MAPLE: Efficient and Diverse Multi-Alpha Generation for Portfolio Construction

Kendro Vincent; Kuan-Yu Chen; Tien-Hao Chang; Yu-Chen Den

REVIEW 4 major objections 5 minor 51 references

MAPLE: Efficient and Diverse Multi-Alpha Generation for Portfolio Construction

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A single training pass can produce a diverse set of stock-ranking alphas that beats nine baselines on risk-adjusted return.

desk verdict A genuinely useful multi-alpha framework, but the headline edge rests on hyperparameters whose selection procedure is not documented; worth a serious look with revision. read the letter →

arxiv 2607.24131 v1 pith:NLWI3HBD submitted 2026-07-27 cs.LG cs.CE

Yu-Chen Den , Kuan-Yu Chen , Kendro Vincent , Tien-Hao Chang This is my paper · ORCID

classification cs.LGcs.CE

keywords multi-alphagenerationportfolioconstructionlearningtorankalphadiversitySharperatioCalmarcapacityscalingstockranking

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 the classic quant principle—combine many low-correlated signals instead of one—can be engineered into a single deep ranking model, eliminating the need for separately trained ensemble members or multi-stage pipelines. MAPLE writes $N_\alpha$ $\alpha$ scores per stock in one forward pass using a capacity-scaled prediction head, then trains them with an extreme-rank weighted listwise loss plus an explicit penalty on pairwise $\alpha$ correlation. Across four equity markets (US, China, Japan), the equal-weighted portfolio of 24 alphas reports the best average risk-adjusted return among nine baselines—average Sharpe 1.690 and Calmar 2.175—at roughly 186K parameters and 3.6 seconds per epoch. The paper further reports that the gains transfer across five backbone architectures (10–23% Sharpe, 17–43% Calmar) and presents behavioral evidence about why each component works. If the claim holds, a practitioner can obtain diversified multi-signal portfolios from a single model instead of from expensive model ensembles.

What carries the argument

The load-bearing object is the unified multi-$\alpha$ prediction head, which replaces the standard linear output layer after a stock encoder. It produces $\hat{\boldsymbol{Y}}=\hat{\boldsymbol{Y}}_{\text{intra}}+\hat{\boldsymbol{Y}}_{\text{inter}}\in\mathbb{R}^{S\times N_\alpha}$, where the intra path is a two-layer MLP with scaled hidden width and the inter path is multi-head attention over the stock axis with one head per $\alpha$, using a value projection that directly emits the per-stock score. The companion mechanism is the training objective: $\mathcal{L}=\mathcal{L}_{\text{spearman}}+\mathcal{L}_{\text{extreme}}+\lambda\mathcal{L}_{\text{diversity}}$, where the extreme term uses a sigmoidal position weight $v_s^{(i)}$ driven by per-$\alpha$ sharpness $\xi^{(i)}$ and margin $\delta^{(i)}$ to concentrate weight on top-ranked names, and the diversity term penalizes the absolute Spearman correlation between $\alpha$ pairs in rank-surrogate space. The work these pieces do is to redirect gradient flow between the two paths, decorrelate the alphas, and preserve per-$\alpha$ ranking accuracy, which the equal-weighted portfolio aggregation $p_t=\frac{1}{N_\alpha}\sum_m p_t^{(m)}$ then converts into ensemble-level risk-adjusted return.

What would settle it

Re-run the comparison on a hold-out period after 2024 with $\lambda$ selected on the validation year only, with each baseline given per-market hyperparameter tuning, and with intraday-2008 index constituents used to define each universe; if MAPLE's average Sharpe and Calmar gaps against the best baselines shrink to near zero or reverse, the central claim is not supported.

Watch

Extended reading notes

Core claim

MAPLE's central discovery is that $\alpha$ diversity can be manufactured inside a single model by choosing where the learning signal is concentrated and how much capacity each $\alpha$ gets. The head sums an intra-stock MLP and an inter-stock attention path (one attention head per $\alpha$) directly in prediction space, with hidden widths that grow with $N_\alpha$ so that additional alphas do not starve one another. The loss adds a global Spearman term, an extreme-rank weighted Spearman term whose per-$\alpha$ learnable sharpness $\xi^{(i)}$ and margin $\delta^{(i)}$ focus each $\alpha$ on the true top-ranked stocks, and a diversity regularizer $\lambda\sum_{i\neq j}|\rho(\phi(\hat{\boldsymbol{y}}_i),\phi(\hat{\boldsymbol{y}}_j))|$ that discourages redundant $\alpha$ pairs. In the paper's evaluation, the equal-weighted portfolio of the resulting alphas achieves the best average Sharpe ratio (1.690) and Calmar ratio (2.175) among nine baselines on CSI300, CSI500, NI225, and SP500, using fewer parameters than the leading baselines and roughly 2–2.5x less training time. The paper also demonstrates that the same head and losses generalize across TCN, GRU, LSTM, Mamba, and Transformer backbones, and that a single model matches the risk-adjusted performance of an explicit multi-seed ensemble while producing all $N_\alpha$ alphas in one forward pass.

