Deep Learning Enhanced Multi-Day Turnover Quantitative Trading Algorithm for Chinese A-Share Market

Yimin Du

REVIEW 5 major objections 5 minor 21 references

Deep Learning Enhanced Multi-Day Turnover Quantitative Trading Algorithm for Chinese A-Share Market

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A five-module pipeline claims 15.2% annualized returns with a 1.87 Sharpe on Chinese A-shares.

desk verdict A coherent five-module A-share pipeline whose reported 15.2% return and Sharpe 1.87 do not survive contact with its own cost table and validation/test overlap. read the letter →

arxiv 2506.06356 v1 pith:NQEEMIPY submitted 2025-06-03 cs.CE cs.LG

Yimin Du This is my paper

classification cs.CEcs.LG

keywords quantitativetradingdeeplearningChineseA-sharemarketcross-sectionalstockpredictionopeningauctionsignalspositionsizinggrid-searchoptimizationtiming

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 tries to establish that a five-module, daily-rebalanced algorithm can combine deep cross-sectional stock ranking with opening-auction timing, liquidity-aware position sizing, grid-optimized exits, and volatility-based market timing to generate scalable alpha in the Chinese A-share market. The out-of-sample results for 2021-2024 are reported as 15.2% annualized return, 4.8% maximum drawdown, and a 1.87 Sharpe ratio, with estimated capacity of 8-12 billion RMB across 50-100 positions. A sympathetic reader would care because the claimed performance sits between high-frequency strategies that cannot scale and slow fundamental strategies that miss intermediate opportunities. The paper interprets the result as evidence that microstructure inefficiencies in the A-share opening auction remain exploitable after transaction costs, and that each module contributes measurable value on top of the others.

What carries the argument

The central machinery is the daily rebalancing pipeline in Algorithm 1, whose load-bearing entry rule is the opening signal of Eq. (5): a weighted combination of the overnight gap, pre-market volume ratio, GARCH(1,1) volatility, and a sentiment term, modeled as a three-component Gaussian mixture and gated by a dynamic threshold that tightens with realized volatility. Cross-sectional scores from the deep network filter the candidate universe, the position-sizing formula turns scores into market-impact-constrained weights, the grid search selects profit-taking and stop-loss levels by volatility regime, and the final timing signal scales overall exposure. The claim is that this pipeline as a whole, not any single component, is what converts raw A-share data into the reported risk-adjusted returns.

What would settle it

Re-run the backtest with trade signals computed only from data available before the opening auction—previous day's close, pre-open indicative volume, and prior volatility—and execute at the opening auction; if the 15.2% annualized return and 1.87 Sharpe ratio fall materially, the reported edge depends on using the opening price simultaneously as signal and execution price.

Watch

Extended reading notes

Core claim

The author's central claim is that the integrated pipeline—a temperature-softmax cross-sectional score from a deep feedforward network, a three-Gaussian mixture model of opening signals built from the overnight gap, pre-market volume ratio, GARCH volatility, and sentiment, market-cap and liquidity position sizing with 0.5-2.0% position limits, a 1,344-combination grid search over profit-taking, stop-loss, holding-period, and trailing-stop parameters, and a regime-aware volatility timing overlay—produces 15.2% annualized out-of-sample returns with 8.1% volatility, 4.8% maximum drawdown, and a 1.87 Sharpe ratio over 2021-2024. The ablation table carries the main evidence: starting from a random baseline at 3.2% return, the cross-sectional module adds 8.2 percentage points, opening signals add 1.7, position sizing adds 1.2, grid optimization adds 0.5, and market timing adds 0.4. The author takes this ordering to show that cross-sectional stock selection does the main work while the other modules improve timing, risk, and drawdown control.

Load-bearing premise

The entry signal uses the opening price and pre-market volume to decide a trade, but the paper does not explain how that signal is computed before the opening price is set, so the backtest may be buying at a price that already contains the information used to trigger the buy.

