PaperlayerView exact revision on arXiv

Source-linked AI summary

STOCK-JEPA: Prior-Anchored Latent Revision Representation Learning in Equity Markets

Yizhi Luo, Jiahe Yi, Jianhui Zhang, Shuo Sun

arXiv:2610.07006v1cs.LGq-fin.ST

TL;DR

Noisy financial data challenge representation learning, as deep models can overfit and lack explicit economic structure while linear priors leave nonlinear signals uncaptured. STOCK-JEPA anchors future-state representations with financial statistics and learns predictable revisions from historical context. It outperforms 13 baselines across large-scale China and U.S. equity universes, with the optimal revision also yielding a precise population-risk reduction.

  • Problem

    Financial representation learning must characterize market structure and return-risk dynamics from noisy data, while deep models may overfit or lack clear economic interpretation and linear priors can miss nonlinear signals.

    Source: §1 · pp. 1–2

  • Method

    STOCK-JEPA maps fixed financial-prior statistics to a latent anchor and separately learns the predictable revision from historical context.

    Source: §1 · p. 2

  • Results

    STOCK-JEPA ranks best on 4 of 5 evaluation metrics in each of the China and U.S. markets, improving RankIC by 14.0% and 15.9% over the second-best method.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Frozen representations achieve the highest mean RankIC and net Sharpe among evaluated methods, while ablations and simulation support the revisions’ incremental and complementary signal value.

    Source: §1 · p. 2, §6 · p. 10

    The population analysis fixes the target encoder and compares minimum squared prediction risk under prior and historical-context information.

    Source: §3 · p. 3

Abstract

from arXiv · show

Learning effective representations helps characterize the structure and dynamics of equity markets from financial data with a low signal-to-noise ratio. Black-box deep models can capture complex patterns but may overfit sample noise and lack explicit economic structure. Meanwhile, classic linear financial models provide interpretable references, but their oversimplified assumptions leave non-linear signals uncaptured. To combine the strengths of these two directions, we propose Stock-JEPA, a joint-embedding predictive framework that learns predictable incremental revisions relative to a point-in-time financial prior. First, we leverage a low-complexity financial model to produce fixed statistics summarizing multi-horizon return and risk. A prior projector then maps these statistics into the target encoder's latent space as an anchor. Second, we design a context-conditioned revision predictor to estimate the future representation's predictable displacement from the anchor. Separate losses update the two branches: the anchor learns from prior statistics, while the revision captures additional predictable information from historical context. Third, we freeze all representation modules and train a downstream readout, evaluating its forecasts through cross-sectional ranking and portfolio performance. Theoretically, we prove that optimal revision reduces the prior anchor's expected squared error for the same future representation by exactly $\mathbb{E}[\|\boldsymbol{\Delta}\|_2^2]$. This non-negative gain is the expected squared magnitude of the additional signal predictable from historical context. Experimentally, Stock-JEPA outperforms 13 strong baselines across large-scale China and U.S. equity universes on 5 key evaluation metrics. Ablation studies and representation analysis further demonstrate the value of the learned revisions for representation learning in equity markets.

1 INTRODUCTION

Financial representation learning must reconcile noisy, low-signal data with the flexibility of deep models and the structured interpretability of financial priors. STOCK-JEPA uses a financial prior as an anchor and learns context-predictable revisions, with experiments and theory supporting this approach.

1 INTRODUCTION

  • Deep models can capture nonlinear patterns but may lack clear economic interpretation and overfit realized future outcomes, while linear financial priors can leave nonlinear interactions unexplained.These limitations motivate combining structured financial models with deep learning.Source: §1 · pp. 1–2
  • The framework maps fixed multi-horizon return and risk statistics into a latent-space anchor, then predicts its context-conditioned displacement toward the future representation.Separate objectives train the prior anchor and the additional predictable component; representation modules are frozen for downstream readout evaluation.Source: §1 · p. 2
  • Across China and U.S. equity markets, STOCK-JEPA ranks best on 4 of 5 metrics in each market and improves RankIC by 14.0% and 15.9% over the second-best method.Frozen representations also achieve the highest mean RankIC and net Sharpe among evaluated methods.Source: §1 · p. 2
  • The paper frames financial representation learning as separating what history can help predict beyond a structured prior from what remains unpredictable.Its evaluation includes realistic settings, including chronological splits, transaction costs, and long-short constraints.Source: §1 · p. 2
  • Optimal revisions reduce squared prediction error for the same future representation by exactly the non-negative expected squared magnitude of the predictable revision.The result isolates the additional predictability supplied by historical context beyond the prior.Source: §1 · p. 2

2 RELATED WORK

Prior work combines financial structure, residual learning, and predictive representations in different ways. STOCK-JEPA distinguishes itself by treating a financial prior as an anchor for predictable revisions in future-state representation space.

