparis2012Distributed APACrefauthors Chasparis, G.C. \ Shamma, J. APACrefauthors \ 2012 . Distributed Dynamic Reinforcement of Efficient Outcomes in Multiagent Coordination and Network Formation Distributed dynamic reinforcement of efficient outcomes in multiagent coordinat...

    2012 doi

  8. [16]

    , Gulcehre, C

    Chung2014 APACrefauthors Chung, J. , Gulcehre, C. , Cho, K. Bengio, Y. APACrefauthors \ 2014 . Empirical evaluation of gated recurrent neural networks on sequence modeling Empirical evaluation of gated recurrent neural networks on sequence modeling . NIPS 2014 Workshop on Deep...

    2014

  9. [17]

    , Sun, R

    Cui2022 APACrefauthors Cui, B. , Sun, R. Su, J. APACrefauthors \ 2022 . A Novel Deep Reinforcement Learning Strategy in Financial Portfolio Management A novel deep reinforcement learning strategy in financial portfolio management

    2022

  10. [18]

    , Yang, Z

    Dai2020 APACrefauthors Dai, Z. , Yang, Z. , Yang, Y. , Carbonell, J. , Le, Q.V. Salakhutdinov, R. APACrefauthors \ 2020 . Transformer-XL: Attentive language models beyond a fixed-length context Transformer-xl: Attentive language models beyond a fixed-length context

    2020

  11. [19]

    \ Niv, Y

    Dayan2008Reinforcement APACrefauthors Dayan, P. \ Niv, Y. APACrefauthors \ 2008 . Reinforcement learning: The Good, The Bad and The Ugly Reinforcement learning: The good, the bad and the ugly . Current Opinion in Neurobiology 18 185-196, APACrefDOI doi:10.1016/j.conb.2008.08.0...

    2008 doi

  12. [20]

    DJIA Dow Jones Industrial Average Dow jones industrial average . 2023 . APACrefURL https://us.spindices.com/indices/equity/dow-jones-industrial-average APACrefURL

    2023

  13. [21]

    \ Kristoufek, L

    Fil2020Pairs APACrefauthors Fil, M. \ Kristoufek, L. APACrefauthors \ 2020 . Pairs Trading in Cryptocurrency Markets Pairs trading in cryptocurrency markets . IEEE Access 8 172644-172651, APACrefDOI doi:10.1109/ACCESS.2020.3024619 APACrefDOI

    2020

  14. [22]

    \ Tzaferi, D

    Fousekis2021Returns APACrefauthors Fousekis, P. \ Tzaferi, D. APACrefauthors \ 2021 . Returns and volume: Frequency connectedness in cryptocurrency markets Returns and volume: Frequency connectedness in cryptocurrency markets . Economic Modelling 95 13-20, APACrefDOI doi:10.10...

    2021 doi

  15. [23]

    \ Zhang, W

    Gao2013 APACrefauthors Gao, L. \ Zhang, W. APACrefauthors \ 2013 . Weighted moving average passive aggressive algorithm for online portfolio selection Weighted moving average passive aggressive algorithm for online portfolio selection . ( 1)

    2013

  16. [24]

    Gao2022 APACrefauthors Gao, R. , Gu, F. , Sun, R. , Stefanidis, A. , Ren, X. Su, J. APACrefauthors \ 2022 . A Novel DenseNet-based Deep Reinforcement Framework for Portfolio Management A novel densenet-based deep reinforcement framework for portfolio management . ( 158-165)

    2022

  17. [25]

    , Gao, Y

    Gao2020 APACrefauthors Gao, Z. , Gao, Y. , Hu, Y. , Jiang, Z. Su, J. APACrefauthors \ 2020 . Application of Deep Q-Network in Portfolio Management Application of deep q-network in portfolio management

    2020

  18. [26]

    , Blei, D

    Gershman2010Context APACrefauthors Gershman, S. , Blei, D. Niv, Y. APACrefauthors \ 2010 . Context, learning, and extinction. Context, learning, and extinction. Psychological review 117 1 197-209, APACrefDOI doi:10.1037/a0017808 APACrefDOI

    2010 doi

  19. [27]

    , Pouget-Abadie, J

    Goodfellow2014 APACrefauthors Goodfellow, I.J. , Pouget-Abadie, J. , Mirza, M. , Xu, B. , Warde-Farley, D. , Ozair, S. Bengio, Y. APACrefauthors \ 2014 . Generative adversarial nets Generative adversarial nets . ( 3)

    2014

  20. [28]

    , Jiang, Z

    Gu2021 APACrefauthors Gu, F. , Jiang, Z. Su, J. APACrefauthors \ 2021 . Application of Features and Neural Network to Enhance the Performance of Deep Reinforcement Learning in Portfolio Management Application of features and neural network to enhance the performance of deep re...