Editorial extensions

If this is right

If the 2021-2024 out-of-sample numbers hold, the Chinese A-share opening auction still contains a tradeable anomaly that can be captured by mixture-model probabilities rather than fixed entry rules.
The ablation ordering implies the deep cross-sectional ranker contributes the largest share of alpha, so a simpler implementation that kept only that module might retain much of the reported performance.
A 9-day maximum holding period with an average of 6.2 days means capital cycles roughly 17 times per year, so maintaining the stated 22.1 basis points of per-trade cost is a make-or-break operational condition.
The capacity estimate of 8-12 billion RMB follows from 50-100 positions, 0.5-2.0% position limits, and a 10% average-daily-volume participation cap, and is therefore directly testable in live execution.
Regime-based results showing positive returns even in high-volatility bear markets would make the strategy a diversifying allocation inside a broader equity portfolio.

Reading between the lines

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

The paper's own limitations section concedes that performance depends on continued market inefficiencies; a direct implication is that the alpha is a decaying resource, so live forward monitoring is the only way to measure how fast it decays.
Because the ablation table isolates each module's contribution, a testable extension is to run the cross-sectional ranker alone with a simple liquidity-weighted entry rule; if it retains most of the 1.87 Sharpe ratio, the opening-signal and timing modules are refinements rather than necessary components.
The same modular pipeline could be transferred to other auction-based markets such as Hong Kong or Japan, though the paper does not test this; the A-share-specific features would need re-estimation, but the architecture is specified in market-neutral terms.
A deployment-oriented check is to reconstruct gross-to-net returns from the stated 22.1 basis points per side and 2100% annual turnover, since the reported net return depends on that reconciliation holding in live execution.

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.

Desk Editor's Note

read the letter

Dear colleague,

Short version: the paper is a five-module A-share trading pipeline with a headline backtest that does not hold up against its own cost table. The architecture is coherent, and the ablation study is a nice touch, but the core performance claim is unreliable.

What is new and useful: the integration of cross-sectional deep ranking, Gaussian-mixture opening signals, liquidity-scaled position sizing, grid-searched exits, and volatility timing into one daily pipeline is a reasonable engineering package. The modular design is clearly described, and the paper does a fair job of situating itself in the ML-and-returns literature.

The problems are load-bearing. First, Table VI says total transaction costs are 464 bp per year, which is 30.5% of gross return. If gross return were 15.2%, net return would be about 10.6%, not the reported 15.2%. Either the return is gross and labeled net, or the costs are wrong. Second, the validation period 2020-2021 overlaps the reported out-of-sample window 2021-2024, so at least one year of the test is in-sample. Third, the entry signal in Eq. (5) uses the day's open price and pre-market volume, which are only known when the opening auction concludes. No mechanism is given to compute the signal before the fill price is set; the phrase "slightly lagged entry timing" does not repair this. Fourth, the stress tests are run on periods inside the training/validation data (2015, 2018, 2020, 2022), so they are not out-of-sample robustness checks. Add a long list of free hyperparameters selected on validation, and the 15.2% and Sharpe 1.87 look like a selected outcome rather than an independent test.

The paper provides no code or data, so none of this can be checked or reproduced. That is a separate problem, but it matters for the credibility of the claim.

Who is this for? Someone who wants a catalog of modules for an A-share strategy might skim the method section, but the empirical contribution is not usable as evidence. I would not cite it in the next twelve months. It could serve a reading group as an example of backtest pitfalls, but not as a serious empirical paper.

Recommendation on peer review: desk reject. The internal inconsistencies and the validation/test overlap are not fixable with minor revisions; they cut to the core of the reported performance. If the author supplied code, data, and corrected cost accounting, it could be reconsidered.

