acceptodds
Under review as a conference paper at ICLR 2027

RAVEN: A Regime-Aware Variable-context Expert Network for Financial Time Series Forecasting

Abstract

Financial time series forecasting presents structural challenges absent from standard benchmarks. Log-returns are non-stationary, exhibit exceptionally low signal-to-noise ratios, and follow regime-dependent temporal dependencies. Recent forecasting models typically rely on a fixed context window, which is mismatched with the time-varying optimal look-back of log-return processes. To address this limitation, we propose the Regime-Aware Variable-context Expert Network (RAVEN), a Mixture-of-Experts framework that adaptively determines the temporal context for each input sample. Specifically, RAVEN constructs a hierarchy of nested contiguous windows determined by the data itself. The model scores historical patches in reverse chronological order and applies Cumulative Importance Thresholding (CIT) to route each prefix to a scale-specialized expert. To overcome representation redundancy caused by overlapping inputs, it penalizes pairwise expert cosine similarity prior to aggregation. Meanwhile, a parallel global compression branch is incorporated to preserve macro-level temporal coherence. Experiments on cumulative log-return prediction (HS300, S&P500) and retail fund sales forecasting demonstrate that RAVEN achieves state-of-the-art performance, which translates into consistent risk-adjusted alpha gains in market backtests. Beyond financial applications, evaluations on four PEMS benchmarks confirm that its adaptive look-back generalizes effectively to broader physical dynamic systems.

Primary area:
learning on time series and dynamical systems
Keywords:
Financial Time Series, Mixture-of-Experts, Adaptive Context Selection, Non-stationary Financial Data

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Gr1iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper argues that fixed look-back windows mismatch non-stationary financial log-returns. A Mixture-of-Experts framework called RAVEN was proposed. RAVEN scores patches in reverse chronological order, and Cumulative Importance Thresholding (CIT) cuts nested contiguous contexts at thresholds, each fed to its own expert. CAW reduces redundancy among experts, while GCR sees the full window. Experiments on financial and traffic time series validate the performance of the proposed RAVEN.

Critical strengths
  1. The idea is simple and sensible: per-sample look-back with chronological contiguity preserved.

  2. Broad evaluation (two equity markets, fund sales, traffic) with a rolling out-of-sample protocol, ablations, sensitivity analyses, efficiency results, and routing visualisation.

  3. Clear writing and a candid limitations appendix.

Critical weaknesses
  1. The paper does not adequately explain how gradients propagate through the hard thresholding operation in CIT to train the scoring function (\phi). Although the entropy regularizer can update patch importance scores, it is unclear whether and how the forecasting loss provides predictive supervision for context selection. The authors should clarify the gradient flow and specify whether a differentiable approximation or an alternative optimization strategy is used.

  2. The "adaptive'' and "regime-aware'' claims are weakly supported. With only 7--8 patches, CIT effectively chooses among a few prefix lengths. Since Lent\mathcal{L}_{\text{ent}} pushes the importance distribution toward uniform, CIT approaches fixed windows (Ψi≈i/N\Psi_i \approx i/N), and the empirical distribution of ℓk\ell_k is not reported. Figure~5 shows only that scores vary across inputs, which any input-dependent scorer would do.

  3. The experimental evaluation does not fully isolate the contribution of adaptive context selection and expert routing. Although the w/o AR ablation replaces adaptive routing and learned context selection with predefined windows, the construction of these windows and the individual contributions of the two components could be described and evaluated more clearly. The authors should consider comparisons with fixed nested windows combined with learned soft expert weights, recency-biased attention, and multi-window forecasting ensembles. In addition, comparisons with more relevant financial forecasting methods and time-series mixture-of-experts models would help establish the method’s advantages. The paper should also clarify whether the baselines were tuned under comparable experimental budgets.

  4. The reported improvements over strong baselines are relatively modest in absolute terms, particularly in MSE. However, the paper does not report sufficient repeated-run statistics or significance tests to establish the robustness of these gains. The authors should provide results across multiple random seeds, including means and standard deviations, and conduct appropriate statistical tests for the main comparisons.

  5. No Sharpe ratio, information ratio, drawdown, or turnover is reported for the financial time series backtest, so the abstract's "risk-adjusted alpha gains'' has weak supporting metrics.

Questions for authors

Please see the weaknesses.

3 comments

Thank you for the constructive assessment. We clarify the mechanism and provide additional evidence below.

Q1. Predictive supervision through CIT.