    2021

  21. [29]

    \ Liu, X Y

    Guan2021 APACrefauthors Guan, M. \ Liu, X Y. APACrefauthors \ 2022 . Explainable Deep Reinforcement Learning for Portfolio Management: An Empirical Approach Explainable deep reinforcement learning for portfolio management: An empirical approach . Proceedings of the Second ACM ...

    2022

  22. [30]

    , Singal, G

    Gupta2021Deep APACrefauthors Gupta, S. , Singal, G. Garg, D. APACrefauthors \ 2021 . Deep Reinforcement Learning Techniques in Diversified Domains: A Survey Deep reinforcement learning techniques in diversified domains: A survey . Archives of Computational Methods in Engineeri...

    2021 doi

  23. [31]

    , Lugosi, G

    Laszlo2006 APACrefauthors Györfi, L. , Lugosi, G. Udina, F. APACrefauthors \ 2006 . Nonparametric kernel-based sequential investment strategies Nonparametric kernel-based sequential investment strategies . Mathematical Finance 16 , APACrefDOI doi:10.1111/j.1467-9965.2006.00274...

    2006

  24. [32]

    , Zhou, A

    2018Soft APACrefauthors Haarnoja, T. , Zhou, A. , Abbeel, P. Levine, S. APACrefauthors \ 2018 . Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a s...

    2018

  25. [33]

    , Zhao, Y

    Hu2021A APACrefauthors Hu, Z. , Zhao, Y. Khushi, M. APACrefauthors \ 2021 . A Survey of Forex and Stock Price Prediction Using Deep Learning A survey of forex and stock price prediction using deep learning . Applied System Innovation , APACrefDOI doi:10.3390/asi4010009 APACrefDOI

    2021 doi

  26. [34]

    , Zhou, J

    Huang2016 APACrefauthors Huang, D.J. , Zhou, J. , Li, B. , Hoi, S.C. Zhou, S. APACrefauthors \ 2016 . Robust Median Reversion Strategy for Online Portfolio Selection Robust median reversion strategy for online portfolio selection . IEEE Transactions on Knowledge and Data Engin...

    2016

  27. [35]

    APACrefauthors \ 2016

    Hung2016 APACrefauthors Hung, N.H. APACrefauthors \ 2016 . Various moving average convergence divergence trading strategies: A comparison Various moving average convergence divergence trading strategies: A comparison . Investment Management and Financial Innovations 13 , APACr...

    2016 doi

  28. [36]

    \ Liang, J

    Jiang2018 APACrefauthors Jiang, Z. \ Liang, J. APACrefauthors \ 2018 . Cryptocurrency portfolio management with deep reinforcement learning Cryptocurrency portfolio management with deep reinforcement learning . ( \ 2018-January)

    2018

  29. [37]

    Jiang2017 APACrefauthors Jiang, Z. , Xu, D. Liang, J. APACrefauthors \ 2017 6 . A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem A deep reinforcement learning framework for the financial portfolio management problem . APACrefDOI doi:10.485...

    2017

  30. [38]

    \ Martin, J.H

    1214993 APACrefauthors Jurafsky, D. \ Martin, J.H. APACrefauthors \ 2009 . Speech and Language Processing (2nd Edition) Speech and language processing (2nd edition) . USA Prentice-Hall, Inc

    2009

  31. [39]

    , Muñoz, M.A

    Kandanaarachchi2020 APACrefauthors Kandanaarachchi, S. , Muñoz, M.A. , Hyndman, R.J. Smith-Miles, K. APACrefauthors \ 2020 . On normalization and algorithm selection for unsupervised outlier detection On normalization and algorithm selection for unsupervised outlier detection ...

    2020 doi

  32. [40]

    \ Shadwick, W.F

    Keating2002 APACrefauthors Keating, C. \ Shadwick, W.F. APACrefauthors \ 2002 . An introduction to Omega An introduction to omega . The Finance Development Centre Ltd ,

    2002

  33. [41]

    , Yin, H

    Khader2021Learning APACrefauthors Khader, S.A. , Yin, H. , Falco, P. Kragic, D. APACrefauthors \ 2021 . Learning Deep Energy Shaping Policies for Stability-Guaranteed Manipulation Learning deep energy shaping policies for stability-guaranteed manipulation . IEEE Robotics and A...