Best

Referee Report

5 major / 5 minor

Summary. This paper proposes a five-module daily-rebalancing quantitative trading algorithm for the Chinese A-share market, combining a deep cross-sectional stock-ranking network, opening-signal mixture modeling, position sizing, grid-searched profit-taking/stop-loss rules, and a volatility-based market-timing overlay. The authors train on 2010-2020 data, use a stated validation period of 2020-2021, and report out-of-sample results for 2021-2024 with 15.2% annualized return, 4.8% maximum drawdown, and a 1.87 Sharpe ratio, along with an 8-12 billion RMB capacity claim. The empirical case rests on Table I and a sequence of ablation, regime, cost, and stress-test analyses.

Significance. If the reported performance were credible, the paper would demonstrate a rare combination of deep-learning stock selection and microstructure-based entry timing with institutional-scale capacity in a major emerging market. The modular architecture is clearly described and the main equations are explicit, and the authors make a serious attempt to model transaction costs and capacity. However, the central empirical claims are undermined by a likely look-ahead bias in the entry signal, a validation/test period overlap, and an internally inconsistent cost table; the paper provides no code, no data source, and no statistical inference, so the headline numbers cannot be independently checked. The significance of the contribution is therefore not assessable from the manuscript as written.

major comments (5)

[Section III-C, Eq. (5), Algorithm 1] The entry signal S_{i,t} is a function of Gap_{i,t} = (P_open,t - P_close,t-1)/P_close,t-1 and the pre-market volume ratio VR_{i,t}; both quantities are resolved only when the opening auction concludes. Algorithm 1 calls OpeningSignalAnalysis before any fill, but the paper never specifies how S_{i,t} can be computed from data available before the opening price is set, nor does it state that the backtest fills at a later price. If the backtest fills at the opening price using a signal that embeds that same price, the reported 15.2% return and 1.87 Sharpe ratio are the result of look-ahead information. The closing remark in Section VII ('slightly lagged but stable entry timing') does not resolve this, because a later fill still leaves the opening price inside the signal and would require an unmodeled cost for the delay.
[Section IV-B] The validation period is stated as 2020-2021 and the out-of-sample test as 2021-2024. The year 2021 is therefore present in both the validation set used for hyperparameter selection and the reported out-of-sample period. Since Table I reports performance over 2021-2024, one of the four years is in-sample, so the headline 15.2% / 4.8% / 1.87 figures are not a pure out-of-sample evaluation. The paper should either move the validation boundary to 2020 or report 2022-2024 separately.
[Table VI, Section V-F] Table VI labels the total annual cost of 464 bp as 30.5% of gross return. If the annualized gross return is 15.2%, then 30.5% of that is 464 bp, leaving a net return of 10.6%, which contradicts the reported net return of 15.2%. Alternatively, if 15.2% is the net return, the gross return must be 21.9%, in which case the total cost in basis points would be 670 bp, not 464 bp. The two stated quantities are mutually inconsistent, and the table needs to be recomputed or clarified.
[Section V-G, Table VII] The stress test in Table VII includes the 2015 A-share crisis, the 2018 trade tensions, and the 2020 COVID pandemic. These events lie inside the 2010-2020 training period, so the stress-test results are in-sample and do not provide evidence of resilience in unseen regimes. Only the 2022 lockdown scenario is out-of-sample. The stress-test section should be restricted to post-2020 events or explicitly re-labelled as an in-sample robustness check.
[Section III-E, Eq. (16)] The grid-search objective in Eq. (16) explicitly maximizes cumulative return divided by maximum drawdown (with weight w2=0.35) together with win rate, turnover efficiency, and consistency over 1,344 parameter combinations on the validation window. Because the validation window overlaps the reported test window (see the earlier comment on Section IV-B) and because many additional hyperparameters (dropout, temperature, loss weight, thresholds, exponents, market-impact parameters) are also selected on the same or adjacent data, the headline Sharpe ratio is the maximum of a large search rather than an unbiased estimate. The paper needs either a nested validation design or an explicit statement that the reported figures are in-sample optima after correcting the overlap.

minor comments (5)

