acceptodds
Withdrawn from ICLR 2027

Multi-Horizon Uncertainty-Aware Latent Scenario Learning for Real-Time Sequential Financial Analytics and Decision-Making with Asynchronous Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Machine learning in the finance domain always involves multi-task sequential quantitative analytics and decision-making, in which research analysts extract predictive signals from diverse firm characteristics, form prospects of future financial disclosure and valuations of stock prices, then construct investment portfolios based on their return expectations. In reality, they have to deal with asynchronous financial data and thereby obsolete information, e.g., accounting-based fundamentals are released quarterly or annually with lags while market-based characteristics such as trading activities can be collected in a more timely manner. This imposes huge challenges in the sequential investment process, where input uncertainty can propagate through multi-task learning, leading to (i) distorted beliefs about firm fundamentals in upstream data analytics, (ii) misinformed return expectations in midstream asset pricing, and (iii) suboptimal downstream portfolio decisions. To tackle these challenges, we propose a novel unified multi-horizon uncertainty-aware latent scenario learning network that simultaneously (i) infers real-time multi-horizon probabilistic firm fundamentals jointly from the stale prints and the latest market data, (ii) estimates probability-weighted systematic risk exposures to common risk factors that produce realistic multi-horizon cross-sectional expected returns, and (iii) constructs various multi-horizon robust investment portfolios, which take the generative latent scenarios into account in optimization. Extensive experiments, ablation studies, and sensitivity analyses on the most comprehensive open-source stock market dataset show that our proposed model consistently outperforms seminal baselines in relevant literature across all evaluation metrics for sequential financial tasks. The results also emphasize the importance of real-time uncertainty-aware design of the latent scenario learning mechanism.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Asset Pricing, Investment Portfolio, Data Completion, Real-Time Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_4sHoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses state uncertainty arising from asynchronous financial data and its propagation into investment portfolio decision-making, proposing a multi-horizon, uncertainty-aware latent scenario learning framework. Overall, the methodological design is relatively comprehensive, and the modeling of multiple horizons and the mechanism for propagating uncertainty offer potential value. However, the independent contributions of the key modules have not been sufficiently validated. The relationship between latent-space anchoring and consistency with raw observations remains unclear, and the actual scope of uncertainty modeling, as well as the probabilistic predictive quality of the generated scenarios, lacks direct empirical evidence. Overall, the paper presents a promising research framework, but its methodological justification and empirical validation remain insufficient. Further targeted validation of the core design choices is therefore recommended.

Critical strengths
  1. The research problem is practically relevant. The paper explicitly distinguishes among actual disclosure status, information staleness, and missing information, and attempts to reconstruct firms' real-time states under strict temporal constraints. The research motivation is clearly articulated.
  2. The ablation studies cover multiple components across the pipeline. The paper compares several configurations, including raw lagged features, deterministic imputation, removal of hierarchical encoding, removal of the staleness channel, and replacement of the return prediction mechanism. These experiments help elucidate the contributions of different design choices. Reporting means and standard deviations across multiple runs also demonstrates attention to the reliability of the reported performance results.
Critical weaknesses
  1. The claim of strict freedom from look-ahead bias is insufficiently substantiated. Equation (15) uses the historical risk premium λₜ₋₁, whereas Appendix Equation (16) and the algorithm use λₜ. Algorithm 2 states that its input information set is Iₜ₋₁, yet computes factors using information through time t. Furthermore, the paper does not clearly explain whether forward-return labels with horizons of up to twelve months are properly isolated across the training, validation, and test boundaries. These inconsistencies make it difficult for readers to verify the paper's repeated claims of strict causality and the absence of look-ahead bias.

  2. The independent contributions of the core modules have not been adequately validated. Table 3 provides some evidence for the advantages of the unified horizon-decoding design, but lacks targeted experiments that remove latent-space anchoring, vary the dimensionality of horizon compression, or unfreeze the horizon decoder. Table 2 reports predictive performance across different horizons; however, stronger long-horizon RankIC values alone do not establish that shared horizon representations improve cross-horizon consistency. In addition, the paper lacks a direct comparison with the most closely related conditional diffusion-based imputation methods.

