acceptodds
Under review as a conference paper at ICLR 2027

DIVINE: Simple Cross-Market Stock Pretraining via Diverse Indicator Reconstruction

12 views

Abstract

Financial time-series pretraining typically learns from masked observations, contrastive relations, or future outcomes—yet existing objectives struggle to simultaneously avoid future-supervision uncertainty and maintain return-prediction alignment. We propose DIVINE (DIVerse INdicator rEconstruction), a simple cross-market pretraining framework that reconstructs technical indicators from raw OHLCV history. Computed from observed price-volume history, technical indicators provide consistently defined supervision across markets while summarizing diverse market dynamics with established relevance to return prediction. Pretrained jointly on six-equity market datasets, DIVINE reconstructs 77 targets derived from 16 standard indicators and transfers only the learned encoder to downstream stock ranking. Across all six markets, DIVINE achieves the strongest average portfolio performance with a lightweight 0.05M-parameter encoder, outperforming pretraining baselines and matching or exceeding substantially larger financial foundation models, while remaining robust and data-efficient. Systematic analyses show that indicator diversity and market diversity provide complementary gains in transfer. Together, these results suggest that supervision design and cross-market diversity—rather than model scale—are the key drivers of strong, transferable financial representations.

Primary area:
learning on time series and dynamical systems
Keywords:
Financial time series, Self-supervised learning, Time-series pretraining, Cross-market transfer, Technical indicator reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9VvpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes DIVINE, a cross-market pretraining framework whose self-supervised objective is to reconstruct technical indicators from raw OHLCV history. Given a length-L window, a shared encoder g_theta (default: a 2-layer GRU, 0.05M params) produces a representation, and a linear head predicts D=77 targets derived from 16 standard indicators (SMA, EMA, MACD, RSI, ATR, MFI, etc.) at periods {3,5,10,14,20} (Eq. 2-3), with careful three-stage normalization. The encoder is pretrained jointly on six markets (CSI300, CSI500, NI225, SP500, FTSE100, TWSE300; 2008-2019 train, 2020 val, 2021-2024 test), the reconstruction head is discarded, and only the encoder is transferred (default: frozen) to a per-market MLP head trained on a differentiable Spearman ranking objective. Validation is via portfolio backtests (top-k, SR/CR, 25 seed-offset evaluations). Headline results: average SR/CR 1.396/1.790, beating the strongest pretraining baseline PatchTST (+39%/+49%) and Kronos-small (1.256/1.404) with ~490x fewer parameters. Ablations claim: indicator-type diversity matters much more than period-scale diversity; cross-market pretraining is the largest single gain; the encoder hidden state outperforms the reconstructed indicators themselves; and target-excluded (OOD) pretraining recovers 64% of the joint-pretraining gain; robustness to transfer protocol, context length, transaction cost, and data efficiency are also reported.

Critical strengths
  1. Thorough, controlled empirical study. Six markets, 25 matched seed-offset evaluations, three encoder backbones, both frozen/fine-tuned transfer, transaction-cost and data-efficiency sweeps, and IC/RankIC cross-checks make the empirical results reproducible and hard to dismiss as tuning artifacts. This is the paper's main strength, though it does not by itself establish the claimed mechanism.
  2. Informative analysis of what drives transfer. The indicator- vs period-diversity ablation and the target-excluded pretraining analysis (Table 4) jointly articulate a coherent "diversity in what and where you reconstruct" design principle that is more valuable than the raw benchmark numbers.
Critical weaknesses
  1. The mechanistic story is undercut by the paper's own results. The central claim is that indicator reconstruction supplies return-aligned supervision, yet (a) the hidden representation beats the reconstructed indicators (1.396 vs 1.333), and (b) MFI reconstruction fails entirely on CSI markets (R2<0) while performance there is the strongest. This suggests gains derive from cross-market pretraining generally rather than from the indicator reconstruction target specifically; the paper's interpretation over-reaches without a same-capacity control (e.g., cross-market masked OHLCV reconstruction under identical architecture/context).
  2. Evaluation validity tempers the headline superiority. The fixed, survivor-filtered universe and cost-excluded main SR/CR are acknowledged, and FTSE100 goes negative under costs. The foundation-model comparison is explicitly uncontrolled: Kronos is pretrained through June 2024, overlapping the 2021-2024 test window, and FinCast's temporal coverage is unknown. Superiority over models that may have seen test-period data is not a clean claim, weakening the "beats 490x larger models" headline.
Questions for authors
  1. Can you add a same-capacity, same-context cross-market control (e.g., masked OHLCV reconstruction or future-return-free alternatives) to isolate the specific contribution of indicator targets versus cross-market pretraining at large?
  2. Given MFI's failure and hidden-state > reconstructed-indicator results, how much of the gain is attributable to the 120-day pretraining context vs the indicator target? (Your own 20-bar result is reassuring but is not the same as matching baselines' objectives.)
  3. Can the Kronos/FinCast comparison be made strictly out-of-sample (post-June-2024), or the temporal overlap quantified, to validate the scale-independence claim?
  4. Are the conclusions stable under a point-in-time universe and after transaction costs, given the FTSE100 reversal?
Reviewer_w4jQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a pre-train-and-downstream framework for stock encoders across multiple equity datasets. The cross-market evaluation is notable. However, the overall academic contribution remains limited. First, the methodological novelty is low. Second, the empirical findings lack a theoretical foundation. Third, core terms are undefined and lack essential citations. Finally, minor editorial flaws persist, such as undefined acronyms.