[Section IV-A] The universe-filter description is truncated: the sentence 'Exclusion of stocks with extreme price movements (>30' ends without a threshold or a closing parenthesis. The criterion should be completed.
[Abstract and throughout] The text contains numerous typographical errors, including 'a dvanced', 'signi ﬁcant', and 'techni ques' in the abstract and introduction. A careful proofreading pass is needed.
[Section III-C, Eq. (5)] The sentiment term Sentiment_{i,t} is not defined anywhere; a precise formula or a reference to a data source for this term is necessary for reproducibility.
[Section IV-A] The paper does not state the data vendor (e.g., Wind, CSMAR, or other) and provides no code or data release. Adding this information and at least a pseudo-code or reproducibility note would substantially increase the value of the empirical section.
[Acknowledgments] The paper is authored by a single person, but the acknowledgments thank 'the authors' and 'we'; the wording should be aligned with the actual authorship.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline 'out-of-sample' performance is partially in-sample: the validation window (2020-2021) overlaps the reported test window (2021-2024), and Eq. (16) maximizes the same return/drawdown metric that Table I presents as the result.

fitted input called prediction [Section IV.B (Training/Validation/Out-of-Sample periods); Table I]
"V alidation Period: 2020-2021 (1 year) - Used for model selection, parameter tuning, and strategy component integration. Out-of-Sample T esting: 2021-2024 (4 years) - Pure out-of-sample evaluation with no parameter adjustments."

The validation interval 2020-2021 and the out-of-sample interval 2021-2024 overlap in calendar year 2021. The grid-search module (Section III.E) selects PT/SL/MHP/TSA by maximizing Eq. (16) on the validation data, and Table I reports 15.2% annualized return, 4.8% maximum drawdown, and Sharpe 1.87 for 2021-2024 as 'out-of-sample.' Because one of the four reported test years was used for parameter selection, the reported performance is not purely predicted; a quarter of the test window is, by the paper's own period definitions, a fitted input.
fitted input called prediction [Section III.E.2 Eq. (16); Table I]
"Objective =w1·WinRate +w2· CumReturn MaxDrawdown +w3·TurnoverEfﬁciency +w4·Consistency"

The metric advertised as the achievement—'maximum drawdown constrained below 5%' with 15.2% annualized return—is exactly the ratio w2 maximizes over the grid. The paper then presents the realized drawdown and return as an empirical discovery rather than as the value of the selection criterion. This would be acceptable if the test set were disjoint; because 2021 is shared with the validation period, the headline numbers are in part the optimized objective itself, not an independent confirmation.

full rationale

The paper contains no self-citation chain or imported uniqueness theorem; its references are standard literature. The central empirical claim is not derived analytically from an ansatz but is a backtest result. However, the claimed out-of-sample test is not clean: Section IV.B defines Validation as 2020-2021 and Out-of-Sample Testing as 2021-2024, so calendar 2021 is used both for parameter selection and for reported performance. Equally, the grid-search objective in Eq. (16) maximizes CumReturn/MaxDrawdown, the exact pair reported in Table I. Thus at least part of the headline '15.2% return, 4.8% drawdown, Sharpe 1.87' is the fitted objective evaluated partly on data used for fitting, not a pure out-of-sample prediction. The opening-price signal look-ahead question is a validity and correctness risk, not a circularity, so it is noted but not scored. I therefore assign a partial-circularity score of 6.

Assumptions & free parameters 16 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a large number of explicit hyperparameters and estimated weights, many of which are not reported with values. The tradeability of the open-price signal and the accuracy of the market impact model are unvalidated domain assumptions. The walk-forward with grid search assumes that parameter selection on validation does not overfit the out-of-sample path, an assumption that is not tested. These items are the true cost of the paper: the performance numbers are the product of many choices, and no external prediction is offered that would break the circularity.

free parameters (16)