  3. The relationship between latent-space anchoring and consistency with observations in the original feature space remains unclear. Equations (4)–(6) encode the original features into a latent space, while Equation (10) fixes a subset of latent variables. However, the paper does not explain how the original-feature disclosure mask is mapped to latent coordinates. Because the decoding process may mix information across latent dimensions, fixing selected latent variables does not, in itself, guarantee that the corresponding disclosed features remain unchanged after decoding.

  4. The scope of uncertainty modeling and the supporting evaluation evidence are insufficient. FactorNet uses the posterior mean, the risk premium is estimated using a historical rolling average, and the randomness of the return scenarios arises primarily from upstream firm-state uncertainty. Consequently, the generated scenarios more closely resemble a distribution of conditional expected returns induced by uncertainty in firm states than a complete predictive distribution of future returns. The reported RMSE, MAE, and RankIC metrics are also insufficient to directly assess the reliability of probabilistic predictions.

Questions for authors

See weakness.

Reviewer_T8hROpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Problem. Cross-sectional asset pricing and portfolio construction when firm characteristics are disclosed asynchronously (monthly market variables alongside quarterly and annual accounting variables carried forward with lags).

Method. Three sequential, separately trained stages.

  • State inference. A hierarchical VAE over characteristic groups and disclosure frequencies, with a latent diffusion model whose reverse steps anchor disclosed dimensions and sample undisclosed ones. One model performs backcasting, nowcasting, and forecasting. It outputs S=100S=100 firm-state scenarios per month.
  • Factor pricing. FactorNet, a VAE on trailing 1- to 12-month returns only, compresses the 12 horizons to 3 and extracts K=16K=16 latent factors. Its horizon decoder is frozen. BetaNet maps each state scenario to factor loadings. Forecasts are R^=Dec(β⊙λt)\hat R = \mathrm{Dec}(\beta \odot \lambda_t) with λt\lambda_t a 48-month rolling mean of past factors.
  • Allocation. Scenario-averaged one-month forecasts, equal-weight top decile, 10 bps cost. A robust allocation template with ambiguity sets (Eq. 20) is described.

Evaluation. OSAP data, 405 U.S. stocks, 96 characteristics, monthly. Single chronological split (train 1980–2007, validation 2008–2009, test 2010–2022), all parameters frozen at test time. Twelve baselines (SFM, ALSTM, TRA, AdaRNN, GAT, AGDL, CAE, FactorVAE, CQVAE, GAN, AP-Tree, FactorGCL). Metrics are RankIC, RankICIR, AR, SR, CR, MDD over five seeds.

Claims. Highest RankIC, RankICIR, AR, CR, and lowest MDD at one month. Ranking strength grows with horizon. Ablations show that raw lagged features, deterministic imputation, recursive rollout, and contemporaneous factors all degrade results.

Critical strengths

1. The economic problem and the pipeline are well motivated. Treating stale and undisclosed fundamentals as a state-uncertainty problem, propagating that uncertainty through conditional betas rather than collapsing it to a point estimate, and keeping factor premia strictly causal (trailing returns only, rolling λt\lambda_t) is a coherent design that matches how practitioners think about point-in-time data.

2. Point-in-time discipline is explicit. Disclosure indicators and staleness are modelled as separate inputs, trailing and forward return panels are separated, and the appendix documents the information set at each origin. The ablations target the right questions (stochastic vs. deterministic imputation, causal vs. contemporaneous factors).