Load-bearing premise

The load-bearing premise is that the comparison is fair out-of-sample: the diversity weight, alpha count, and design choices were not tuned on the 2021–2024 test window, the nine baselines were given competitive hyperparameters, and the stock universes are free of survivorship bias.

Editorial extensions

If this is right

A practitioner who already has a trained stock encoder can attach the MAPLE head and losses to it; the paper reports consistent Sharpe and Calmar gains (10–23% and 17–43% on average) across five different temporal backbones.
The diversity–quality trade-off is controllable through $\lambda$: risk-adjusted return peaks near $\lambda\in[0.1,0.15]$ and falls for larger $\lambda$, so the method does not require a delicate or luck-dependent setting to obtain its diversification benefit.
The explicit comparison to a multi-seed ensemble implies that the common practice of training several models with different random seeds and averaging their outputs can be replaced by one MAPLE model at a fraction of the training cost, without sacrificing risk-adjusted performance.
Because the head operates on the final hidden representation of any backbone, the framework transfers to new markets or new feature sets without architecture redesign, as long as the input is a cross-sectional panel of stocks-by-features-by-timesteps.

Reading between the lines

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

Editorial inference: if the result survives out-of-sample, the same recipe—capacity-scaled multi-head prediction plus a correlation-penalizing loss—could replace ensembles in other listwise ranking applications such as recommendation or document retrieval, where the goal is also to produce several diverse orderings of the same items.
Editorial inference: the paper's sensitivity curves suggest a practical rule—run an unregularized multi-seed ensemble once to read off its natural alpha correlation, then tune $\lambda$ so that the single model's correlation approaches but does not go below that level; overshooting sacrifices ranking quality without extra return.
Editorial inference: since the authors state the diversity regularizer operates within one prediction target, asset class, and feature set, the most direct stress test is to widen one of those axes (e.g., multiple horizons or multiple asset classes) and check whether the regularizer still produces nonredundant alphas or whether diversity must be re-sourced from the auxiliary inputs.
Editorial inference: the paper's gradient-ratio analysis raises a testable diagnostic—monitor the inter/intra gradient ratio on a new market during training; a model that converges to a balanced ratio (near 1:1) should yield more diversified alphas, so the ratio could serve as an early indicator of whether the diversity regularizer is working before any test-set backtest.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Yu-Chen Den et al. have a real result: adding a capacity-scaled multi-alpha head, an extreme-rank weighted Spearman loss, and an explicit pairwise-correlation regularizer yields a single-model ensemble that beats nine baselines on average Sharpe and Calmar across four markets. The ablations actually support the claim that each piece does something. The behavioral analysis in Section 5 is better than usual - the path-dominance R^2 and gradient-ratio evidence, and the risk-return decomposition in Appendix D.2, give a concrete mechanism rather than a hand-wave. The authors also report transaction-cost numbers and concede the diversity mechanism is bounded to one target, asset class, and feature set. That is honest. Algorithm 1 and the loss definitions are explicit enough to reproduce from the text. The paper earns a serious referee. The soft spots are the ones the stress-test flags, and they are real. The headline lambda=0.1 and N_alpha=24 sit at the peak of the Section 5.4 sensitivity curves, computed on what looks like the same 2021-2024 window used for Table 1. The paper never says these were fixed on the 2020 validation set. If they were chosen after inspecting test-period curves, the average SR gap of 0.236 and CR gap of 0.241 are optimistically biased. This is the weakest load-bearing point. Related: five seeds are averaged but no error bars or significance tests are shown; on NI225 the Calmar gap to TIPS is 0.001, so 'beats every baseline on average' is doing a lot of work. Baselines use published hyperparameters without per-market tuning while MAPLE components were designed on these datasets; that asymmetry is acknowledged but not addressed. Appendix A.1 doesn't specify how the 295/514/209/525-stock universes were filtered from the indices, leaving survivorship-bias risk unquantified. No code or data is released. None of this is fatal; all of it is curable. The central architecture and loss design are sound, and the mechanism story is the strongest part. The paper needs a validation-only lambda selection rule, error bars or paired tests, a same-breadth baseline, and code/data release before I'd trust the exact gap. But the framework is worth a serious referee.