2.1 FINANCIAL MODELS AND RESIDUAL LEARNING

  • Residual-based strategies derive signals from factor-adjusted returns, whereas STOCK-JEPA uses a financial anchor to define revisions in representation space.Prior examples include mean-reverting idiosyncratic returns and signals from histories of factor-residual portfolios.Source: §2.1 · pp. 2–3

2.2 DEEP FINANCIAL FACTOR MODELS

  • Deep financial factor methods organize learned representations as factors or exposures; STOCK-JEPA retains a financial anchor without requiring revisions to have a factor interpretation.Examples include conditional autoencoders, FactorVAE, FactorGCL, and PRISM-VQ.Source: §2.2 · p. 3

2.3 PREDICTIVE REPRESENTATION LEARNING

  • Predictive representation methods predict latent targets or partition latent states, but do not themselves distinguish prior-predictable information from what richer context adds.STOCK-JEPA addresses that distinction through a financial anchor and separate prior-prediction and context-revision objectives.Source: §2.3 · p. 3

3 PROBLEM FORMULATION

The problem formulation separates prediction from fixed prior statistics and prediction using prior information together with historical context. The optimal context-conditioned prediction adds the predictable revision to the prior anchor, reducing population squared risk by the revision’s expected squared magnitude.

3 PROBLEM FORMULATION

  • For the same target representation, conditioning on more historical information can maintain or reduce minimum achievable mean squared error.The population analysis fixes the target encoder and uses prior statistics and historical context at the forecast origin.Source: §3 · p. 3
  • The prior anchor captures what the prior predicts, while the revision captures what historical context adds; together they equal the target representation’s conditional mean given context.The decomposition is a conditional-mean restriction, not an independence assumption.Source: §3 · p. 3
  • The learned displacement targets its predictable component: under squared loss, its optimal predictor from historical context is the conditional mean revision.This target need not correspond to a particular omitted factor or require the prior to span the true factor space.Source: §3 · p. 3
  • The optimal revision reduces squared prediction risk by exactly its non-negative expected squared magnitude; if history adds no conditional-mean predictability, the revision is zero and the prior prediction is recovered.The non-increase guarantee also holds for parameterized revision classes containing the zero function when the target encoder and anchor are fixed.Source: §3 · p. 4

4 METHODOLOGY

STOCK-JEPA builds a point-in-time financial prior, learns a latent anchor and context-predicted revision, then evaluates frozen representations with a trainable return readout.

4.2 PRIOR-ANCHORED REPRESENTATION LEARNING

  • The target encoder and prior projector form future targets and financial anchors in a shared latent space, while the context encoder and revision predictor estimate predictable displacements from historical information.The projector maps prior statistics into the target latent space; the revision branch uses richer historical context than the anchor branch.Source: Figure 2 · p. 5, §4.2 · pp. 5–6

4.1 POINT-IN-TIME STRUCTURED FINANCIAL PRIOR

  • The point-in-time prior uses Ridge forecasts of return, volatility, drawdown, and market-relative return at 5- and 21-trading-day horizons, fixed during representation learning.Historical fitting errors and residual scales use only outcomes available before each forecasting block begins.Source: §4.1 · p. 5

4.2 PRIOR-ANCHORED REPRESENTATION LEARNING

  • Separate anchor and revision losses assign distinct prediction roles, and stop-gradient prevents revision gradients from directly updating the prior projector.EMA target encoder and anchor copies provide stable training targets and receive no gradients.Source: §4.2 · p. 6

4.3 RETURN READOUT WITH FROZEN REPRESENTATIONS

  • After selecting a checkpoint, the method freezes representation modules and trains an MLP on 21-day log-return labels using the historical summary, prior, anchor, and predicted revision.At inference, the readout uses historical observations and precomputed prior statistics available at the forecast origin.Source: §4.3 · pp. 6–7

5 EXPERIMENT

Across China and U.S. equity universes, STOCK-JEPA leads most reported ranking and portfolio metrics. Ablations and representation analyses examine revisions’ predictive value, structure, and signal recovery.