Critical strengths

None.

Critical weaknesses
  1. Incremental Novelty: The framework essentially mimics a standard pre-training and fine-tuning pipeline. Expanding the evaluation to six equity datasets does not constitute a breakthrough in terms of core methodology or academic contribution.
  2. Absence of Theoretical Analysis: The authors highlight the empirical performance of the pretrained encoder as a major contribution; however, they fail to provide any theoretical backing or explanation for this phenomenon.
  3. Imprecise Terminology and Missing Literature: Critical concepts (e.g., "future-supervision," "past-only supervision," and "future return") lack formal discussion, mathematical grounding, and appropriate literature citations.
  4. Formatting Oversight: The paper uses important technical terms (e.g., OHLCV) as acronyms in the abstract and introduction without spelling them out on first use.
Questions for authors
  1. How does the proposed method balance two competing goals: mitigating future-supervision uncertainty and maintaining return-prediction alignment?
  2. What are the core innovations and main academic contributions of this work?
  3. Technical indicators are used solely during pre-training. Despite employing fewer parameters and less data, the framework achieves superior performance. What is the theoretical justification for these gains?
Reviewer_S3wXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DIVINE, a cross-market financial pretraining framework designed to learn transferable representations for cross-sectional stock return ranking. Instead of relying on masked reconstruction or future return supervision, DIVINE learns representations by reconstructing 77 targets that are derived from 16 standard technical indicators across multiple time periods, using raw OHLCV historical data. After pretrained jointly on historical OHLCV data from six global equity markets, this framework discards the linear reconstruction head and only transfers the learned encoder for the downstream cross-sectional stock ranking tasks. The extensive experiments adopt 2008–2019 for training, 2020 for validation, and 2021–2024 for testing. The results demonstrate that DIVINE outperforms evaluated financial pretraining baselines, and it matches or exceeds larger financial foundation models with hundreds of times fewer parameters.

Critical strengths
  1. DIVINE uses technical indicators computed from observed price-volume history as reconstruction targets, which avoids future-supervision uncertainty while maintaining alignment with stock ranking. This may provide a principled direction for self supervised financial representation learning.
  2. By using a lightweight encoder and historical signals, DIVINE delivers robust portfolio performance across six markets and achieves good average IC and RankIC, which outperforms massive foundation models. Data efficiency results demonstrate practical value for real world quantitative use cases.
Critical weaknesses
  1. DIVINE and the baselines differ in architecture, normalization, context length, and training budget. As such, the gains attributed to market diversity are confounded with data volume and training resources, making it difficult to isolate its incremental contribution and identify the reasons of the improvements. This makes the claim that market diversity drives transfer insufficiently validated.
  2. DIVINE processes each stock independently using a per-stock encoder and does not natively model cross-sectional inter-stock relationships in its backbone, making this architecture fails to capture systemic inter-stock correlations. This limitation causes DIVINE to underperform inter-stock modeling methods on smaller stock universes such as FTSE100.
Questions for authors
  1. Sections 5.3 and 6.2 compare single-market pretraining with joint pretraining across several markets. Joint pretraining also provides more stock histories and potentially more optimization steps under the fixed-epoch schedule. And are all the pretraining baselines trained jointly on the same 6 market corpus? Can you provide a comparison using the same GRU, preprocessing, context, and optimization budget for indicator reconstruction v.s. masked reconstruction and future-return supervision?
  2. DIVINE uses a per-stock encoder without modeling cross-stock interactions. In the experiment of FTSE100, models capturing stock relations can close the performance gap. Could you quantify how much performance degradation we should expect for markets with small stock universes?
  3. Given that the 77 targets are generated from 16 technical indicators through overlapping rolling-window transformations, which can make many targets collinear and redundant. Though the ablation study shows nearby periods bring marginal gain, it remains unclear whether performance improvements come from diverse market dynamics or just over-parameterized fitting of correlated price features? Provide additional justification for the technical indicator targets. And explain how sensitive is the downstream ranking performance to the exact choice and number of the 16 technical indicators included in the pretraining target set?
1 comment

Thanks for the response. The overall contribution still feels somewhat weak, and I will keep my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.