Referee Report

4 major / 5 minor

Summary. This paper proposes MAPLE, a backbone-agnostic framework that generates N_alpha stock-ranking signals in a single forward pass using a unified prediction head (intra-stock MLP plus per-alpha inter-stock attention), an extreme-rank weighted Spearman loss, and a pairwise-correlation diversity regularizer. The authors report experiments on four equity markets (CSI300, CSI500, NI225, SP500) over 2021-2024, claiming that MAPLE achieves the best average Sharpe (1.690) and Calmar (2.175) ratios among nine baselines with 186K parameters and 3.60s per epoch, that it generalizes across five backbones with 10-23% SR and 17-43% CR gains, and that ablations and behavioral analyses attribute the gains to diversity and capacity scaling rather than architectural complexity.

Significance. If the empirical claims hold, the contribution is useful: MAPLE demonstrates that a single model with explicit diversity control and capacity scaling can replace separate multi-model ensembles for multi-alpha generation, and the prediction head is backbone-agnostic, which is a practical advantage for portfolio construction pipelines. The paper does not release code, but the algorithmic descriptions, training objective, and evaluation protocol are detailed enough to reimplement. The strongest parts are the component-wise ablations and the gradient-path analysis, which give mechanistic evidence rather than only aggregate metrics. The significance is conditional on the evaluation being free of test-set hyperparameter selection and on the baseline comparison being fair; the average SR gap of 0.236 and CR gap of 0.241 over the best baseline are modest enough that these issues could change the conclusion.

major comments (4)

[§5.4 / Fig. 5 / Table 10] The headline configuration lambda=0.1 and N_alpha=24 sits at or near the peak of the sensitivity curves reported in Fig. 5 and Table 10, and those curves appear to be computed on the same 2021-2024 test window used for Table 1. For alpha scaling, Table 10 shows a CR of 2.175 at both lambda=0.10 and lambda=0.15, and the paper nowhere states that lambda and N_alpha were fixed on the 2020 validation set rather than chosen after inspecting these curves. Because the central claim is the average SR/CR gap over the nine baselines, this is a test-set selection concern that could materially shrink the reported edge. Please document the model-selection procedure (e.g., selecting lambda and N_alpha on 2020 and re-running 2021-2024 as a purely confirmatory window), or report the main result across lambda in a neighborhood of 0.1, with the selection rule stated.
[§4.1 / §A.3 / Table 1] All headline numbers are point estimates averaged over five seeds, with no standard deviations, confidence intervals, or significance tests. The NI225 Calmar gap against TIPS is 0.001 (0.783 vs 0.784), and without an error bar it is impossible to tell whether even the unselected average gap of 0.236 SR / 0.241 CR is reliably nonzero. Please report per-seed ranges or confidence intervals, and at least a paired test for the average SR/CR comparison.
[§A.2 / §A.3] Baselines are run with their original published hyperparameters without per-market tuning, while MAPLE's hyperparameters (including lambda and N_alpha) are reported as fixed choices. This asymmetry is a fairness risk for the headline comparison. At minimum, please state explicitly that no MAPLE hyperparameter was selected based on the 2021-2024 test window, and ideally include a matched comparison where a reasonable baseline (e.g., the vanilla single-alpha Transformer) receives validation-based tuning on the same 2020 set.
[§A.1] The construction of the four universes (295, 514, 209, 525 stocks) is not specified: the text does not state how constituents are filtered from the indices, when the constituent lists are snapped, or how delisted stocks are handled. Without this, survivorship-bias risk in the constructed universes is unaddressed. Please provide the exact universe construction and a statement on survivorship bias.

minor comments (5)