Loss weight α = 0.7
Equation (4), weighting between ranking and regression loss, chosen by hand.
Softmax temperature T = 2.0
Equation (3), set to 2.0 for output normalization.
Dropout rates = 0.3 hidden, 0.1 input
Section III-B2, regularization hyperparameters.
Number of mixture components K = 3
Equation (6), Gaussian components for opening signals.
Opening signal weights α1-α4 = not reported
Equation (5), estimated by time-varying regression; values not disclosed.
Entry threshold parameters θ0, β = not reported
Equation (9), adaptive threshold; values not disclosed.
Position sizing exponents = 0.5, 0.2, 0.3, 0.5
Equation (10), exponents for market cap, momentum, ADV, volatility.
MaxParticipation = 10%
Equation (11), liquidity constraint for market impact.
Portfolio constraints = 0.5-2.0% individual, 25% sector, 20-60% cap
Equations (12)-(14), diversification limits.
VIX scaling coefficient = 0.5
Equation (15), volatility-based position scaling.
Grid-search parameter sets = PT 8 levels, SL 7, MHP 6, TSA 4 (1,344 combos)
Section III-E1, exit rule parameters.
Objective weights w1-w4 = 0.25, 0.35, 0.25, 0.15
Equation (16), optimization objective weights.
Regime smoothing factor = 0.7 current, 0.3 previous
Equation (17), adaptive parameter smoothing.
Volatility combination weights = not reported
Equation (18), Kalman-estimated weights; values not disclosed.
Timing signal weights β1-β3 = not reported
Equation (20), weights on timing components; values not disclosed.
Transaction cost inputs = commission 5bp, stamp 10bp, impact coefficient 0.5
Section IV-C and Eq. (21), assumptions for cost model.

assumptions (5)

domain assumption Historical market data contains predictive information about future relative returns
The cross-sectional prediction module assumes this; no external evidence is provided.
domain assumption The opening price in the signal is tradeable at that same price
Eq. (5) uses the open price to construct the signal, but the paper does not show the signal is available before the open price is fixed.
domain assumption The market impact model (Eq. 21) accurately represents real execution costs
The net return and capacity claims depend on this model, which is not calibrated to actual trades.
domain assumption A 3-state HMM with Viterbi decoding identifies market regimes correctly
Section III-E3 uses the regime classification to switch parameters; no validation of regime accuracy is provided.
ad hoc to paper The grid-search on validation does not overfit the out-of-sample period
Many parameters are selected on validation and then a single 4-year out-of-sample backtest is reported without multiple-testing correction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep Learning Enhanced Multi-Day Turnover Quantitative Trading Algorithm for Chinese A-Share Market." pith.science (2026). https://pith.science/paper/NQEEMIPY

@misc{pith2026250606356,
  author       = {Pith},
  title        = {Pith review of: Deep Learning Enhanced Multi-Day Turnover Quantitative Trading Algorithm for Chinese A-Share Market},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NQEEMIPY}},
  note         = {Machine review of arXiv:2506.06356}
}

read the original abstract

This paper presents a sophisticated multi-day turnover quantitative trading algorithm that integrates advanced deep learning techniques with comprehensive cross-sectional stock prediction for the Chinese A-share market. Our framework combines five interconnected modules: initial stock selection through deep cross-sectional prediction networks, opening signal distribution analysis using mixture models for arbitrage identification, market capitalization and liquidity-based dynamic position sizing, grid-search optimized profit-taking and stop-loss mechanisms, and multi-granularity volatility-based market timing models. The algorithm employs a novel approach to balance capital efficiency with risk management through adaptive holding periods and sophisticated entry/exit timing. Trained on comprehensive A-share data from 2010-2020 and rigorously backtested on 2021-2024 data, our method achieves remarkable performance with 15.2\% annualized returns, maximum drawdown constrained below 5\%, and a Sharpe ratio of 1.87. The strategy demonstrates exceptional scalability by maintaining 50-100 daily positions with a 9-day maximum holding period, incorporating dynamic profit-taking and stop-loss mechanisms that enhance capital turnover efficiency while preserving risk-adjusted returns. Our approach exhibits robust performance across various market regimes while maintaining high capital capacity suitable for institutional deployment.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

[1]