  • Patch scores select windows and weight their embeddings. Eq. (5) expands as:

    • Ekw=[exp⁡(si)∑j∈Gkexp⁡(sj)⋅ei]i∈GkE_{k}^{w} = \left[ \frac{\exp(s_i)}{\sum_{j\in G_k}\exp(s_j)} \cdot e_i \right]_{i\in G_k}
    • Uk=AttnBlockk(Ekw)U_{k} = AttnBlock_{k} \left(E_{k}^{w}\right)
    • hk=FFNBlockk(Uk)h_{k} = FFNBlock_{k} \left(U_{k}\right)

    The blocks include residual and normalization operations. Indices are fixed during backpropagation. Forecasting gradients reach the scorer through weighted embeddings and within-window softmax, without differentiating through the hard threshold or using a straight-through estimator. The scorer thus receives predictive supervision alongside entropy regularization. We will clarify this in the revised PDF.

Q2. Adaptive contexts, patch count, and regime awareness.

  • Uniform scores yield fixed lengths 2/4/6 for seven patches. Entropy regularization encourages uniformity but does not enforce it under the joint objective. Table 1 shows multiple lengths per expert, with the most frequent lengths being 1/3/5 on HS300 and 2/3/5 on S&P500.

    Table 1. Selected window lengths (%).

    Market/expert123456
    HS300/E166.429.73.8.100
    HS300/E27.920.436.428.96.1.3
    HS300/E3.11.82.710.850.534.1
    S&P500/E147.749.23.0.100
    S&P500/E26.819.637.529.36.4.4
    S&P500/E30.82.413.944.138.8
  • Table A.4 evaluates N=6/7/10/15, yielding HS300 Corr of .0372/.0390/.0384/.0384. Additional fixed-window experiments show that the best length varies across years in both markets.

  • These distributions support input-adaptive selection. The ablations below assess its predictive benefit.

Q3. Component isolation, baselines, and tuning.

  • Fixed windows. In w/o AR, we remove CIT and patch importance learning. Assuming equal patch importance, we predefine three nested windows containing the most recent 2/4/6 patches for Experts 1–3. These windows are fixed across inputs, while expert encoders, learned aggregation, and other components are retained.

  • The w/o CAW variant separately evaluates correlation-aware aggregation. Global confidence parameters αk\alpha_k are learned from equal initialization and fixed after training. Final weights remain input-dependent through CAW.

    Table 2. HS300 ablations (mean ± SD).

    VariantCorr ↑
    RAVEN.0391±.0008
    w/o AR.0351±.0020
    w/o CAW.0357±.0019

    Both controls reduce overall Corr, supporting the contributions of adaptive contexts and correlation-aware aggregation. We'll add full results in revised pdf.

  • Baselines. Patch-Ensemble already provides a multi-window ensemble comparison. We add financial baselines and recent time-series methods, including LightGBM(with raw input and also with Alpha360 factors), TFPS, TimeSqueeze.

    Table 3. Additional HS300 comparisons.

    ModelCorr ↑CS-IC ↑
    RAVEN.0391.0576
    LightGBM (raw).0171.0249
    LightGBM (Alpha360).0254.0406
    TRA.0233.0352
    TFPS.0276.0436
    TimeSqueeze-base.0315.0464

    Table 3 also shows that RAVEN performs best on both metrics among these methods.

  • Tuning protocol. We tune all reproduced SOTA methods using their official implementations, released scripts, and hyperparameter configurations as references. We will document the configurations explored and the model-selection procedure in the revision.

Q4. Reproducibility and gain magnitude.

  • Table 4 reports 10-seed mean ± SD for 2020–2024. RAVEN achieves better means and lower seed variation than MLF on all reported metrics in both markets.

    Table 4. Ten-seed comparison.

    MarketModelCorr ↑MSE ↓CS-IC ↑
    HS300RAVEN.0391±.0008.9904±.0009.0576±.0006
    HS300MLF.0357±.0021.9932±.0025.0504±.0018
    S&P500RAVEN.0364±.00081.001±.0009.0473±.0006
    S&P500MLF.0302±.00221.016±.0028.0387±.0019
  • Corr improves by 9.5%/20.5% and CS-IC by 14.3%/22.2% on HS300/S&P500. Full annual results will be added to the appendix.

Q5. Risk-adjusted performance.

  • HS300 Qlib backtests use the same strategy and transaction-cost settings across models.

    Table 5. Portfolio performance. MDD is drawdown magnitude.

