acceptodds
Under review as a conference paper at ICLR 2027

HiMarket: A HIERARCHICAL MARKET WORLD MODEL FOR STOCK SELECTION

10 views

Abstract

Stock selection plays a pivotal role in quantitative investment, requiring models to identify stocks likely to outperform under changing market conditions. Recently, representation learning and pretraining methods have shown promise for this task. However, market states are implicit, hierarchical, and time-varying, making it difficult to separate shared dynamics from stock-specific variation. Moreover, predicting shared dynamics alone does not identify which stocks will outperform. To address these challenges, we propose HiMarket, a Hierarchical Market world model for stock selection. Specifically, a hierarchical market-state construction module decomposes encoded stock representations into global, dynamic-group, and residual components. Then, during pretraining, a market-conditioned residual prediction objective uses forecast shared market states to predict future stock residual representations. Finally, a stock-conditioned factor-query readout combines encoded stock representations with their residual components to form stock-specific queries that retrieve context from historical global and group memories. Experiments on three real-world stock-market datasets demonstrate that HiMarket achieves state-of-the-art (SOTA) performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Stock Selection, World Models, Financial Time Series, Self-Supervised Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_wXxAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HiMarket is a hierarchical predictive representation learning framework for stock selection, dividing stock representation into three parts: global, dynamic grouping, and individual stock residuals. In the pre-training phase, the model first predicts the future shared market state, then uses this as a condition to predict individual stock residuals, and uses an EMA teacher model to provide the training objective. In the downstream phase, the model uses a Factor-Query module to extract information from historical market states by combining stock representations and residuals, and performs fine-tuning of stock ranking.

The paper compares HiMarket with 14 baseline configurations on NASDAQ, NYSE, and CSMD300. Results show that HiMarket achieves the highest cumulative return on all three datasets and the highest Sharpe ratio on NASDAQ and NYSE.

Critical strengths
  1. Coherent Methodology. Global, grouping, and residual decomposition, combined with market state condition prediction and individual stock query modules, form a clear framework. The paper explicitly distinguishes between pre-training and downstream ranking, demonstrating that these representations are not necessarily independent or correspond to identifiable economic factors.

  2. Strong Experimental Results and Some Ablation Evidence. HiMarket achieved competitive results on three datasets. Table 3 shows that performance decreased after removing the pre-training or Factor-Query modules, indicating that both contribute to the overall model. The paper's architecture diagram and training details are also clear.

Critical weaknesses
  1. The contributions of the core mechanisms are not adequately differentiated. Table 3 demonstrates the effectiveness of the overall framework, but fails to separately prove the individual roles of hierarchical decomposition and market state condition residual prediction. Specifically, "Plain JEPA + Linear Readout" alters both the pre-training method and the downstream prediction head. Furthermore, the paper lacks ablation experiments that separately remove residual information from dynamic grouping, predicted state conditions, and queries. Therefore, it remains unclear which core designs truly contribute to performance improvements.

  2. The evaluation protocol and baseline comparability remain unclear. Table 2 does not clearly distinguish which baselines were reproduced by the authors and which directly cite results from existing papers. It is also unclear whether the methods use the same input features, preprocessing, data partitioning, stock validity screening, and portfolio rules. Section 5.1 defines the candidate stock pool using effective next-day return labels. If the validity of future labels affects stock selection, it may introduce forward bias. Additionally, the daily adjusted Top-5 portfolio does not account for transaction costs, slippage, and execution latency, limiting the practical interpretation of the return results. These issues do not prove that the results are wrong, but they do indicate that there are still important details in the assessment that need to be verified.

Questions for authors
  1. Can you remove the dynamic grouping, market state prediction conditions, and residual information from stock queries while keeping other settings unchanged to verify the contribution of each module?

  2. Which baselines were rerun by the authors, and which cited existing papers? Did they use the same data partitioning, features, preprocessing, stock efficiency screening, and portfolio rules?

  3. Is the Top-5 candidate stock pool entirely determined by information known at the time of decision-making? How do you handle the issue of missing next-day return labels? Can you supplement the performance under turnover rate and non-zero transaction fees?