[§5.4] The text says that for the alpha-scaled configuration 'CR plateaus,' but Fig. 5 shows a clear rise-then-decline shape; please rephrase to describe the peak and subsequent decline.
[Algorithm 1] The comment 'r=0 for top stock' and the use of d = clip(c - r, min=0) should be clarified, since d=0 for both the center and bottom ranks; a short note on why the bottom region receives zero weight would help.
[§3.2] Equation (7) and Algorithm 1 compute a weighted dot product after demeaning and L2-normalization but call it a Spearman correlation; the relationship to the usual rank correlation should be made precise.
[Appendix] The paper does not mention code or data release; a reproducibility appendix with the exact data preprocessing steps and random seeds would strengthen the empirical claims.
[Appendix A.4] The evaluation protocol produces W rebalancing-phase return streams and averages metrics across phases; please make this explicit in the main text, since it affects comparability with baselines that may use a different protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline SR/CR claim is an external held-out benchmark comparison, and the self-citations are not load-bearing.

full rationale

The derivation chain runs from a defined loss objective (Eqs. 6-9) to an inference-time equal-weighted portfolio (Eq. 10), and the load-bearing evidence is the 2021-2024 test-window comparison against nine external baselines (Table 1). Nothing in that chain defines the reported Sharpe or Calmar ratios in terms of MAPLE's own training objective; the portfolio returns come from realized stock returns and top-k selections, not from the loss values. The only self-citations are TIPS (reference [7]) as a baseline and as part of the dataset-construction protocol; neither supplies a premise that the paper's own equations presuppose, and the TIPS numbers are re-computed in this paper rather than imported as a theorem. The behavioral sections measure quantities that the losses directly target (inter-alpha correlation, top-rank precision), but this is objective-aligned validation rather than a prediction derived from its own inputs. The conclusion's limitation statement that the diversity regularizer decorrelates alphas within a single target, asset class, and feature set is a scope caveat, not a circular step, and Appendix B explicitly labels the coverage correction as heuristic rather than an exact normalization. A residual concern that lambda=0.1 coincides with the peak of a sensitivity curve computed on the same test window (Fig. 5 and Table 10) is an evaluation-selection risk, not a definitional reduction, because no passage states that lambda was fitted on the test set.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The paper is an empirical ML study, so its ledger is dominated by domain assumptions about data, evaluation, and baseline fairness rather than mathematical axioms. The principal free parameters are the diversity weight lambda, the alpha count, the position-profile parameters, the top-k basket size, the prediction horizon, the rank-surrogate constant, the capacity-scaling rule, and the heuristic coverage correction. The main domain assumptions are that the four stock universes are tradable and bias-free, that the portfolio protocol captures realized strategy performance, that score decorrelation transfers to return diversification, and that untuned baselines provide a fair bar. No new entities (particles, forces, mediators) are introduced.

free parameters (8)

Diversity weight lambda = 0.1
Set in Section 3.3; Section 5.4 shows the Calmar ratio peaks at lambda in [0.1, 0.15] on the same-looking evaluation window, so the chosen value sits at the measured peak.
Number of alphas N_alpha = 24
Final configuration chosen after scaling experiments (Sections 4.3 and 5.3); performance varies materially with N_alpha (Tables 2 and 9).
Initial sharpness xi and margin gamma = xi=10, gamma=0.8
Hand-set initial values for the learnable per-alpha position-profile parameters (Table 6); they shape where each alpha's extreme-rank weighting acts during training.
Top-k basket size = 5 stocks
Each alpha selects its top-5 stocks (Appendix A.4); the ensemble's effective breadth (union size about 10 in Table 4) depends on this choice.
Prediction horizon and holding period q = 5 trading days
Labels are forward 5-day returns (Appendix A.1) and the holding or rebalancing interval is W (Appendix A.4).
Rank surrogate scale constant = 1.83
Fixed coefficient in the sigmoid rank surrogate in Algorithm 1; adopted from prior work [11] rather than derived.
Capacity scaling coefficients = intra MLP hidden floor(D*N_alpha/8); attention embed floor(2*D*N_alpha/8)
Hand-chosen scaling rule (Section 3.1) that keeps per-head key dimension approximately constant as N_alpha grows.
Coverage correction C_i = 1 + 2*delta_i/S = computed from learned delta_i
Heuristic gradient-magnitude compensation (Eq. 16, Appendix B), admitted by the authors as 'heuristic rather than an exact normalization'.