ModelSharpe ↑IR ↑MDD ↓Turnover
RAVEN1.9871.79716.10%15.11%
Patch-Ensemble1.0381.02823.91%17.02%
PatchTST1.1051.06621.03%18.96%
PathFormer1.4181.31214.29%14.24%
PathConcat1.4251.30717.70%15.77%
FiLM1.3011.33419.75%16.27%
NHits1.5371.40317.61%15.33%
MLF1.7231.66317.56%16.43%
  • Sharpe/IR exceed runner-up MLF by 15.3%/8.1%. RAVEN has the second-lowest drawdown among evaluated models, with turnover within the baseline range.

Thank you for your response and for providing the detailed experimental settings and results that were missing from the original submission. While your responses have addressed some of my concerns, I have decided to maintain my original rating, considering the explanations of the model's underlying mechanisms and its performance advantages in financial time-series forecasting.

Thank you for your feedback.

We would appreciate any further feedback on which aspects of the underlying mechanisms or performance advantages remain insufficiently addressed. Your suggestions and guidance would help us provide more clarifications or additional experimental results, and also further improve the quality of the paper.

Thank you again for your time and consideration.

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

This paper proposes RAVEN, a regime-aware variable-context mixture-of-experts network for financial time series forecasting. It addresses the problem that existing forecasting models use a fixed look-back window, which does not match the time-varying dynamics of non-stationary financial data. To solve this, the authors use Cumulative Importance Thresholding (CIT) to build nested, variable-length windows for experts at different time scales, Correlation-Aware Weighting (CAW) to reduce redundancy among experts, and a global branch to keep information from the full sequence. Experiments on stock return prediction (HS300, S&P500), fund sales prediction, and four PEMS traffic datasets show better results than the baselines, together with gains in portfolio backtests.

Critical strengths
  • The paper proposes a clear method that adaptively selects the look-back window for each sample through nested contiguous contexts, which keeps the time order better than token-level MoE routing.
  • RAVEN shows consistent improvements over the baselines on stock return prediction, fund sales prediction, and traffic forecasting, and the gains are also reflected in portfolio backtests.
Critical weaknesses
  • According to Eq. (4)–(5), the windows are built by hard thresholds, and the encoder inputs are not weighted by the importance scores, so the forecasting loss does not seem to pass gradients to the scoring function. In that case, the scores are mainly driven by the entropy regularizer toward a uniform distribution, which would make the window lengths nearly fixed. The raw routing confidence αkα_k is also never defined.
  • The reported gains are very small (e.g., Corr 0.0390 vs. 0.0357 over MLF) and are based on single runs without standard deviations or significance tests.
  • Important baselines are missing, including financial models such as MoE models for time series such as TFPS and Time-MoE, and adaptive patching methods such as TimeSqueeze.
Questions for authors

Please refer to weakness.

1 comment

We appreciate the reviewer's careful assessment. Below, we clarify CIT's training mechanism and present additional experiments addressing the reliability and relevance of the reported gains.

Q1. How does CIT receive predictive supervision, and does entropy regularization lead to fixed windows?

  • Predictive supervision. Patch scores both select context windows and weight the selected embeddings. Eq. (5) expands as:

    • Ekw=[exp⁡(si)∑j∈Gkexp⁡(sj)⋅ei]i∈GkE_{k}^{w} = \left[ \frac{\exp(s_i)}{\sum_{j\in G_k}\exp(s_j)} \cdot e_i \right]_{i\in G_k}
    • Uk=AttnBlockk(Ekw)U_{k} = AttnBlock_{k} \left(E_{k}^{w}\right)
    • hk=FFNBlockk(Uk)h_{k} = FFNBlock_{k} \left(U_{k}\right)
  • The blocks include residual and normalization operations. Selected indices are fixed during backpropagation. Predictive gradients reach the scorer through the weighted embeddings and within-window softmax, without differentiating through the hard threshold or using a straight-through estimator. The scorer therefore receives forecasting supervision alongside entropy regularization. We will clarify this computation in the revised PDF.

  • Entropy regularization. This term encourages more uniform scores but does not enforce uniformity under the joint training objective. With seven patches, uniform scores yield fixed lengths 2/4/6. Table 1 instead shows multiple selected lengths within every expert. The most frequent lengths are 1/3/5 on HS300 and 2/3/5 on S&P500. Thus, the learned selections do not collapse to fixed windows.

    Table 1. Selected window lengths (%). Length seven has zero frequency.

    Market/expert123456
    HS300/E166.429.73.8.100
    HS300/E27.920.436.428.96.1.3
    HS300/E3.11.82.710.850.534.1
    S&P500/E147.749.23.0.100
    S&P500/E26.819.637.529.36.4.4
    S&P500/E30.82.413.944.138.8
  • Expert confidence. The αk\alpha_k are global learnable confidence parameters, initialized at 1/K1/K and fixed after training. Final aggregation weights remain input-dependent through CAW's redundancy adjustment. We will state this definition explicitly.