3 comments
AuthorsOpenReview ↗

Response to Reviewer wXxA (Part 1/2)

We thank the reviewer for recognizing the motivation and coherence of HiMarket. We address the requests for controlled ablations, baseline provenance, and a verifiable portfolio evaluation protocol below. IRR denotes cumulative return and SR denotes annualized Sharpe ratio.

Q1. Controlled ablations

Can the authors provide controlled ablations that separately remove dynamic grouping, market conditioning in residual prediction, and the residual component of the stock-specific query, while keeping the remaining architecture and training protocol unchanged? This would help identify the contributions of the proposed hierarchical mechanisms.

Response. We have extended the controlled ablations to all three datasets (NASDAQ, NYSE, and CSMD300), using consistent chronological splits, input processing, and validation-based checkpoint selection. All results follow the original zero-fee, daily equal-weighted Top-5 close-to-close protocol. We distinguish the matched encoder comparison from the component ablations, as they use different transfer settings.

1. Pretrained encoder comparison. To address the readout confound in the original "Plain JEPA + Linear Readout" comparison, we transfer only the pretrained encoder from each method. The downstream hierarchical aggregator, statistics projection, and stock-ranking readout use identical architectures and initializations, with the same fine-tuning protocol.

Pretrained encoderNASDAQ IRR / SRNYSE IRR / SRCSMD300 IRR / SR
HiMarket84% / 2.6457% / 1.9725.02% / 1.807
Plain JEPA64% / 2.3818% / 0.734.84% / 0.581

HiMarket achieves higher IRR and SR across all three datasets under matched downstream conditions, providing controlled evidence for the effectiveness of its pretraining scheme. This comparison transfers only the encoder and therefore uses a different reference from the Full-model component ablations below.

2. Hierarchical component ablations. We separately examine dynamic grouping, group-level information, forecast market-state conditioning, and residual-aware stock queries. The affected training stages are retrained rather than modifying an already fine-tuned Full model only at inference time.

SettingNASDAQ IRR / SRNYSE IRR / SRCSMD300 IRR / SR
Full HiMarket83% / 2.6383% / 2.7846% / 2.39
Fixed Grouping43% / 1.8537% / 1.6114% / 1.10
w/o Group Information61% / 1.89−17% / −0.6037% / 2.51
w/o Forecast Conditioning72% / 1.6535% / 1.4531% / 2.35
w/o Residual in Query32% / 1.9471% / 2.16−12% / −0.77
  • Dynamic grouping: Fixed Grouping retains the group-information branch but uses predefined stock assignments that remain unchanged throughout pretraining and fine-tuning. Its lower IRR and SR across all three datasets support the value of adaptive group assignments.
  • Group-level information: The w/o Group Information variant removes the effective group-information branch during both pretraining and downstream fine-tuning. Removing group information reduces IRR across all datasets and SR on NASDAQ and NYSE, supporting the contribution of group-level modeling.
  • Forecast market-state conditioning: We replace the forecast market-state input to the residual predictor with zeros throughout pretraining while retaining the market-state prediction head and its loss. The ablated model is pretrained from scratch and subsequently fine-tuned using the same downstream protocol. Removing conditioning reduces IRR across all three datasets, with substantial SR declines on NASDAQ and NYSE.
  • Residual-aware query: We retain the same pretrained weights as Full but replace the stock-query input [Z,R] with [Z,0] throughout downstream fine-tuning and evaluation. The readout is trained from scratch, the transferred modules are jointly fine-tuned, and the query projection dimensionality is preserved. This reduces both IRR and SR across all three datasets, demonstrating the contribution of residual information to stock-conditioned factor retrieval.

Together, these experiments provide complementary evidence for HiMarket's hierarchical design. The matched encoder comparison supports the effectiveness of its pretraining scheme, while the component ablations distinguish the contributions of adaptive grouping, group-level representations, forecast market-state conditioning, and residual-aware queries.

AuthorsOpenReview ↗

Response to Reviewer wXxA (Part 2/2)

Q2. Baseline sources and comparability

Were all baseline results in Table 2 obtained using identical data splits, input features, preprocessing procedures, stock-availability masks, and portfolio evaluation rules? If some results were taken from previous studies, please clarify their sources and comparability.

Response. We follow the SSPT benchmark for NASDAQ and NYSE and adopt the baseline and Market results from Table 2 of SSPT. All CSMD300 baseline results are obtained from our own experiments. HiMarket is trained and evaluated by us on all three datasets.