assumptions (7)

domain assumption Five-day forward cross-sectional returns are predictable from OHLCV and moving-average features.
The whole ranking task presumes learnable signal in the labels; standard in this literature but not verified independently here. Enters at Appendix A.1 label definition.
domain assumption The stock universes (295, 514, 209, 525 names) faithfully represent the CSI300, CSI500, NI225, and SP500 without survivorship or lookahead bias.
Appendix A.1 specifies features and labels but not constituent selection or filtering; this is load-bearing for all reported Sharpe and Calmar numbers.
domain assumption The portfolio protocol (top-5 softmax-weighted baskets per alpha, equal weighting across alphas, W rebalancing phases) is a valid proxy for realized strategy performance.
Defined in Appendix A.4; the protocol also increases portfolio breadth relative to single-alpha baselines, which mechanically lowers volatility.
domain assumption Decorrelating predicted alpha scores transfers to diversification of realized portfolio returns.
The regularizer in Eq. (8) penalizes predicted-score correlation; Section 5.2 uses overlap and downside correlation as proxies, but a share of the Sharpe gain may come from holding a wider set of stocks.
domain assumption Baseline hyperparameters from the original papers remain appropriate for these four datasets and this train/validation/test split.
Appendix A.2 states baselines are not tuned on these datasets, while MAPLE's lambda, N_alpha, and capacity rule were developed with knowledge of this evaluation.
standard math The differentiable rank surrogate (demeaned, L2-normalized sigmoid) preserves Spearman-like ranking gradients.
Used in Eqs. (6)-(8) and Algorithm 1, borrowed from prior work [11]; no derivation is given, which is appropriate for a borrowed tool.
standard math Standard deep learning machinery (LayerNorm, softmax attention, AdamW, gradient clipping) behaves as expected.
Background assumptions for all models; unremarkable but relied upon.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MAPLE: Efficient and Diverse Multi-Alpha Generation for Portfolio Construction." pith.science (2026). https://pith.science/paper/NLWI3HBD

@misc{pith2026260724131,
  author       = {Pith},
  title        = {Pith review of: MAPLE: Efficient and Diverse Multi-Alpha Generation for Portfolio Construction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NLWI3HBD}},
  note         = {Machine review of arXiv:2607.24131}
}

read the original abstract

Classical alpha mining achieves strong risk-adjusted returns by combining many low-correlated predictive signals, yet deep learning stock-ranking methods typically produce a single alpha per stock, rely on increasingly complex architectures with diminishing gains, and obtain diversity only through separate models or implicit routing, without explicitly controlling inter-alpha correlation. We introduce MAPLE (Multi-Alpha Position-aware Listwise Ensembling), a backbone-agnostic framework that recovers this diversity principle within a single training pass. MAPLE combines a unified, capacity-scaled prediction head with an extreme-rank weighted listwise ranking loss and a diversity regularizer that explicitly penalizes pairwise correlation across alphas. Across four equity markets spanning the US, China, and Japan, MAPLE achieves the best average Sharpe and Calmar ratios among nine baselines, using up to 55x fewer parameters and 2.5x less training time, and generalizes across five backbone architectures with Sharpe and Calmar Ratio gains of 10-23% and 17-43%, respectively. Behavioral analysis further shows why each component works: the unified head already reduces inter-alpha correlation before any diversity loss is applied, and the extreme-rank loss lets diversity regularization improve rather than erode per-alpha ranking quality as capacity scaling sustains this balance at scale. These results show that principled loss design and capacity allocation, rather than architectural complexity, drive diverse and effective multi-alpha generation.

Figures

Figures reproduced from arXiv: 2607.24131 by the authors.

**Figure 2.** Position-aware weight 𝑣𝑠 shown across different sharpness 𝜉 and scale 𝛾 values. Larger 𝜉 produces a sharper transition, while larger 𝛾 shifts the margin 𝛿 further from the center, narrowing the range of top-ranked stocks that receive nonzero weight. Extreme-Rank Weighted Spearman Loss. To resolve this conflict, we concentrate each alpha’s learning signal on the extremerank positions, so that diversity regularizati… view at source ↗

**Figure 3.** Average Sharpe Ratio (left) and Calmar Ratio (right) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Annual Return (left) and Sharpe Ratio (right) across [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 6.** Figure 6: Gradient trajectory of different attention structure [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

**Figure 7.** Figure 7: Ablation of the extreme loss ℒext and the diversity regularizer ℒdiv. Each term isolates one side of the risk– return trade-off: ℒext improves mean return but inflates volatility, while ℒdiv reverses both. Their combination attains the highest risk-adjusted score, in…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 34 canonical work pages

[1]

Viral V Acharya, Lasse H Pedersen, Thomas Philippon, and Matthew Richardson

work page
[2]

Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normaliza- tion.arXiv preprint arXiv:1607.06450(2016)

arXiv 2016
[3]

Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007. Learning to rank: from pairwise approach to listwise approach. InInternational Conference on Machine Learning (ICML)(Corvalis, Oregon, USA)(ICML ’07). Association for Computing Machinery, New York, NY, USA, 129–136. doi:10.1145/1273496. 1273513

doi:10.1145/1273496 2007
[4]

Zhiyuan Cao, Jiayu Xu, Chengqi Dong, Peiwen Yu, and Tian Bai. 2024. MATCC: A Novel Approach for Robust Stock Price Prediction Incorporating Market Trends and Cross-time Correlations. InACM International Conference on Information and Knowledge Management (CIKM). 187–196

work page 2024
[5]

Weijun Chen and Yanze Wang. 2025. DHMoE: Diffusion Generated Hierarchi- cal Multi-Granular Expertise for Stock Prediction. InConference on Artificial Intelligence (AAAI), Vol. 39. 11490–11499

work page 2025
[6]

Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555(2014)

arXiv 2014
[7]

Yu-Chen Den, Kuan-Yu Chen, Kendro Vincent, and Darby Tien-Hao Chang. 2026. Integrating Inductive Biases in Transformers via Distillation for Financial Time Series Forecasting.arXiv preprint arXiv:2603.16985(2026)

work page Pith review arXiv 2026
[8]

Eugene F Fama and Kenneth R French. 1992. The cross-section of expected stock returns.Journal of Finance47, 2 (1992), 427–465

work page 1992

Show all 51 references

[9]

Eugene F Fama and Kenneth R French. 2018. Choosing factors.Journal of Financial Economics128, 2 (2018), 234–252

2018
[10]

Jinyong Fan and Yanyan Shen. 2024. StockMixer: a simple yet strong MLP-based architecture for stock price forecasting. InConference on Artificial Intelligence (AAAI). AAAI Press. doi:10.1609/aaai.v38i8.28681

2024 doi
[11]

Jie Fang, Shutao Xia, Jianwu Lin, Zhikang Xia, Xiang Liu, and Yong Jiang. 2019. Alpha discovery neural network based on prior knowledge.arXiv preprint arXiv:1912.11761(2019)

2019 arXiv
[12]

Fuli Feng, Huimin Chen, Xiangnan He, Ji Ding, Maosong Sun, and Tat-Seng Chua
[13]

Fuli Feng, Xiangnan He, Xiang Wang, Cheng Luo, Yiqun Liu, and Tat-Seng Chua
[14]

Tingsheng Feng, Zhihao Shen, Xi Zhao, Xiaoni Lu, and Yuyang Zhou. 2025. CryptoMixer: Fine-grained market information-aware MLP Networks for Indi- vidual Cryptocurrency Trading Prediction. InACM International Conference on Knowledge Discovery & Data Mining (KDD). 603–614

2025
[15]

Itay Goldstein and Liyan Yang. 2015. Information diversity and complementarities in trading and information acquisition.Journal of Finance70, 4 (2015), 1723–1765

2015
[16]

Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752(2023)

2023 arXiv
[17]

Lars Kai Hansen and Peter Salamon. 2002. Neural network ensembles.IEEE Transactions on Pattern Analysis and Machine Intelligence12, 10 (2002), 993–1001

2002
[18]

Yi-Ling Hsu, Yu-Che Tsai, and Cheng-Te Li. 2021. Fingat: Financial graph atten- tion networks for recommending top-𝑘 k profitable stocks.IEEE Transactions on Knowledge and Data Engineering (TKDE)35, 1 (2021), 469–481

2021
[19]

Anders Krogh and Jesper Vedelsby. 1994. Neural network ensembles, cross validation, and active learning.Advances in Neural Information Processing Systems (NeurIPS)7 (1994)

1994
[20]

Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in Neural Information Processing Systems (NeurIPS)30 (2017)

2017
[21]

Yanyan Lan, Yadong Zhu, Jiafeng Guo, Shuzi Niu, and Xueqi Cheng. 2014. Position-Aware ListMLE: A Sequential Learning Process for Ranking.. InUAI, Vol. 14. 449–458

2014
[22]

Tong Li, Zhaoyang Liu, Yanyan Shen, Xue Wang, Haokun Chen, and Sen Huang
[23]

Zhige Li, Derek Yang, Li Zhao, Jiang Bian, Tao Qin, and Tie-Yan Liu. 2019. In- dividualized indicator for all: Stock-wise technical indicator optimization with stock embedding. InACM International Conference on Knowledge Discovery & Data Mining (KDD). 894–902

2019
[24]