Q2. Are the gains meaningful and reproducible?

  • Repeated experiments. We evaluate RAVEN and MLF, the strongest baseline in overall Corr, over 10 random seeds. Table 2 reports mean ± SD for the 2020–2024 evaluation. RAVEN achieves better means and lower standard deviations across all four metrics in both markets.

    Table 2. Ten-seed comparison (mean ± SD).

    MarketModelCorr ↑CS-IC ↑MSE ↓ICIR ↑
    HS300RAVEN.0391±.0008.0576±.0006.9904±.0009.3938±.0145
    HS300MLF.0357±.0021.0504±.0018.9932±.0025.3582±.0352
    S&P500RAVEN.0364±.0008.0473±.00061.001±.0009.5986±.0148
    S&P500MLF.0302±.0022.0387±.00191.016±.0028.5581±.0365
  • Gain magnitude. The absolute Corr differences correspond to relative improvements of 9.5% on HS300 and 20.5% on S&P500 over MLF. Cross-sectional IC, which measures predictive association across stocks, improves by 14.3% and 22.2%, respectively. These results extend the advantage beyond temporal Corr and show lower sensitivity to random initialization in both markets. Full annual results will be included in the appendix.

  • Financial relevance. We also assess whether the predictive gains translate into portfolio performance. The HS300 Qlib backtests use the same trading strategy and transaction-cost settings across models.

    Table 3. HS300 portfolio performance. MDD is drawdown magnitude.

    ModelSharpe ↑IR ↑MDD ↓Turnover
    RAVEN1.9871.79716.10%15.11%
    MLF1.7231.66317.56%16.43%
    PathFormer1.4181.31214.29%14.24%

    RAVEN improves Sharpe and IR over MLF, the runner-up on both metrics, by 15.3% and 8.1%. It achieves the second-lowest drawdown among all evaluated models, with turnover within the baseline range. These results provide evidence of improved risk-adjusted performance in the evaluated backtest.

Q3. How does RAVEN compare with additional relevant methods?

  • We expand the evaluation to include TFPS and TimeSqueeze-base, alongside the financial model TRA and LightGBM using raw inputs or Alpha360 features.

    Table 4. Additional HS300 comparisons.

    ModelCorr ↑CS-IC ↑MSE ↓ICIR ↑
    RAVEN.0391.0576.9904.3938
    LightGBM (raw).0171.02491.0058.2405
    LightGBM (Alpha360).0254.04061.0014.3201
    TRA.0233.03521.0021.3034
    TFPS.0276.0436.9972.3204
    TimeSqueeze-base.0315.0464.9951.3546
  • RAVEN performs best across all four metrics among these additional methods. The results strengthen the comparison with both recent time-series models and established financial forecasting baselines, including a feature-based LightGBM pipeline.

The revision will include these additional results and explicit descriptions of the scoring and aggregation computations.

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

This paper studies dynamic history-window selection in financial time series forecasting. The authors argue that financial returns have a low signal-to-noise ratio and time dependencies that keep changing, so a fixed history window may struggle to adapt to different market conditions.

To address this, the authors propose RAVEN, a mixture-of-experts framework for variable-context forecasting. The method uses learnable segment importance scores and a cumulative importance threshold mechanism (CIT) to build continuous, nested history windows, which are then processed by different experts. The model also uses correlation-aware weighting (CAW) to reduce redundancy among the experts' representations, and a global compressed representation branch (GCR) to retain the complete history. The paper reports performance comparisons, investment backtests, a component ablation, and an efficiency analysis on HS300, S&P500, fund sales, and four PEMS datasets.

Critical strengths
  1. The motivation is reasonable, and the method has a certain degree of novelty. The paper addresses dynamic history-window selection in non-stationary time series, uses CIT to build nested windows that preserve temporal order, and fuses information through different experts, CAW, and a global branch. The overall design is fairly intuitive, and the role of each component is relatively clear.

  2. The experiments cover a wide range, and the forecasting results are competitive. On HS300 and S&P500, RAVEN's overall Corr is 0.0390 and 0.0363 respectively, higher than MLF's 0.0357 and 0.0302. It also does well on fund sales forecasting and the four PEMS datasets, achieving the best result on 14 of the 16 PEMS metrics. The ablation in Table 5 and the other experiments further examine the main components, routing behavior, and computational efficiency, giving some support to the method's effectiveness.