Common risk factors in the ret urns on stocks and bonds,

E. F. Fama and K. R. French, "Common risk factors in the ret urns on stocks and bonds," Journal of Financial Economics , vol. 33, no. 1, pp. 3-56, 1993

work page 1993
[2]

On persistence in mutual fund performanc e,

M. M. Carhart, "On persistence in mutual fund performanc e," The Journal of Finance , vol. 52, no. 1, pp. 57-82, 1997

work page 1997
[3]

Consumer credit-r isk models via machine-learning algorithms,

A. E. Khandani, A. J. Kim, and A. W. Lo, "Consumer credit-r isk models via machine-learning algorithms," Journal of Banking & Finance , vol. 34, no. 11, pp. 2767-2787, 2010

work page 2010
[4]

Deep learning with long short- term memory networks for ﬁnancial market predictions,

T. Fischer and C. Krauss, "Deep learning with long short- term memory networks for ﬁnancial market predictions," European Journal of Oper- ational Research, vol. 270, no. 2, pp. 654-669, 2018

work page 2018
[5]

Financia l time series forecasting with deep learning: A systematic litera ture review: 2005–2019,

O. B. Sezer, M. U. Gudelek, and A. M. Ozbayoglu, "Financia l time series forecasting with deep learning: A systematic litera ture review: 2005–2019," Applied Soft Computing , vol. 90, p. 106181, 2020

work page 2005
[6]

Informer: Beyond efﬁcient transformer f or long se- quence time-series forecasting,

H. Zhou et al., "Informer: Beyond efﬁcient transformer f or long se- quence time-series forecasting," Proceedings of the AAAI Conference on Artiﬁcial Intelligence , vol. 35, no. 12, pp. 11106-11115, 2021

work page 2021
[7]

and the cross-sectio n of expected returns,

C. R. Harvey, Y . Liu, and H. Zhu, "... and the cross-sectio n of expected returns," The Review of Financial Studies , vol. 29, no. 1, pp. 5-68, 2016

work page 2016
[8]

The real value of China’s stock market,

J. N. Carpenter, F. Lu, and R. F. Whitelaw, "The real value of China’s stock market," Journal of Financial Economics , vol. 139, no. 3, pp. 679-703, 2021

work page 2021

Show all 21 references

[9]

Machine learning f or stock selection and portfolio optimization in Chinese A-sh are market,

G. Wang, C. Tang, C. Kang, and J. Chen, "Machine learning f or stock selection and portfolio optimization in Chinese A-sh are market," Financial Innovation, vol. 6, no. 1, pp. 1-23, 2020

2020
[10]

Dynamic conditional correlation: A simpl e class of multivariate generalized autoregressive conditional het eroskedasticity models,

R. F. Engle, "Dynamic conditional correlation: A simpl e class of multivariate generalized autoregressive conditional het eroskedasticity models," Journal of Business & Economic Statistics , vol. 20, no. 3, pp. 339-350, 2002

2002
[11]

Financial time se ries prediction using hybrids of chaos theory, multi-layer perceptron and m ulti-objective evolutionary algorithms,

V . Ravi, D. Pradeepkumar, and K. Deb, "Financial time se ries prediction using hybrids of chaos theory, multi-layer perceptron and m ulti-objective evolutionary algorithms," Swarm and Evolutionary Computation , vol. 36, pp. 136-149, 2017

2017
[12]

A new interpretation of information rat e,

J. L. Kelly Jr, "A new interpretation of information rat e," Bell System Technical Journal, vol. 35, no. 4, pp. 917-926, 1956

1956
[13]

The Kelly criterion in blackjack sports be tting, and the stock market,

E. O. Thorp, "The Kelly criterion in blackjack sports be tting, and the stock market," Handbook of Asset and Liability Management , pp. 385- 428, 2006

2006
[14]

Optimal execution of portfol io transactions,

R. Almgren and N. Chriss, "Optimal execution of portfol io transactions," Journal of Risk , vol. 3, pp. 5-39, 2001

2001
[15]