We will identify result sources by dataset in the table caption and describe the corresponding experimental settings in the appendix.

The U.S. evaluation follows SSPT's chronological benchmark and daily equal-weighted Top-5 close-to-close protocol. The local CSMD300 experiments share the same frozen panel, chronological splits, and portfolio evaluation rules, with model-specific feature processing. The appendix will specify the input features, preprocessing, and stock-availability rules. We have also added local supplementary comparisons under a common execution and fee protocol, as discussed in Q3.

Q3. Stock eligibility, missing labels, turnover, and costs

Is stock eligibility for the daily Top-5 portfolio determined exclusively using information available before the target trading day? In particular, how is missing next-day return-label availability handled without affecting the eligible universe at decision time? Can the authors also report portfolio turnover and performance under nonzero transaction costs?

Response. The original portfolio evaluator required an available next-day return label before a stock could enter the Top-5 holdings. This makes the evaluation universe depend on future label availability. We have added an evaluation in which historical observations determine eligibility, return-label availability is used for retrospective prediction metrics, and opening quotes determine whether orders can execute.

For the executable-price backtest, inputs end at the close of day t−1, orders execute at the open of t, and rebalancing occurs at the open of t+1. We reuse the original prediction scores. A missing opening quote prevents the corresponding order from executing. Unspent funds remain in cash, unsold holdings are retained, and existing positions are valued using the last available quote. Orders, cash, holdings, and fees are recorded in a daily ledger.

The table reports HiMarket's results under this open-to-open schedule. Each performance cell gives cumulative IRR / annualized SR.

DatasetZero feesBuy 5 / sell 15 bps
NASDAQ49.39% / 1.72828.71% / 1.044
NYSE55.50% / 2.3599.55% / 0.568
CSMD30039.64% / 1.82917.12% / 1.001

The 5/15 bps buy/sell schedule follows the Qlib CSI300 example and is applied here as a common cost-sensitivity scenario across markets. Fees are charged on actual traded amounts, including initial entry, rebalancing, and final liquidation. Under this fee schedule, HiMarket has positive net returns and SR on all three datasets.

Thank you for the response and additional experiments. The new ablations largely address my concerns about the individual components. I also appreciate the authors' acknowledgment of the future-label issue in the original evaluation and the revised backtest.

However, I will maintain my rating of 2 (Weak rejection) for two reasons. First, under the revised protocol with transaction costs, HiMarket's advantage is less convincing. On CSMD300, StockMixer achieves higher returns and Sharpe ratio (18.22% / 1.083 vs. 17.12% / 1.001). On NYSE, HiMarket only slightly outperforms HIST (9.55% / 0.568 vs. 9.46% / 0.517). Second, the new component ablations still use the original zero-cost, close-to-close protocol. They support the proposed design, but do not establish whether the same gains hold under the corrected evaluation. The response strengthens the paper, but the remaining evidence does not change my overall assessment.

Reviewer_fKU6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HiMarket splits stock representations into global, learned-group, and residual components. It pretrains by predicting future shared states and stock residuals, then fine-tunes a stock-specific attention readout for return ranking. The evaluation uses daily top-five portfolios on NASDAQ, NYSE, and CSMD300.

Critical strengths

Conditioning residual prediction on shared market states gives pretraining a stock-specific target that fits the ranking task.

Critical weaknesses

The plain-JEPA ablation changes both pretraining and readout, so Table 3 does not isolate the benefit of the hierarchical decomposition or market-conditioned residual prediction.

Questions for authors
  1. What happens when plain JEPA uses the same factor-query head, and when the hierarchical model predicts residuals without conditioning on forecast shared states?
  2. Do the main gains survive repeated seeds and paired uncertainty estimates over test dates, especially the small NASDAQ return difference?
1 comment

We thank the reviewer for recognizing the alignment between conditioned residual prediction and stock ranking. We have added controlled ablations that hold the downstream architecture fixed, together with repeated-seed comparisons, to strengthen the evidence for HiMarket's design.