Critical weaknesses

1. Regime robustness of the frozen factor model is untested, and no walk-forward evaluation exists.

All networks are trained on 1980–2007 and frozen for the entire 2010–2022 test period (Appendix B). Only λt\lambda_t updates, as a 48-month rolling mean. A FactorNet fitted on pre-2008 returns is applied without re-estimation through the 2020 crash and the 2022 rate cycle, yet no evidence is given that the latent factors or the horizon decoder remain valid. There is no sub-period breakdown, no plot of ftf_t or λt\lambda_t over time, no re-training schedule, and no test of whether the factor latents alone carry predictive content. Because the whole method rests on the factor representation, this is the central open question.

2. The reported improvements are not statistically established, and the ablation results raise rather than settle doubts about the pricing stack.

MetricOursBest baseline
RankIC (%)1.93 ± 0.47AGDL 1.68 ± 0.31
AR (%)14.33 ± 1.31GAT 13.84 ± 2.46
SR0.825 ± 0.035ALSTM 0.836 ± 0.085

Every headline gap is within one standard deviation across five seeds, Sharpe is below ALSTM, and no significance test is reported. Table 2's multi-horizon returns overlap, so the rise in RankICIR from 15 to 40 is partly a mechanical autocorrelation effect that requires HAC-type correction. Table 3 reports single numbers with no dispersion at all, and its Full Model row (RankIC 1.799, SR 0.842, MDD 26.74) does not match the Ours row of Table 1 (1.927, 0.825, 29.46) with no explanation. Most importantly, the Raw Lag Feat. variant (SR 0.593, RankIC 0.712) is far below ALSTM and SFM, which consume the same raw lagged features. This implies that the FactorNet plus BetaNet pricing structure underperforms a plain sequence model unless the diffusion imputation is present, which is the opposite of what a sound factor model should show.

3. The closest comparators and the claimed robust-allocation stage are missing, and nothing can be verified.

STABLE (Soun et al., ICLR 2026), which also feeds generated scenarios into portfolio allocation, is not compared. Recent portfolio-weight methods such as DeepTrader (Wang et al., AAAI 2021), Diffolio (Jeon et al., WSDM 2026), and the direct forecasters MASTER and StockMixer that the paper itself cites are not run. The robust allocation of Eq. 20 (max-min, CVaR, box, ellipsoidal, KL) is listed as a contribution but has no results anywhere. Table 1's note says portfolios use a mean–variance rule while Appendix C and Table 5 say equal-weight top decile. The selection rule for the 405 stocks is not stated, so survivorship bias over a 42-year panel cannot be excluded. No code is provided.

Questions for authors

Q1. Regime robustness of the factor latents (critical). Plot ftf_t and λt\lambda_t over 2010–2022 and report sub-period results (e.g. 2010–2014, 2015–2019, 2020–2022). Report a simple model (ALSTM, logistic regression, or ridge) that predicts returns from the factor latents alone, so that the information content of FactorNet can be assessed separately from the diffusion stage. Report results with FactorNet re-estimated on an expanding or rolling window instead of frozen at 2007.

Q2. Statistical significance with overlap correction (critical). Report paired tests of Ours vs. each baseline for RankIC, AR, and SR using HAC (Newey–West) standard errors on the monthly series. For Table 2, correct for the hh-month overlap. Add dispersion and paired tests to every row of Table 3.

Q3. Explain the ablation pattern. Why does the Raw Lag Feat. variant fall well below ALSTM and SFM, which use the same raw features? Why does Sharpe not improve over simple baselines when RankIC does? Reconcile the Full Model row of Table 3 with the Ours row of Table 1.

Q4. Comparators under the same protocol. Add STABLE (Soun et al., ICLR 2026), DeepTrader, Diffolio (Jeon et al., WSDM 2026), MASTER, and StockMixer under the identical data, split, universe, and cost convention.

Q5. Robust allocation results. Report results for the ambiguity-set allocations of Eq. 20, since this is a stated contribution. Also state which portfolio rule produced Table 1, mean–variance or equal-weight top decile.

Q6. Universe and data. How were the 405 stocks selected from OSAP, and are they required to have full 1980–2022 histories?

Q7. Code. Release the code so that hyperparameter and checkpoint selection on the 2008–2009 validation period, the frozen-at-test protocol, and the point-in-time construction can be verified.