    2021

  34. [42]

    APACrefauthors \ 2020

    Koziarski2020 APACrefauthors Koziarski, M. APACrefauthors \ 2020 . Radial-Based Undersampling for imbalanced data classification Radial-based undersampling for imbalanced data classification . Pattern Recognition 102 , APACrefDOI doi:10.1016/j.patcog.2020.107262 APACrefDOI

    2020

  35. [43]

    , Yadav, S

    Kumar2021A APACrefauthors Kumar, A. , Yadav, S. , Gupta, P. Mehlawat, M.K. APACrefauthors \ 2021 . A Credibilistic Multiobjective Multiperiod Efficient Portfolio Selection Approach Using Data Envelopment Analysis A credibilistic multiobjective multiperiod efficient portfolio s...

    2021

  36. [44]

    , Bengio, Y

    2015Deep APACrefauthors Lecun, Y. , Bengio, Y. Hinton, G. APACrefauthors \ 2015 . Deep learning Deep learning . Nature 521 7553 436,

    2015

  37. [45]

    , Littman, M.L

    Lehnert2020Reward-predictive APACrefauthors Lehnert, L. , Littman, M.L. Frank, M.J. APACrefauthors \ 2020 . Reward-predictive representations generalize across tasks in reinforcement learning Reward-predictive representations generalize across tasks in reinforcement learning ....

    2020 doi

  38. [46]

    \ Hoi, S.C

    Li2012 APACrefauthors Li, B. \ Hoi, S.C. APACrefauthors \ 2012 . On-line portfolio selection with moving average reversion On-line portfolio selection with moving average reversion . ( 1)

    2012

  39. [47]

    , Hoi, S.C

    Lii2011 APACrefauthors Li, B. , Hoi, S.C. Gopalkrishnan, V. APACrefauthors \ 2011 . CORN: Correlation-driven nonparametric learning approach for portfolio selection Corn: Correlation-driven nonparametric learning approach for portfolio selection . ACM Transactions on Intellige...

    2011

  40. [48]

    , Hoi, S.C

    Li2011 APACrefauthors Li, B. , Hoi, S.C. , Zhao, P. Gopalkrishnan, V. APACrefauthors \ 2011 . Confidence Weighted Mean Reversion strategy for on-line portfolio selection Confidence weighted mean reversion strategy for on-line portfolio selection . ( 15)

    2011

  41. [49]

    , Wang, J

    Li2018 APACrefauthors Li, B. , Wang, J. , Huang, D. Hoi, S.C. APACrefauthors \ 2018 . Transaction cost optimization for online portfolio selection Transaction cost optimization for online portfolio selection . Quantitative Finance 18 , APACrefDOI doi:10.1080/14697688.2017.1357...

    2018

  42. [50]

    , Zhao, P

    Lii2012 APACrefauthors Li, B. , Zhao, P. , Hoi, S.C. Gopalkrishnan, V. APACrefauthors \ 2012 . PAMR: Passive aggressive mean reversion strategy for portfolio selection Pamr: Passive aggressive mean reversion strategy for portfolio selection . Machine Learning 87 , APACrefDOI d...

    2012 doi

  43. [51]

    , Ventre, C

    Liu101145 APACrefauthors Liu, C. , Ventre, C. Polukarov, M. APACrefauthors \ 2022 . Synthetic Data Augmentation for Deep Reinforcement Learning in Financial Trading Synthetic data augmentation for deep reinforcement learning in financial trading . Proceedings of the Third ACM ...

    2022

  44. [52]

    , Yang, H

    Liu2021 APACrefauthors Liu, X Y. , Yang, H. , Chen, Q. , Zhang, R. , Yang, L. , Xiao, B. Wang, C. APACrefauthors \ 2021 . FinRL: A Deep Reinforcement Learning Library for Automated Stock Trading in Quantitative Finance Finrl: A deep reinforcement learning library for automated...

    2021 doi

  45. [53]

    APACrefauthors \ 2018

    2018Investor APACrefauthors Liu, Z.T.Y. APACrefauthors \ 2018 . Investor-Imitator: A Framework for Trading Knowledge Extraction Investor-imitator: A framework for trading knowledge extraction . SIGKDD explorations Udisk ,

    2018

  46. [54]

    , Harikrishnan, K.P

    M.2022Recurrence APACrefauthors M., K. , Harikrishnan, K.P. Ambika, G. APACrefauthors \ 2022 . Recurrence measures and transitions in stock market dynamics Recurrence measures and transitions in stock market dynamics . Physica A: Statistical Mechanics and its Applications , AP...