Q1. Plain JEPA with the same factor-query head, and residual prediction without forecast shared states

1. Plain JEPA with the same factor-query head. Both methods use the same encoder backbone. Plain JEPA predicts future full-stock latents; HiMarket trains the encoder through hierarchical decomposition and shared-state-conditioned residual prediction. We transfer the pretrained encoder weights from each method into identical downstream ranking architectures. The hierarchical aggregator, statistics projection, and factor-query head are initialized identically in both arms, then jointly fine-tuned with their respective encoders using the same training and validation-selection rules. Results on CSMD300 are:

Pretrained encoderIRRSR
HiMarket25.02%1.807
Plain JEPA\underline{4.84\\%}0.581‾\underline{0.581}

With the same factor-query head, HiMarket pretraining improves cumulative IRR by 20.18 percentage points and SR by 1.226. This supports the contribution of HiMarket's pretraining with the readout held fixed. This pair transfers only the encoder; Full in the next comparison also transfers the pretrained aggregator and statistics projection.

2. Removing forecast conditioning. We replace the forecast shared-state input to the residual predictor with zeros throughout pretraining, retaining the hierarchical architecture, shared-state prediction head, and loss terms. We pretrain this variant from scratch and use the same downstream architecture and fine-tuning protocol as Full. Results on CSMD300 are:

SettingIRRSR
HiMarket Full45.54%2.394
Without forecast shared-state input\underline{31.17\\%}2.354‾\underline{2.354}

Forecast conditioning increases cumulative IRR by 14.37 percentage points, with Full also attaining a higher SR. Together, these comparisons support the benefits of HiMarket pretraining and forecast-conditioned residual prediction under matched downstream architectures.

Q2. Repeated-seed evaluation

Repeated seeds. Across HiMarket, SSPT-comb, StockMixer, and MASTER, HiMarket ranks first in mean SR on NASDAQ and NYSE and second on CSMD300, while ranking first or second in mean IRR on all three datasets. We include all runs for seeds 1111, 2222, and 3333, with independent HiMarket pretraining and fine-tuning and validation-based checkpoint selection. Entries are mean ± sample standard deviation. IRR is cumulative return (SD in percentage points); SR is annualized. Bold and underlined entries indicate the best and second-best mean performance for each metric within each dataset.

NASDAQ

ModelIRRSR
HiMarket\underline{61.15\\% \pm 28.01}2.296 ± 0.519
SSPT-comb48.35% ± 32.721.478 ± 0.713
StockMixer62.93% ± 31.092.254±0.443‾\underline{2.254 \pm 0.443}
MASTER50.00% ± 4.841.561 ± 0.145

NYSE

ModelIRRSR
HiMarket51.04% ± 28.362.000 ± 0.809
SSPT-comb\underline{16.34\\% \pm 21.39}0.683±0.736‾\underline{0.683 \pm 0.736}
StockMixer9.55% ± 13.560.496 ± 0.596
MASTER9.71% ± 4.760.560 ± 0.172

CSMD300

ModelIRRSR
HiMarket30.13% ± 16.701.717±0.763‾\underline{1.717 \pm 0.763}
SSPT-comb0.96% ± 16.720.155 ± 0.839
StockMixer10.95% ± 20.920.828 ± 1.077
MASTER\underline{24.55\\% \pm 13.80}3.280 ± 0.528

The NASDAQ and NYSE baseline results in Table 2 are taken from SSPT, whereas the three-seed results reported here are obtained from our local adaptations. These repeated-seed results support HiMarket's competitive return and risk-adjusted performance across the three datasets. Its mean IRR and SR exceed local SSPT-comb on every dataset.

These additional experiments strengthen the empirical support for HiMarket: matched-head comparisons demonstrate the benefit of its pretraining, the conditioning ablation supports forecasting stock residuals with shared states, and repeated-seed evaluations show top-two mean IRR and SR among the four compared methods across all three datasets.

Reviewer_df8cOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiMarket, a hierarchical predictive framework for learning stock representations and performing cross-sectional stock selection. The model decomposes stock embeddings into global market, dynamic group, and stock-specific residual components. During JEPA pretraining, it first predicts future shared market states and then uses these predictions to condition future stock residual prediction. For downstream selection, stock-specific queries retrieve historical global and group information, and the model is fine-tuned using realized return rankings. Experiments cover NASDAQ, NYSE, and CSMD300.

Critical strengths
  1. Separating market-wide, group-level, and stock-specific information is reasonable, and using predicted market states to guide stock representation learning offers an interesting application of JEPA to stock selection.