5.1 EXPERIMENTAL SETUP

  • Experiments cover 5,669 China and 10,152 U.S. historically tradable securities, with chronological training, validation, and test periods and fixed priors within each prediction block.The common evaluation uses a 21-day horizon and charges 5 bps per unit of executed notional at entry and exit.Source: §5.1 · p. 7, §5.2 · p. 7

5.2 CROSS-SECTIONAL RANKING AND PORTFOLIO EVALUATION

  • Table 1 reports first place in four of five metrics in each market; RankIC is 11.22% on CN ALL and 12.96% on US ALL, 14.0% and 15.9% above the second-best method.Mean net annualized returns and net Sharpe ratios are also highest in both markets: 45.66% and 1.50 in China, and 68.20% and 1.24 in the U.S.Source: §5.2 · p. 7, Table 1 · p. 8

5.3 UNDERSTANDING THE PRIOR-ANCHORED REPRESENTATION

  • Matched readouts show that adding revisions raises mean RankIC by 2.02/3.67 percentage points on CN ALL/US ALL; Separate-EMA has the highest mean portfolio measures in both markets.Revisions outperform equally sized context features on all four mean metrics in both markets, and Separate-EMA outperforms Ordinary-JEPA on all four mean metrics in both markets.Source: §5.3 · p. 8, Table 2 · p. 8
  • On US ALL, revisions have mean off-diagonal absolute correlation of approximately 0.30 and average daily effective rank of 21.1, while market-average revision direction changes across dates.Large daily adjustments occur across stocks around monetary-policy repricing and trade tensions; the event statistic uses the median stock-level change.Source: §5.3 · pp. 8–9, Figure 3 · p. 9
  • In simulation, Separate-EMA’s anchor achieves 66.14% mean test R2 on the slow signal, while its revision achieves 49.25% on the fast signal.Their concatenation retains both signals; all configurations score below the constant reference on average for independent future noise.Source: §5.3 · p. 9, Figure 4 · p. 9

6 CONCLUSION

STOCK-JEPA learns prior-relative revisions with separate anchor and revision objectives. Frozen representations lead evaluated methods in mean RankIC and net Sharpe in both markets, while diagnostics and simulation support revisions’ incremental value and complementary signal recovery.

6 CONCLUSION

  • STOCK-JEPA learns prior-relative revisions through separate anchor and revision objectives, and frozen representations lead evaluated methods in mean RankIC and net Sharpe in both markets.Ablations support incremental predictive value; diagnostics show stock-level and temporal variation, and simulation demonstrates complementary signal recovery from noisy targets.Source: §6 · p. 10

A.1 PROOF OF PROPOSITION 1

The proof characterizes the best prior-only and context-informed predictions of a fixed future representation, showing when revisions reduce squared prediction risk. The structured prior is built from point-in-time financial statistics and fixed before representation learning.

A.1 PROOF OF PROPOSITION 1

  • The prior-only optimum is the conditional mean of the target representation given the prior information, and its squared-loss minimizer is unique up to almost-sure equality.With the target and anchor fixed, a revision class containing the zero function retains the anchor-only predictor; the exact risk reduction requires attaining the conditional projections.Source: §A.1 · pp. 12–13
  • When history reveals predictable variation beyond the prior, the optimal revision reduces squared prediction risk by exactly E[∥∆∥2 2]; otherwise, the risk is unchanged.The reduction is non-negative and equals the expected squared magnitude of the predictable revision.Source: §A.1 · p. 12

B STRUCTURED FINANCIAL PRIOR

  • At each forecast origin, the financial prior uses 40 descriptors: 28 asset-level and 12 market-level statistics available through that time.Source: B STRUCTURED FINANCIAL PRIOR · p. 13
  • For each prediction block, a multi-output ridge prior is fit on historical samples whose complete 21-day outcomes end before the block begins.Block-specific histories also determine input scaling and raw residual scales.Source: B STRUCTURED FINANCIAL PRIOR · pp. 13–14
  • The structured prior combines eight normalized return-and-risk predictions with eight log residual-scale features, all precomputed and held fixed during representation learning.The prior projector maps these statistics into the target latent space.Source: B STRUCTURED FINANCIAL PRIOR · p. 14

C REPRESENTATION ARCHITECTURE AND OPTIMIZATION

The representation and readout pipeline uses masked historical context to predict revisions relative to a prior anchor, then freezes learned representations for return forecasting. The appendix also specifies the datasets, prior-fitting schedule, baselines, and evaluation procedures.