Critical weaknesses
  1. The optimization mechanism of CIT is unclear, which affects how the core adaptive routing can be judged. Eqs. (3)–(4) use learnable segment scores and compare the cumulative weight against a fixed threshold to determine each expert's input window. Since window selection is a discrete operation, the prediction loss seemingly cannot update the segment importance scoring function through ordinary backpropagation. Although the paper claims end-to-end training, it does not state whether a differentiable approximation, a straight-through estimator, or some other gradient path is used. Entropy regularization can shape the score distribution, but it tends to raise the entropy of that distribution and cannot by itself explain how the scoring function learns window lengths that benefit prediction. In addition, Eq. (4) may produce empty windows, and the raw routing confidence in Eq. (7) is not clearly defined.

  2. The financial experiments do not yet adequately support the adaptive mechanism or the actual investment value. Figure 5 shows how routing changes across years and stocks, but it cannot yet prove that these window selections correspond to more useful forecasting history. The ablation in Table 5 generally supports the role of each component, but it lacks results on variation across random seeds or significance tests, and the Corr gains from some components are relatively small. Moreover, Appendix A.3 uses the per-stock time-series correlation as the main metric, while the backtest uses daily cross-sectional Top-K stock selection, and the two are not fully aligned. Although the paper reports a cumulative return advantage and applies a uniform transaction cost setting in the backtest, it does not adequately show portfolio risk metrics such as Sharpe ratio and maximum drawdown, so the conclusion about risk-adjusted returns still lacks direct evidence. The financial experiments also lack same-setting comparisons with some finance-specific models and strong traditional baselines, which limits how convincing the claim of best performance is.

Questions for authors
  1. The hard-threshold window selection in Eq. (4) is usually non-differentiable. How does the prediction loss pass gradients to the segment importance scoring function? Is a differentiable approximation, a straight-through estimator, or some other mechanism used? If not, how does the scoring function learn history lengths that are relevant to the prediction target?

  2. When the cumulative importance of the most recent segments already exceeds the minimum CIT threshold, how does the algorithm avoid empty windows? Also, how is the raw routing confidence in Eq. (7) computed, and can it provide an additional gradient path to the segment scoring function?

  3. Drawing on the existing experiments, can the authors explain why dynamically selected history windows suit the current forecasting tasks better than fixed windows? Were the performance gains of each component in Table 5 obtained from repeated experiments or tested for significance? Furthermore, why were more finance-specific forecasting methods or strong traditional baselines such as LightGBM not included in the comparison?

  4. In the rolling evaluation, how are the 10-day cumulative return labels that cross the training, validation, and test boundaries handled? Since the backtest uses cross-sectional stock selection, can the authors report cross-sectional IC or RankIC? In addition, can the existing cumulative return results support the conclusion that risk-adjusted returns improve, and are there results such as Sharpe ratio or maximum drawdown?

5 comments

Thank you for your constructive feedback. We explain these issues in detail below.

Q1. How does the forecasting loss supervise CIT?

  • Patch scores select windows and weight their embeddings. Eq. (5) expands as:

    • Ekw=[exp⁡(si)∑j∈Gkexp⁡(sj)⋅ei]i∈GkE_{k}^{w} = \left[ \frac{\exp(s_i)}{\sum_{j\in G_k}\exp(s_j)} \cdot e_i \right]_{i\in G_k}
    • Uk=AttnBlockk(Ekw)U_{k} = AttnBlock_{k} \left(E_{k}^{w}\right)
    • hk=FFNBlockk(Uk)h_{k} = FFNBlock_{k} \left(U_{k}\right)
  • The blocks include residual and normalization operations. Predictive gradients reach the scorer through weighted embeddings and within-window softmax. Indices are fixed during backpropagation. Gradients do not pass through the hard threshold Eq. (4). We will clarify this in the revised PDF.

Q2. Nonempty windows and expert confidence.

  • Every window includes the most recent patch, retaining only that patch if its importance exceeds the threshold. Windows remain nonempty, contiguous, and nested.
  • The αk\alpha_k are global learnable confidence parameters, initialized at 1/K1/K and fixed after training. Final weights remain input-dependent through CAW's redundancy adjustment.

Q3. Adaptive contexts, robustness, and baselines.

  • Adaptive contexts. Uniform importance gives fixed lengths 2/4/6 under Eq. (4). Table 1 shows multiple lengths per expert despite entropy regularization. The most frequent lengths are 1/3/5 on HS300 and 2/3/5 on S&P500.

    Table 1. Selected window lengths (%).