  2. The model performs well across the three datasets, achieving the best results in five of the six comparisons, while the ablations support the overall benefits of pretraining, joint fine-tuning, and the stock scoring module.

Critical weaknesses
  1. JEPA-based market representation learning has already been explored by Mahns et al. (2026), and the separation of shared and stock-specific information is related to HIST and HireVAE. The paper makes targeted extensions, but without direct comparisons with these related models, the added value remains unclear. The hierarchy also contains several modules whose motivation and necessity are not sufficiently established by the current explanations and ablations.

  2. The model introduces several hyperparameters, including the number of groups, assignment temperature, and multiple loss weights. The appendix provides their values but no sensitivity analysis. It is therefore unclear whether performance depends on a particular parameter combination or remains stable when these settings change.

  3. The test periods are too limited to establish the model’s ability to handle changing market conditions. Both U.S. datasets are evaluated only in 2017, while the CSMD300 test period covers less than five months. These results do not establish whether the gains persist across different market regimes.

  4. The backtest uses close-to-close returns with zero transaction costs and does not account for slippage or execution delays. Predictions use the previous day’s complete price and volume information, while entry is assumed at that same day’s closing price. This is an idealized execution assumption. For a daily rebalanced Top-5 strategy, the current results provide limited evidence of practical trading value.

Questions for authors
  1. Does HiMarket retain its advantage over related hierarchical models such as HIST and HireVAE when evaluated using the same data and stock selection protocol?

  2. How were the number of groups, assignment temperature, and main loss weights chosen? In particular, does performance remain stable when the group count and pretraining loss weights change?

  3. Please specify the date ranges used for pretraining and supervised fine-tuning, and explain how complete pretraining windows are assigned to the training, validation, and test splits.

  4. Do the main findings hold in other test years or rolling evaluation periods, particularly under different volatility levels and market trends?

  5. Does the advantage over the baselines remain after including reasonable transaction costs and using execution prices available after the input information becomes observable?

7 comments
AuthorsOpenReview ↗

Response to Reviewer df8c (Part 1/4)

We thank the reviewer for the careful and constructive feedback. We address the concerns below with additional experiments and clearer descriptions of the architectural contribution and evaluation protocols.

Architectural contribution

HiMarket extends JEPA-based market representation learning to model shared market dynamics and stock-specific variation jointly for cross-sectional stock selection. Mahns et al. (2026) combine factor-token aggregation with temporal JEPA to learn a daily market-state representation. Cross-sectional selection additionally requires distinguishing how individual stocks evolve within the same market context. HiMarket addresses this requirement by decomposing stock representations into global, group, and stock-specific residual components, then conditioning future stock-residual prediction on forecast shared market states. A residual-aware, stock-conditioned factor-query readout retrieves relevant historical global and group context to adapt these pretrained representations to stock ranking.

Controlled Comparisons Across Three Datasets

Response. We extend the controlled comparisons to NASDAQ, NYSE, and CSMD300. All experiments use identical chronological splits, input processing, and validation-based checkpoint selection, following the zero-fee, daily equal-weighted Top-5 close-to-close protocol. We distinguish encoder comparisons from component ablations because they use different transfer settings.

1. Pretrained Encoder Comparison.

Only the pretrained encoder is transferred; all downstream modules share identical architectures and initializations.

EncoderNASDAQNYSECSMD300
Plain JEPA0.64/2.380.18/0.730.05/0.58
HiMarket0.84/2.640.57/1.970.25/1.81

HiMarket outperforms Plain JEPA in both IRR and SR across all datasets, supporting its pretraining scheme under matched downstream conditions.

2. Component Ablations.

These experiments transfer the pretrained encoder, hierarchical aggregator, and statistics projection, while training a new stock-ranking readout. Full is the reference.