Cartea, S

Á. Cartea, S. Jaimungal, and J. Penalva, Algorithmic and high-frequency trading. Cambridge University Press, 2015

2015
[16]

A reﬁnement to the Sharpe ratio and inf ormation ratio,

C. L. Israelsen, "A reﬁnement to the Sharpe ratio and inf ormation ratio," Journal of Asset Management , vol. 5, no. 6, pp. 423-427, 2005

2005
[17]

Risks and portfolio decision s involving hedge funds,

V . Agarwal and N. Y . Naik, "Risks and portfolio decision s involving hedge funds," The Review of Financial Studies , vol. 17, no. 1, pp. 63- 98, 2004

2004
[18]

Big data and AI s trategies: Machine learning and alternative data approach to investin g,

M. Kolanovic and R. T. Krishnamachari, "Big data and AI s trategies: Machine learning and alternative data approach to investin g," J.P . Morgan Global Quantitative and Derivatives Strategy , 2017

2017
[19]

Twitter mood predicts th e stock market,

J. Bollen, H. Mao, and X. Zeng, "Twitter mood predicts th e stock market," Journal of Computational Science , vol. 2, no. 1, pp. 1-8, 2011

2011
[20]

Forecastin g the equity risk premium: the role of technical indicators,

C. J. Neely, D. E. Rapach, J. Tu, and G. Zhou, "Forecastin g the equity risk premium: the role of technical indicators," Management Science , vol. 60, no. 7, pp. 1772-1791, 2014

2014
[21]

Empirical asset pricing via machine learning,

S. Gu, B. Kelly, and D. Xiu, "Empirical asset pricing via machine learning," The Review of Financial Studies , vol. 33, no. 5, pp. 2223- 2273, 2020

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Common risk factors in the ret urns on stocks and bonds,

E. F. Fama and K. R. French, "Common risk factors in the ret urns on stocks and bonds," Journal of Financial Economics , vol. 33, no. 1, pp. 3-56, 1993

work page 1993

[2] [2]

On persistence in mutual fund performanc e,

M. M. Carhart, "On persistence in mutual fund performanc e," The Journal of Finance , vol. 52, no. 1, pp. 57-82, 1997

work page 1997

[3] [3]

Consumer credit-r isk models via machine-learning algorithms,

A. E. Khandani, A. J. Kim, and A. W. Lo, "Consumer credit-r isk models via machine-learning algorithms," Journal of Banking & Finance , vol. 34, no. 11, pp. 2767-2787, 2010

work page 2010

[4] [4]

Deep learning with long short- term memory networks for ﬁnancial market predictions,

T. Fischer and C. Krauss, "Deep learning with long short- term memory networks for ﬁnancial market predictions," European Journal of Oper- ational Research, vol. 270, no. 2, pp. 654-669, 2018

work page 2018

[5] [5]

Financia l time series forecasting with deep learning: A systematic litera ture review: 2005–2019,

O. B. Sezer, M. U. Gudelek, and A. M. Ozbayoglu, "Financia l time series forecasting with deep learning: A systematic litera ture review: 2005–2019," Applied Soft Computing , vol. 90, p. 106181, 2020

work page 2005

[6] [6]

Informer: Beyond efﬁcient transformer f or long se- quence time-series forecasting,

H. Zhou et al., "Informer: Beyond efﬁcient transformer f or long se- quence time-series forecasting," Proceedings of the AAAI Conference on Artiﬁcial Intelligence , vol. 35, no. 12, pp. 11106-11115, 2021

work page 2021

[7] [7]

and the cross-sectio n of expected returns,

C. R. Harvey, Y . Liu, and H. Zhu, "... and the cross-sectio n of expected returns," The Review of Financial Studies , vol. 29, no. 1, pp. 5-68, 2016

work page 2016

[8] [8]

The real value of China’s stock market,

J. N. Carpenter, F. Lu, and R. F. Whitelaw, "The real value of China’s stock market," Journal of Financial Economics , vol. 139, no. 3, pp. 679-703, 2021

work page 2021