    2022

  47. [55]

    \ Martellini, L

    Maeso2020Maximizing APACrefauthors Maeso, J. \ Martellini, L. APACrefauthors \ 2020 . Maximizing an equity portfolio excess growth rate: a new form of smart beta strategy? Maximizing an equity portfolio excess growth rate: a new form of smart beta strategy? Quantitative Financ...

    2020

  48. [56]

    \ Shen, J

    Mousavi2021A APACrefauthors Mousavi, A. \ Shen, J. APACrefauthors \ 2021 . A penalty decomposition algorithm with greedy improvement for mean‐reverting portfolios with sparsity and volatility constraints A penalty decomposition algorithm with greedy improvement for mean‐revert...

    2021 doi

  49. [57]

    APACrefauthors \ 1999

    Murphy1999 APACrefauthors Murphy, J.J. APACrefauthors \ 1999 . Technical analysis of the financial markets Technical analysis of the financial markets \ ( 77)

    1999

  50. [58]

    , Rizvi, S.T.H

    Naeem2020A APACrefauthors Naeem, M. , Rizvi, S.T.H. Coronato, A. APACrefauthors \ 2020 . A Gentle Introduction to Reinforcement Learning and its Application in Different Fields A gentle introduction to reinforcement learning and its application in different fields . IEEE Acces...

    2020

  51. [59]

    \ Soliman, A.M

    Nasir2014Aspects APACrefauthors Nasir, M. \ Soliman, A.M. APACrefauthors \ 2014 . Aspects of Macroeconomic Policy Combinations and Their Effects on Financial Markets Aspects of macroeconomic policy combinations and their effects on financial markets . Banking& Insurance eJournal ,

    2014

  52. [60]

    , Zoha, A

    Onireti2016 APACrefauthors Onireti, O. , Zoha, A. , Moysen, J. , Imran, A. , Giupponi, L. , Imran, M.A. Abu-Dayya, A. APACrefauthors \ 2016 . A cell outage management framework for dense heterogeneous networks A cell outage management framework for dense heterogeneous networks...

    2016

  53. [61]

    APACrefauthors \ 2020

    Parisi2020Reinforcement APACrefauthors Parisi, S. APACrefauthors \ 2020 . Reinforcement Learning with Sparse and Multiple Rewards Reinforcement learning with sparse and multiple rewards . APACrefDOI doi:10.25534/TUPRINTS-00011372 APACrefDOI

    2020 doi

  54. [62]

    , James, H.J

    2018CONTINUOUS APACrefauthors Paul, L.T. , James, H.J. , David, S. , Tom, E. , Yuval, T. , Otto, H.N.M. Alexander, P. APACrefauthors \ 2018 . CONTINUOUS CONTROL WITH DEEP REINFORCEMENT LEARNING Continuous control with deep reinforcement learning

    2018

  55. [63]

    , Gómez-Martínez, R

    Paule-Vianez2020A APACrefauthors Paule-Vianez, J. , Gómez-Martínez, R. Prado‐Román, C. APACrefauthors \ 2020 . A bibliometric analysis of behavioural finance with mapping analysis tools A bibliometric analysis of behavioural finance with mapping analysis tools . European Resea...

    2020 doi

  56. [64]

    \ Granger, C.W

    Poon2003 APACrefauthors Poon, S.H. \ Granger, C.W. APACrefauthors \ 2003 . Forecasting volatility in financial markets: A review Forecasting volatility in financial markets: A review \ ( 41)

    2003

  57. [65]

    Qin2022 APACrefauthors Qin, Y. , Gu, F. Su, J. APACrefauthors \ 2022 . A Novel Deep Reinforcement Learning Strategy for Portfolio Management A novel deep reinforcement learning strategy for portfolio management

    2022

  58. [66]

    , Jiang, Z

    Ren2021 APACrefauthors Ren, X. , Jiang, Z. Su, J. APACrefauthors \ 2021 . The Use of Features to Enhance the Capability of Deep Reinforcement Learning for Investment Portfolio Management The use of features to enhance the capability of deep reinforcement learning for investmen...