    Market/expert123456
    HS300/E166.429.73.8.100
    HS300/E27.920.436.428.96.1.3
    HS300/E3.11.82.710.850.534.1
    S&P500/E147.749.23.0.100
    S&P500/E26.819.637.529.36.4.4
    S&P500/E30.82.413.944.138.8
  • Fixed-window sensitivity. No tested length performs best in every year in either market (Table 2), motivating context adaptation. In addition, Fig. 1 also shows similar patterns.

    Table 2. Annual Corr for fixed window lengths.

MarketLength20202021202220232024
HS30060.0351.0298.0332.0305.0381
HS30090.0378.0346.0345.0325.0441
HS300120.0370.0326.0366.0346.0442
HS300150.0367.0321.0350.0309.0438
HS300180.0360.0305.0338.0310.0427
S&P50060.0340.0305.0328.0298.0319
S&P50090.0363.0333.0352.0312.0335
S&P500120.0369.0325.0344.0328.0354
S&P500150.0358.0320.0341.0312.0342
S&P500180.0348.0312.0332.0303.0330
  • Repeated ablations. The w/o AR variant uses predefined nested windows, retaining expert aggregation and other components.

    Table 3. HS300 ablations (mean ± SD).

    VariantCorr ↑
    RAVEN.0391±.0008
    w/o AR.0351±.0020
    w/o CAW.0357±.0019
    w/o GCR.0361±.0018
    w/o entropy loss.0372±.0019
    w/o diversity loss.0370±.0018
  • Robustness. Table 4 reports 10-seed mean ± SD for 2020–2024. CS-IC measures correlation across stocks, complementing temporal Corr. Over MLF, Corr improves by 9.5%/20.5% and CS-IC by 14.3%/22.2% on HS300/S&P500, with lower seed variation. Full annual results will be added to the appendix.

    Table 4. 10-seed comparison with MLF.

    MarketModelCorr ↑CS-IC ↑
    HS300RAVEN.0391±.0008.0576±.0006
    HS300MLF.0357±.0021.0504±.0018
    S&P500RAVEN.0364±.0008.0473±.0006
    S&P500MLF.0302±.0022.0387±.0019
  • Baseline coverage. We add LightGBM with raw inputs and Alpha360 factors, TRA, TFPS, and TimeSqueeze-base.

    Table 5. Additional HS300 comparisons.

    ModelCorr ↑CS-IC ↑MSE ↓ICIR ↑
    RAVEN.0391.0576.9904.3938
    LightGBM (raw).0171.02491.0058.2405
    LightGBM (Alpha360).0254.04061.0014.3201
    TRA.0233.03521.0021.3034
    TFPS.0276.0436.9972.3204
    TimeSqueeze-base.0315.0464.9951.3546

    RAVEN performs best across all four metrics among these recent time-series models and established financial baselines TRA and LightGBM.

Q4. Label boundaries and investment value.

  • For the 10-day return label, we remove the last 10 trading-day prediction samples from the preceding split at train–validation and validation–test boundaries. Retained label intervals do not cross these boundaries. Look-back inputs can use preceding splits' observations available at prediction time.

  • HS300 Qlib backtests use the same strategy and transaction-cost settings across models. MDD is drawdown magnitude.

    Table 6. HS300 portfolio performance.

    ModelSharpe ↑IR ↑MDD ↓Turnover
    RAVEN1.9871.79716.10%15.11%
    MLF1.7231.66317.56%16.43%
    PathFormer1.4181.31214.29%14.24%
    LightGBM (raw)0.3220.40636.88%20.39%
    LightGBM (Alpha360)1.2751.32920.86%19.17%
    TRA1.0371.00724.05%18.83%
    TFPS1.4371.36418.21%17.88%
    TimeSqueeze-base1.6461.63217.35%17.91%
  • RAVEN improves Sharpe/IR over second-best MLF by 15.3%/8.1% and has the second-lowest drawdown among evaluated models, supporting improved risk-adjusted performance.

Thank you for the detailed response and additional experiments. The new results address some of my previous concerns. However, I still have reservations about the CIT mechanism and whether the experiments sufficiently support the claimed benefits of adaptive window selection. I would appreciate further clarification and remain open to reconsidering my rating.

Thank you for the feedback. We’d like to further clarify the CIT mechanism and the benefits of adaptive window selection.

1. Explanation of the CIT mechanism.

Let eie_i be patch ii's embedding and θ\theta the scorer parameters. Patch NN is the most recent. Scores and normalized importance are

si=fθ(ei),pi=exp⁡(si)∑j=1Nexp⁡(sj).s_i=f_\theta(e_i),\qquad p_i=\frac{\exp(s_i)}{\sum_{j=1}^{N}\exp(s_j)}.