Sida Lin, Yankai Chen, Yiyan Qi, Chenhao Ma, Bokai Cao, Yifei Zhang, Xue Liu, and Jian Guo. 2025. CSPO: Cross-Market Synergistic Stock Price Movement Forecasting with Pseudo-volatility Optimization.arXiv preprint arXiv:2503.22740 (2025)

2025 arXiv
[25]

YuJun Liu, Chen-Hui Song, Peiyuan Liu, Naiqi Li, Tao Dai, Jigang Bao, Yong Jiang, and Shu-Tao Xia. 2025. MERA: Mixture of Experts with Retrieval-Augmented Representation for Modeling Diversified Stock Patterns. InProceedings of the ACM on Web Conference (WWW). 1148–1152

2025
[26]

Yong Liu and Xin Yao. 1999. Ensemble learning via negative correlation.Neural networks12, 10 (1999), 1399–1404

1999
[27]

Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

2017 arXiv
[28]

Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al . 2011. Scikit-learn: Machine learning in Python.the Journal of Machine Learning Research12 (2011), 2825–2830

2011
[29]

Molei Qin, Xinyu Cai, Yewen Li, Haochong Xia, Chuqiao Zong, Shuo Sun, Xinrun Wang, and Bo An. 2026. FineFT: Efficient and Risk-Aware Ensemble Reinforce- ment Learning for Futures Trading. InACM International Conference on Knowledge Discovery & Data Mining (KDD). 1194–1203

2026
[30]

Ramit Sawhney, Shivam Agarwal, Arnav Wadhwa, Tyler Derr, and Rajiv Ratn Shah. 2021. Stock selection via spatiotemporal hypergraph attention network: A learning to rank approach. InConference on Artificial Intelligence (AAAI), Vol. 35. 497–504

2021
[31]

Hao Shi, Weili Song, Xinting Zhang, Jiahe Shi, Cuicui Luo, Xiang Ao, Hamid Arian, and Luis Angel Seco. 2025. AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors. InConference on Artificial Intelligence (AAAI), Vol. 39. 12524–12532

2025
[32]

Yu Shi, Yitong Duan, and Jian Li. 2026. Navigating the Alpha Jungle: An LLM- Powered MCTS Framework for Formulaic Alpha Factor Mining. InConference on Artificial Intelligence (AAAI), Vol. 40. 997–1005

2026
[33]

Shuo Sun, Xinrun Wang, Wanqi Xue, Xiaoxuan Lou, and Bo An. 2023. Mastering stock markets with efficient mixture of diversified trading experts. InACM International Conference on Knowledge Discovery & Data Mining (KDD). 2109– 2119

2023
[34]

Ziyi Tang, Zechuan Chen, Jiarui Yang, Jiayao Mai, Yongsen Zheng, Keze Wang, Jin- rui Chen, and Liang Lin. 2025. AlphaAgent: LLM-Driven Alpha Mining with Reg- ularized Exploration to Counteract Alpha Decay.arXiv preprint arXiv:2502.16789 (2025)

2025 arXiv
[35]

2019.Finding Alphas: A quantitative approach to building trading strategies

Igor Tulchinsky. 2019.Finding Alphas: A quantitative approach to building trading strategies. John Wiley & Sons

2019
[36]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in Neural Information Processing Systems (NeurIPS)30 (2017)

2017
[37]

Heyuan Wang, Tengjiao Wang, Shun Li, Jiayi Zheng, Shijie Guan, and Wei Chen
[38]

Fen Xia, Tie-Yan Liu, Jue Wang, Wensheng Zhang, and Hang Li. 2008. Listwise approach to learning to rank: theory and algorithm. InInternational Conference on Machine Learning (ICML). 1192–1199

2008
[39]

Hongjie Xia, Huijie Ao, Long Li, Yu Liu, Sen Liu, Guangnan Ye, and Hongfeng Chai. 2024. CI-STHPAN: pre-trained attention network for stock selection with channel-independent spatio-temporal hypergraph. InConference on Artificial Intelligence (AAAI). AAAI Press. doi:10.1609/aaa...