SettingNASDAQNYSECSMD300
Full0.83/2.630.83/2.780.46/2.39
Fixed Grouping0.43/1.850.37/1.610.14/1.10
w/o Group Information0.61/1.89-0.17/-0.600.37/2.51
w/o Forecast Conditioning0.72/1.650.35/1.450.31/2.35
w/o Residual in Query0.32/1.940.71/2.16-0.12/-0.77
  • Dynamic grouping: Fixed Grouping retains the group branch but uses predefined, unchanged assignments throughout both stages. Its lower IRR and SR support adaptive grouping.
  • Group information: This variant removes the group branch during both pretraining and fine-tuning. IRR decreases on all datasets, while SR decreases on NASDAQ and NYSE.
  • Forecast conditioning: We zero the forecast market-state input throughout pretraining, retain its prediction head and loss, and retrain from scratch. IRR decreases across all datasets.
  • Residual-aware query: Using Full's pretrained weights, we replace [Z,R] with [Z,0] throughout downstream fine-tuning and evaluation. The readout is retrained, reducing both IRR and SR across all datasets.

Together, these results support the contributions of HiMarket's pretraining, adaptive grouping, group-level information, forecast conditioning, and residual-aware queries.

We address your five questions below. IRR denotes cumulative investment return ratio, and SR denotes annualized Sharpe ratio. Bold and underlined values identify the best and second-best results within each ablation table, or within each dataset, metric, and fee setting in the model comparisons.

Q1. Related hierarchical-model comparisons

Does HiMarket retain its advantage over related hierarchical models such as HIST and HireVAE when evaluated using the same data and stock selection protocol?

Response. HiMarket outperforms HIST and HireVAE on NYSE and CSMD300, and HireVAE on NASDAQ, in both IRR and SR. The comparisons use matched test dates, stock universes, availability masks, and the original zero-fee, daily equal-weighted Top-5 close-to-close protocol.

DatasetModelIRRSR
NASDAQHiMarket (ours)82.64%‾\underline{82.64\%}2.631‾\underline{2.631}
NASDAQHIST108.07%2.730
NASDAQHireVAE72.61%1.957
NYSEHiMarket (ours)83.39%2.777
NYSEHIST52.74%‾\underline{52.74\%}1.792‾\underline{1.792}
NYSEHireVAE-3.14%-0.005
CSMD300HiMarket (ours)45.54%2.394
CSMD300HIST26.80%‾\underline{26.80\%}1.771‾\underline{1.771}
CSMD300HireVAE6.97%0.610
AuthorsOpenReview ↗

Response to Reviewer df8c (Part 2/4)

Q2. Hyperparameter sensitivity

How were the number of groups, assignment temperature, and main loss weights chosen? In particular, does performance remain stable when the group count and pretraining loss weights change?

Response. The default hyperparameters were selected through validation-set search. We have added sensitivity analyses for the number of groups, assignment temperature, and the market-state, relation, and diversity loss weights on all three datasets, evaluating five values for each hyperparameter. The results favor a moderate group count, while extreme temperature or loss-weight settings can reduce performance. The complete IRR and SR curves are provided in Appendix C, Figure 2(a) and (b), respectively, on page 18 of the revised manuscript.

Q3. Data splits and pretraining-window assignment

Please specify the date ranges used for pretraining and supervised fine-tuning, and explain how complete pretraining windows are assigned to the training, validation, and test splits.

Response. Pretraining and supervised fine-tuning use the same training period; validation is used for model selection, and the test split is reserved for final evaluation. The effective date ranges below are inclusive. Samples are indexed by trading day t; for supervised fine-tuning, t is the return-label date.

DatasetTraining dates (both stages)Pretraining validation datesFine-tuning validation datesTest dates
NASDAQ2013-02-19 to 2015-12-312016-01-14 to 2016-12-302016-01-04 to 2016-12-302017-01-03 to 2017-12-11
NYSE2013-02-19 to 2015-12-312016-01-14 to 2016-12-302016-01-04 to 2016-12-302017-01-03 to 2017-12-08
CSMD3002021-04-07 to 2024-03-142024-03-27 to 2024-08-072024-03-15 to 2024-08-072024-08-08 to 2024-12-31

Each pretraining sample contains the 32 trading days preceding t, from t−32 to t−1: the first 24 days are visible context and the final 8 days are prediction targets. Training windows lie entirely within the training partition. For pretraining validation, all eight prediction-target days must lie within the validation partition; earlier visible context can come from the preceding training period. This requirement explains the later start of pretraining validation.

Both stages update parameters only on the training split. Fine-tuning uses the preceding 32 trading days to predict the close-to-close return on day t, and each supervised sample is assigned to a partition by its return-label date. Validation is used only for checkpoint selection and early stopping, without parameter updates. No test-period windows are used for pretraining, checkpoint selection, or fine-tuning.