CIT accumulates importance backward from the newest patch. With threshold τk\tau_k,

G_k=\left\\{i\in\\{1,\ldots,N\\}:\sum_{j=i}^{N}p_j\leq\tau_k\right\\}.

The same scores weight embeddings within each selected window:

qk,i=exp⁡(si)∑j∈Gkexp⁡(sj),e~k,i=qk,iei,i∈Gk.q_{k,i}=\frac{\exp(s_i)}{\sum_{j\in G_k}\exp(s_j)},\qquad \tilde{e}_{k,i}=q_{k,i}e_i,\quad i\in G_k.

These embeddings enter the expert encoders and aggregation. During backpropagation, GkG_k is fixed, and

∂qk,i∂sj=qk,i(δij−qk,j),i,j∈Gk,\frac{\partial q_{k,i}}{\partial s_j}=q_{k,i}(\delta_{ij}-q_{k,j}),\qquad i,j\in G_k,

where δij\delta_{ij} is the Kronecker delta. L\mathcal{L} is the forecasting loss, contributions from all windows containing patch jj give

∂L∂sj=∑k:j∈Gk∑i∈Gk(∂L∂e~k,i)⊤ei,qk,i(δij−qk,j).\frac{\partial\mathcal{L}}{\partial s_j}=\sum_{k:j\in G_k}\sum_{i\in G_k}\left(\frac{\partial\mathcal{L}}{\partial\tilde{e}_{k,i}}\right)^\top e_i\\,q_{k,i}(\delta_{ij}-q_{k,j}).

The scorer is therefore updated through the chain rule:

∇θL=∑j=1N∂L∂sj∇θfθ(ej).\nabla_\theta\mathcal{L}=\sum_{j=1}^{N}\frac{\partial\mathcal{L}}{\partial s_j}\nabla_\theta f_\theta(e_j).

Thus, forecasting supervision reaches the scorer through weighted embeddings. Predictive gradients train patch weighting, and subsequent forward passes recompute CIT windows from the updated scores. We will clarify this distinction in the revision.

2. The benefits of adaptive context modeling

Fixed single windows.

Table 1 compares five fixed look-back lengths using metrics computed over the full 2020–2024 period. RAVEN outperforms every tested fixed length on all four metrics in both markets. Against the strongest fixed configuration, 120 days, Corr improves by 5.1%/5.8% on HS300/S&P500.

Table 1. Full-period performance.

MarketMethod / lengthCorr ↑CS-IC ↑MSE ↓ICIR ↑
HS300Fixed, 60.0342.0508.9938.3516
HS300Fixed, 90.0365.0538.9920.3742
HS300Fixed, 120.0372.0546.9915.3798
HS300Fixed, 150.0356.0522.9921.3715
HS300Fixed, 180.0348.0518.9931.3596
HS300RAVEN.0391.0576.9904.3938
S&P500Fixed, 60.0318.04211.0082.5356
S&P500Fixed, 90.0339.04471.0048.5642
S&P500Fixed, 120.0344.04521.0040.5725
S&P500Fixed, 150.0335.04431.0056.5568
S&P500Fixed, 180.0325.04281.0071.5438
S&P500RAVEN.0364.04731.0010.5986

Predefined multiple windows.

The w/o AR variant removes CIT mechanism, using nested windows of the most recent 2/4/6 patches. Expert encoders, learned expert aggregation, and other components are retained.

Table 2. HS300 comparison with predefined multiple windows.

VariantCorr, mean ± SD
RAVEN.0391±.0008
w/o AR: fixed 2/4/6.0351±.0020

RAVEN improves mean Corr by 11.4%.

Together, these comparisons support the adaptive context module over the tested fixed single-window and multi-window alternatives.

Selections do not collapse to 2/4/6.

With seven patches, uniform importance yields fixed lengths 2/4/6. Instead, every expert selects multiple lengths.

  • The most selected patch numbers in different experts of HS300 model are 1/3/5, with frequencies 66.4%/36.4%/50.5%.

  • The most selected patch numbers in different experts of S&P500 model are 2/3/5, with frequencies 49.2%/37.5%/44.1%.

The distributions demonstrate input-dependent selection (as shown in Table 1. Selected window lengths (%) in previous rebuttal content), while the fixed-window comparisons and ablation support the predictive benefit of the full adaptive context.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers and Area Chair

Dear reviewers and AC,

We sincerely thank all your reviewers for the constructive feedback. We have provided additional methodological clarifications and experiments addressing three shared concerns:

  • CIT optimization and adaptivity
  • Adaptive versus fixed historical contexts
  • Statistical reliability and financial relevance.