2024 doi
[40]

2019.Stock Alpha Min- ing Based On Genetic Algorithm.Technical Report

Ziyu Li Xiaoming Lin, Ye Chen and Kang He. 2019.Stock Alpha Min- ing Based On Genetic Algorithm.Technical Report. Huatai Securities Re- search Center, https://crm.htsc.com.cn/doc/2019/10750101/f75b4b6a-2bdd-4694- b6964c62528791ea.pdf

2019
[41]

Chen Yang, Jingyuan Wang, Xiaohan Jiang, and Junjie Wu. 2025. Learning Univer- sal Multi-level Market Irrationality Factors to Improve Stock Return Forecasting. arXiv preprint arXiv:2502.04737(2025)

2025 arXiv
[42]

Jaemin Yoo, Yejun Soun, Yong-chan Park, and U Kang. 2021. Accurate multivariate stock movement prediction via data-axis transformer with multi-level contexts. InACM International Conference on Knowledge Discovery & Data Mining (KDD). 2037–2045

2021
[43]

Shuo Yu, Hongyan Xue, Xiang Ao, Feiyang Pan, Jia He, Dandan Tu, and Qing He. 2023. Generating synergistic formulaic alpha collections via reinforcement learning. InACM International Conference on Knowledge Discovery & Data Mining (KDD). 5476–5486

2023
[44]

Zhaojian Yu, Yinghao Wu, Genesis Wang, and Heming Weng. 2024. MIGA: Mixture-of-Experts with Group Aggregation for Stock Market Prediction.arXiv preprint arXiv:2410.02241(2024)

2024 arXiv
[45]

Tianping Zhang, Yuanqi Li, Yifei Jin, and Jian Li. 2020. Autoalpha: an efficient hierarchical evolutionary algorithm for mining alpha factors in quantitative investment.arXiv preprint arXiv:2002.08245(2020)

2020 arXiv
[46]

Zetao Zheng, Jie Shao, Jia Zhu, and Heng Tao Shen. 2023. Relational Temporal Graph Convolutional Networks for Ranking-Based Stock Prediction. InInterna- tional Conference on Data Engineering (ICDE). 123–136. doi:10.1109/ICDE55515. 2023.00017 MAPLE: Efficient and Diverse Multi-...

2023
[2017]

Measuring systemic risk.Review of Financial Studies30, 1 (2017), 2–47

2017
[2018]

Enhancing stock movement prediction with adversarial training.arXiv preprint arXiv:1810.09936(2018)

2018 arXiv
[2019]

Temporal relational ranking for stock prediction.ACM Transactions on Information Systems (TOIS)37, 2 (2019), 1–30

2019
[2022]

Conference’17, July 2017, Washington, DC, USA Den and Chen et al

Adaptive Long-Short Pattern Transformer for Stock Investment Selection.. Conference’17, July 2017, Washington, DC, USA Den and Chen et al. InInternational Joint Conference on Artificial Intelligence (IJCAI). 3970–3977

2017
[2024]

In Conference on Artificial Intelligence (AAAI), Vol

Master: Market-guided stock transformer for stock price forecasting. In Conference on Artificial Intelligence (AAAI), Vol. 38. 162–170

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Viral V Acharya, Lasse H Pedersen, Thomas Philippon, and Matthew Richardson

work page

[2] [2]

Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normaliza- tion.arXiv preprint arXiv:1607.06450(2016)

arXiv 2016

[3] [3]

Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007. Learning to rank: from pairwise approach to listwise approach. InInternational Conference on Machine Learning (ICML)(Corvalis, Oregon, USA)(ICML ’07). Association for Computing Machinery, New York, NY, USA, 129–136. doi:10.1145/1273496. 1273513

doi:10.1145/1273496 2007

[4] [4]

Zhiyuan Cao, Jiayu Xu, Chengqi Dong, Peiwen Yu, and Tian Bai. 2024. MATCC: A Novel Approach for Robust Stock Price Prediction Incorporating Market Trends and Cross-time Correlations. InACM International Conference on Information and Knowledge Management (CIKM). 187–196

work page 2024

[5] [5]

Weijun Chen and Yanze Wang. 2025. DHMoE: Diffusion Generated Hierarchi- cal Multi-Granular Expertise for Stock Prediction. InConference on Artificial Intelligence (AAAI), Vol. 39. 11490–11499

work page 2025

[6] [6]

Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555(2014)

arXiv 2014

[7] [7]

Yu-Chen Den, Kuan-Yu Chen, Kendro Vincent, and Darby Tien-Hao Chang. 2026. Integrating Inductive Biases in Transformers via Distillation for Financial Time Series Forecasting.arXiv preprint arXiv:2603.16985(2026)

work page Pith review arXiv 2026

[8] [8]

Eugene F Fama and Kenneth R French. 1992. The cross-section of expected stock returns.Journal of Finance47, 2 (1992), 427–465

work page 1992