    2021

  59. [67]

    \ Hoffman, S.T

    Rollinger2015 APACrefauthors Rollinger, T.N. \ Hoffman, S.T. APACrefauthors \ 2015 . Sortino A Sharper Ratio Sortino a sharper ratio . Red Rock Capital ,

    2015

  60. [68]

    , Wolski, F

    Schulman2017 APACrefauthors Schulman, J. , Wolski, F. , Dhariwal, P. , Radford, A. Klimov, O. APACrefauthors \ 2017 . Proximal Policy Optimization Algorithms Proximal policy optimization algorithms . APACrefDOI doi:10.48550 APACrefDOI

    2017

  61. [69]

    , Aloui, C

    Sharif2020COVID-19 APACrefauthors Sharif, A. , Aloui, C. Yarovaya, L. APACrefauthors \ 2020 . COVID-19 pandemic, oil prices, stock market, geopolitical risk and policy uncertainty nexus in the US economy: Fresh evidence from the wavelet-based approach Covid-19 pandemic, oil pr...

    2020

  62. [70]

    APACrefauthors \ 1994

    Sharpe1994 APACrefauthors Sharpe, W.F. APACrefauthors \ 1994 . The Sharpe Ratio The sharpe ratio . The Journal of Portfolio Management 21 , APACrefDOI doi:10.3905/jpm.1994.409501 APACrefDOI

    1994

  63. [71]

    APACrefauthors \ 2014

    Singleton2014 APACrefauthors Singleton, K.J. APACrefauthors \ 2014 . Investor flows and the 2008 boom/bust in oil prices Investor flows and the 2008 boom/bust in oil prices . Management Science 60 , APACrefDOI doi:10.1287/mnsc.2013.1756 APACrefDOI

    2014

  64. [72]

    \ Paquet, E

    Soleymani2020 APACrefauthors Soleymani, F. \ Paquet, E. APACrefauthors \ 2020 . Financial portfolio optimization with online deep reinforcement learning and restricted stacked autoencoder—DeepBreath Financial portfolio optimization with online deep reinforcement learning and r...

    2020

  65. [73]

    , Jiang, Y

    Song2019Observational APACrefauthors Song, X. , Jiang, Y. , Tu, S. , Du, Y. Neyshabur, B. APACrefauthors \ 2019 . Observational Overfitting in Reinforcement Learning Observational overfitting in reinforcement learning . ArXiv abs/1912.02975 ,

    2019 arXiv

  66. [74]

    , Jiang, Z

    Sun2021 APACrefauthors Sun, R. , Jiang, Z. Su, J. APACrefauthors \ 2021 . A Deep Residual Shrinkage Neural Network-based Deep Reinforcement Learning Strategy in Financial Portfolio Management A deep residual shrinkage neural network-based deep reinforcement learning strategy i...

    2021

  67. [75]

    , Qin, M

    sun2023trademaster APACrefauthors Sun, S. , Qin, M. , Xia, H. , Zong, C. , Ying, J. , Xie, Y. others APACrefauthors \ 2023 . TradeMaster: A Holistic Quantitative Trading Platform Empowered by Reinforcement Learning Trademaster: A holistic quantitative trading platform empowere...

    2023

  68. [76]

    \ Barto, A.G

    1998Reinforcement APACrefauthors Sutton, R.S. \ Barto, A.G. APACrefauthors \ 1998 . Reinforcement Learning Reinforcement learning . A Bradford Book volume 15 7 665-685,

    1998

  69. [77]

    \ Kitajima, T

    Takada2022Trend-following APACrefauthors Takada, T. \ Kitajima, T. APACrefauthors \ 2022 . Trend-following with better adaptation to large downside risks Trend-following with better adaptation to large downside risks . PLoS ONE 17 , APACrefDOI doi:10.1371/journal.pone.0276322 ...

    2022 doi

  70. [78]

    , Vedaldi, A

    Ulyanov2014 APACrefauthors Ulyanov, D. , Vedaldi, A. Lempitsky, V. APACrefauthors \ 2014 . Instance Normalization: The Missing Ingredient for Fast Stylization Dmitry Instance normalization: The missing ingredient for fast stylization dmitry . Physical Review D - Particles, Fie...

    2014

  71. [79]

    , Adekoya, O

    Umar2021Media APACrefauthors Umar, Z. , Adekoya, O. , Oliyide, J. Gubareva, M. APACrefauthors \ 2021 . Media sentiment and short stocks performance during a systemic crisis Media sentiment and short stocks performance during a systemic crisis . International Review of Financia...