We summarize the key details below.

    1. CIT Mechanism and Gradient Propagation

CIT optimization. CIT constructs nested contexts using cumulative patch importance. Although the discrete context boundaries are not differentiable, the forecasting loss updates the importance scorer through differentiable weighting within each selected context:

\quad i \in \mathcal{G}_k.$$ The weighted embeddings are processed by the corresponding expert. Thus, predictive gradients reach the scorer through the weighted inputs. The most recent patch is always included to avoid empty contexts. **Expert aggregation.** The expert confidences $\alpha_k$ are initialized uniformly and learned during training. During inference, they remain fixed, while CAW adjusts the aggregation weights using input-dependent redundancy scores. - 2. Why Adaptive Historical Contexts Matter **Fixed-window comparisons.** We evaluate five fixed look-back lengths under the same forecasting setting. | Look-back | HS300 Corr ↑ | S&P500 Corr ↑ | |---|---:|---:| | 60 | 0.0342 | 0.0318 | | 90 | 0.0365 | 0.0339 | | 120 | 0.0372 | 0.0344 | | 150 | 0.0356 | 0.0335 | | 180 | 0.0348 | 0.0325 | | **RAVEN (adaptive)** | **0.0390** | **0.0363** | The best fixed-window length varies across evaluation years. RAVEN outperforms the best overall fixed-window configuration on both markets, supporting input-dependent context selection rather than a single predefined historical range. **Adaptive routing evidence.** The learned context lengths show nonuniform distributions across inputs. Frequently selected window lengths include 1, 3, and 5 patches on HS300 and 2, 3, and 5 patches on S&P500. The w/o AR ablation retains multiple predefined nested windows (2/4/6 patches) and learned expert aggregation but removes adaptive context selection. Its lower performance further supports the contribution of CIT beyond predefined multi-window modeling. - 3. Statistical Reliability and Stronger Baselines **Repeated experiments.** We conduct 10 independent runs and report mean and standard deviation: | Dataset | Model | Corr ↑ | CS-IC ↑ | |---|---|---:|---:| | HS300 | **RAVEN** | **0.0391±0.0008** | **0.0576±0.0006** | | HS300 | MLF | 0.0357±0.0021 | 0.0504±0.0018 | | S&P500 | **RAVEN** | **0.0364±0.0008** | **0.0473±0.0006** | | S&P500 | MLF | 0.0302±0.0022 | 0.0387±0.0019 | RAVEN achieves higher mean Corr and CS-IC with lower variation across random seeds on both markets. **Additional baselines.** On HS300, RAVEN (Corr 0.0391) also outperforms TimeSqueeze-base (0.0315), TFPS (0.0276), TRA (0.0233), and LightGBM using Alpha360 features (0.0254). |Model|Corr ↑|CS-IC ↑|MSE ↓|ICIR ↑| |---|---|---|---|---| |RAVEN|**.0391**|**.0576**|**.9904**|**.3938**| |LightGBM (raw)|.0171|.0249|1.0058|.2405| |LightGBM (Alpha360)|.0254|.0406|1.0014|.3201| |TRA|.0233|.0352|1.0021|.3034| |TFPS|.0276|.0436|.9972|.3204| |TimeSqueeze-base|.0315|.0464|.9951|.3546| - 4. Financial Relevance and Backtesting We additionally report risk-adjusted portfolio performance under a unified Qlib backtesting protocol on HS300 (2020–2024), using the same portfolio strategy and transaction costs. | Model | SR ↑ | IR ↑ | MDD ↓ | Turnover ↓ | |---|---:|---:|---:|---:| | **RAVEN** | **1.987** | **1.797** | 16.10% | 15.11% | | MLF | 1.723 | 1.663 | 17.56% | 16.43% | | PathFormer | 1.418 | 1.312 | 14.29% | 14.24% | | TimeSqueeze-base | 1.646 | 1.632 | 17.35% | 17.91% | | TFPS | 1.437 | 1.364 | 18.21% | 17.88% | RAVEN improves Sharpe/IR over second-best MLF by 15.3%/8.1% and has the second-lowest drawdown among evaluated models, supporting improved risk-adjusted performance. We will incorporate the gradient derivation, context-length distributions, fixed-window and ablation results, repeated-run comparisons, additional baselines, and backtesting details into the revised manuscript and appendix. Reviewer-specific issues, including label-boundary handling, patch-count sensitivity, and tuning protocols, are addressed in the individual responses. We hope these additional results and clarifications help address the remaining concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.