Q4. Additional test years and rolling evaluation

Do the main findings hold in other test years or rolling evaluation periods, particularly under different volatility levels and market trends?

Response. HiMarket achieves positive returns in all three additional NASDAQ test periods in 2015–2016, with a cumulative IRR of 95.16% and an SR of 1.567 over 380 out-of-sample trading days. This expanding-window evaluation provides additional temporal evidence beyond the original 2017 NASDAQ test period.

In each fold, we pretrain and fine-tune the model from scratch, reusing the submitted hyperparameter configuration. The training period expands over time, followed by six months for validation and six months for testing. Checkpoints are selected using each fold’s validation period.

FoldTraining endValidation periodTest periodIRRSR
R12014-12-312015-01-02 to 2015-06-302015-07-01 to 2015-12-3121.45%1.130
R22015-06-302015-07-01 to 2015-12-312016-01-04 to 2016-06-3016.30%1.289
R32015-12-312016-01-04 to 2016-06-302016-07-01 to 2016-12-3038.17%2.871

All folds use the original daily equal-weighted Top-5 close-to-close protocol with zero transaction costs. Aggregate IRR compounds the chronologically concatenated daily returns from the three non-overlapping test periods. SR uses the same 252-trading-day annualization as the main evaluation.

AuthorsOpenReview ↗

Response to Reviewer df8c (Part 3/4)

Q5. Transaction costs and execution after signal availability

Does the advantage over the baselines remain after including reasonable transaction costs and using execution prices available after the input information becomes observable?

Response. HiMarket achieves positive net backtested returns on all three datasets under the specified transaction fees in both evaluation protocols. We report two complementary evaluations: adding fees to the original close-to-close benchmark and using next-open execution. Both use daily equal-weighted Top-5 portfolios, with zero fees or 5 bps buy fees and 15 bps sell fees. Fees apply to actual traded amounts at entry, rebalancing, and final liquidation. Each entry reports cumulative IRR / annualized SR.

1. Transaction costs under the original close-to-close protocol

With fees, HiMarket ranks first in both IRR and SR on NYSE and CSMD300; on NASDAQ, HiMarket attains the best Sharpe ratio while HIST yields the highest IRR. We assess cost sensitivity under the original SSPT close-to-close benchmark, retaining the original predictions, evaluation dates, availability masks, and Top-5 rules. Zero-fee HiMarket results reproduce the main-table values. Because inputs include the starting close, execution at that close is not guaranteed; next-open execution is evaluated separately below.

DatasetModelZero feesBuy 5 / sell 15 bps
NASDAQHiMarket (ours)82.64%‾\underline{82.64\%} / 2.631‾\underline{2.631}34.12%‾\underline{34.12\%} / 1.340
NASDAQHIST108.07% / 2.73038.33% / 1.294‾\underline{1.294}
NASDAQHireVAE72.61% / 1.95718.77% / 0.728
NASDAQSSPT-comb17.52% / 0.748−0.65% / 0.115
NASDAQStockMixer60.39% / 2.08324.87% / 1.052
NASDAQTimeMixer39.10% / 1.219−3.03% / 0.067
NYSEHiMarket (ours)83.39% / 2.77728.77% / 1.226
NYSEHIST52.74%‾\underline{52.74\%} / 1.792‾\underline{1.792}10.45%‾\underline{10.45\%} / 0.523‾\underline{0.523}
NYSEHireVAE−3.14% / −0.005−24.93% / −1.065
NYSESSPT-comb−0.94% / 0.081−18.01% / −0.739
NYSEStockMixer7.15% / 0.414−3.46% / −0.017
NYSETimeMixer33.82% / 1.140−8.39% / −0.140
CSMD300HiMarket (ours)45.54% / 2.39421.82% / 1.361
CSMD300HIST26.80% / 1.771‾\underline{1.771}8.32% / 0.720
CSMD300HireVAE6.97% / 0.610−1.99% / 0.108
CSMD300SSPT-comb−5.34% / −0.041−19.08% / −0.860
CSMD300StockMixer30.08%‾\underline{30.08\%} / 1.70213.00%‾\underline{13.00\%} / 0.911‾\underline{0.911}
CSMD300TimeMixer13.87% / 1.076−0.16% / 0.179

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.