    2021

  72. [80]

    , Shazeer, N

    Vaswani2017 APACrefauthors Vaswani, A. , Shazeer, N. , Parmar, N. , Uszkoreit, J. , Jones, L. , Gomez, A.N. Polosukhin, I. APACrefauthors \ 2017 . Attention is all you need Attention is all you need . ( \ 2017-December)

    2017

  73. [81]

    \ Nagy, Z

    Jos2019 APACrefauthors Vázquez-Canteli, J.R. \ Nagy, Z. APACrefauthors \ 2019 . Reinforcement learning for demand response: A review of algorithms and modeling techniques Reinforcement learning for demand response: A review of algorithms and modeling techniques \ ( 235)

    2019

  74. [82]

    , Yan, X.S

    Wang2020Shorting APACrefauthors Wang, X. , Yan, X.S. Zheng, L. APACrefauthors \ 2020 . Shorting flows, public disclosure, and market efficiency Shorting flows, public disclosure, and market efficiency . Journal of Financial Economics 135 191-212, APACrefDOI doi:10.1016/J.JFINE...

    2020 doi

  75. [83]

    , Tanner, B

    Whiteson2011Protecting APACrefauthors Whiteson, S. , Tanner, B. , Taylor, M.E. Stone, P. APACrefauthors \ 2011 . Protecting against evaluation overfitting in empirical reinforcement learning Protecting against evaluation overfitting in empirical reinforcement learning . 2011 I...

    2011

  76. [84]

    , Yin, H

    Xia2020Fully APACrefauthors Xia, W. , Yin, H. , Dai, X. Jha, N. APACrefauthors \ 2020 . Fully Dynamic Inference With Deep Neural Networks Fully dynamic inference with deep neural networks . IEEE Transactions on Emerging Topics in Computing 10 962-972, APACrefDOI doi:10.1109/TE...

    2020

  77. [85]

    , Liu, X Y

    Yang2020 APACrefauthors Yang, H. , Liu, X Y. , Zhong, S. Walid, A. APACrefauthors \ 2020 . Deep Reinforcement Learning for Automated Stock Trading: An Ensemble Strategy Deep reinforcement learning for automated stock trading: An ensemble strategy . SSRN Electronic Journal , AP...

    2020 doi

  78. [86]

    , Sun, R

    Yang2022 APACrefauthors Yang, X. , Sun, R. , Ren, X. , Stefanidis, A. , Gu, F. Su, J. APACrefauthors \ 2022 . Ghost Expectation Point with Deep Reinforcement Learning in Financial Portfolio Management Ghost expectation point with deep reinforcement learning in financial portfo...

    2022

  79. [87]

    , Ren, X

    Yao2022 APACrefauthors Yao, W. , Ren, X. Su, J. APACrefauthors \ 2022 . An Inception Network with Bottleneck Attention Module for Deep Reinforcement Learning Framework in Financial Portfolio Management An inception network with bottleneck attention module for deep reinforcemen...

    2022

  80. [88]

    , Pei, H

    Ye2020ReinforcementLearningBP APACrefauthors Ye, Y. , Pei, H. , Wang, B. , Chen, P Y. , Zhu, Y. , Xiao, J. Li, B. APACrefauthors \ 2020 . Reinforcement-Learning based Portfolio Management with Augmented Asset Movement Prediction States Reinforcement-learning based portfolio ma...

    2020 arXiv

  81. [89]

    , Ren, X

    Zhang2022 APACrefauthors Zhang, R. , Ren, X. , Gu, F. , Stefanidis, A. , Sun, R. Su, J. APACrefauthors \ 2022 . MDAEN: Multi-Dimensional Attention-based Ensemble Network in Deep Reinforcement Learning Framework for Portfolio Management Mdaen: Multi-dimensional attention-based ...

    2022

  82. [90]

    , Chen, Q

    Zhang2021A APACrefauthors Zhang, W. , Chen, Q. , Yan, J. , Zhang, S. Xu, J. APACrefauthors \ 2021 . A novel asynchronous deep reinforcement learning model with adaptive early forecasting method and reward incentive mechanism for short-term load forecasting A novel asynchronous...

    2021

  83. [91]

    \ Rosendo, A

    Zhu2021A APACrefauthors Zhu, W. \ Rosendo, A. APACrefauthors \ 2021 . A Functional Clipping Approach for Policy Optimization Algorithms A functional clipping approach for policy optimization algorithms . IEEE Access 9 96056-96063, APACrefDOI doi:10.1109/ACCESS.2021.3094566 APACrefDOI

    2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.