Reviewer_4DrXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for financial forecasting with asynchronously disclosed firm characteristics. It combines disclosure-aware latent diffusion for firm-state inference, scenario-conditioned multi-horizon factor pricing, and portfolio construction. Evaluation on the OSAP dataset reports improvements in return ranking and several portfolio metrics, supported by component ablations.

Critical strengths

The framework explicitly accounts for differences in feature-update frequencies, disclosure status, and information age. Propagating inferred firm-state scenarios into return prediction provides a meaningful approach to handling stale financial inputs rather than treating all available characteristics as equally current.

Critical weaknesses

-- Evaluation is restricted to one dataset and the U.S. equity market. Moreover, the long-only strategy’s approximately 14.3% annualized return over 2010–2022 is difficult to interpret without same-universe equal-weight and buy-and-hold benchmarks. Its Sharpe ratio is not the highest among the compared methods. Together, these limitations leave unclear how much economic value comes from stock selection rather than market or conventional factor exposure.

-- The paper does not establish relationships between its latent factors and traditional risk factors, explain the economic meaning of estimated exposures, or report standard pricing-error diagnostics. A factor-structured architecture alone does not substantiate the claimed economic interpretability.

Questions for authors
  1. How does the strategy compare with same-universe equal-weight and buy-and-hold portfolios under identical trading-cost assumptions? Does it deliver statistically significant factor-adjusted excess returns?

  2. What evidence supports the economic interpretation of the learned factors and exposures? Can the authors provide relationships with established factors and corresponding pricing-error diagnostics?

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

This paper presents a three-stage framework for multi-horizon stock return prediction and portfolio construction using asynchronous financial data. Contributions are as follows;

  • A hierarchical encoder and a conditional latent diffusion model fuse mixed-frequency firm features, disclosure indicators, and data age to generate complete firm-state scenarios.
  • A factor pricing module uses a variational autoencoder to extract risk premia from past returns and maps each firm-state scenario to specific factor exposures.
  • A frozen low-rank horizon decoder converts the product of these exposures and risk premia into consistent 1- to 12-month return distributions for portfolio optimization.

Experiments and ablation studies on 405 stocks from the Open Source Asset Pricing dataset show improvements in return ranking and most portfolio metrics compared to twelve baseline models.

Critical strengths
  • The paper addresses two important problems in quantitative investment: combining features with different update frequencies and availability, and making consistent multi-horizon return predictions.
  • The paper is clearly written and well organized. The architecture diagrams and step-by-step algorithms make the three-stage framework easy to understand.
  • The experimental setup is thorough and compares the model against twelve baselines across multiple return horizons. The ablation studies also give useful insights by showing that lower feature reconstruction error alone does not guarantee better portfolio performance, even though the absolute predictive power remains modest.
Critical weaknesses
  • The model predicts total stock returns only through systematic factor exposures (β\beta) and has no residual alpha (α\alpha) term. Forcing all 96 firm features through a 16-factor bottleneck (β⊤λ\beta^\top \lambda) treats the residual return (ϵ\epsilon) as random noise and misses stock-specific alpha.
  • The return predictions are weak across all horizons, and the portfolio optimizer in Stage 3 uses a single horizon for trading. The Rank Information Coefficient is only 1.93% at the one-month horizon. While it reaches 6.24% at the 12-month horizon, a 12-month holding period gives only one independent rebalance per year across 405 stocks (low breath).
  • As a minor presentation issue, the math symbols and time indices between Stage 1 and Stage 2 are slightly unclear. Equation 3 and Equation 11 define the decoded features as X~t(s)\tilde{X}_t^{(s)} and the latent state as z~t(s)\tilde{z}_t^{(s)} at time tt, but Section 3.3 passes an undefined Zt−1(s)Z_{t-1}^{(s)} at time t−1t-1 into BetaNet.
Questions for authors

none.

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

Related papers

Discussion (0)

Sign in to comment.