C REPRESENTATION ARCHITECTURE AND OPTIMIZATION

  • A 126-day context and 21-day future window use the same Transformer encoder architecture, producing 128-dimensional representations; training masks context blocks and feature groups but leaves the target window unmasked.The prior projector maps 16-dimensional prior statistics to the anchor, while the revision predictor uses the context representation and detached EMA anchor.Source: C REPRESENTATION ARCHITECTURE AND… · p. 14

D FROZEN-REPRESENTATION RETURN READOUT

  • After checkpoint selection, the representation modules are frozen; the return readout combines a historical summary, prior, EMA anchor, and revision to predict standardized 21-day cumulative log returns.The readout checkpoint is selected using validation SmoothL1 loss.Source: D FROZEN-REPRESENTATION RETURN READOUT · pp. 15–16

E BASELINE IMPLEMENTATION

  • The 13 baselines share the prepared stock panels, targets, and validation and test periods, with model configurations fixed before test evaluation and checkpoints selected by validation loss.The compared approaches include supervised neural models, factor models, and JEPA representation models.Source: E BASELINE IMPLEMENTATION · p. 16

F.1 DATASET CONSTRUCTION AND STATISTICS.

  • CN ALL and US ALL cover 5,669 and 10,152 historically tradable securities, respectively, with testing from 2024 through June 2026 and split-boundary target windows excluded.Stage 1 priors make representation and readout training possible from 2016, after the 2014–2015 initial fitting history.Source: §F.1 · pp. 16–17

F.2 EXPANDING-WINDOW PRIOR ESTIMATION.

  • Stage 1 generates annual out-of-sample priors for 2016–2021 using only fitting histories whose complete outcome windows end before each prediction period; validation and test priors are fitted on earlier eligible history.The fitted priors remain fixed within validation and test periods.Source: §F.2 · p. 17

G CROSS-SECTIONAL AND PORTFOLIO EVALUATION

  • Cross-sectional IC and RankIC are computed over each date’s asset universe, while portfolio evaluation uses Top-30 selections, overlapping 21-session batches, and transaction costs.The backtest assumes scheduled quoted-price execution and does not model market impact, execution restrictions, stock-borrow availability, borrowing fees, or funding interest.Source: G CROSS-SECTIONAL AND PORTFOLIO… · pp. 17–18

H COMPONENT ABLATION DETAILS

The component ablations compare standalone prior, context, and prior-anchored revision configurations, followed by the full model with a historical summary and MLP readout.

H COMPONENT ABLATION DETAILS

  • S1+S2 supplies the prior and predicted revision to a frozen linear readout, while the full model adds the historical summary and uses an MLP readout.Source: H COMPONENT ABLATION DETAILS · p. 19
  • S2-only is trained without a prior branch or prior loss, whereas S1+S2 uses the prior-anchored checkpoint to predict future latent revisions relative to a financial anchor.Source: H COMPONENT ABLATION DETAILS · p. 19
  • S1+S2 outperforms every standalone configuration across all four mean metrics in both markets; the full model achieves the highest means on all four metrics.Adding the historical summary and MLP raises RankIC by 1.28 and 1.07 percentage points and net Sharpe by 0.40 and 0.24 on CN ALL and US ALL, respectively.Source: H COMPONENT ABLATION DETAILS · p. 19

I MATCHED MLP READOUT ABLATION

The matched-readout ablation tests whether adding the anchor and predicted revision improves frozen-representation forecasts. Revisions improve all four metrics in both markets, and predicted revisions outperform direct context inputs.

I MATCHED MLP READOUT ABLATION

  • The ablation compares frozen inputs R0 [s; P], R1 [s; P; za], R2 [s; P; za; b∆], and R3 [s; P; za; h] using a matched MLP readout.The anchor, revision, and context representation are each 128-dimensional; only the MLP is trained.Source: I MATCHED MLP READOUT ABLATION · pp. 19–20
  • R2 has the highest means across all four metrics in both markets; versus R1, it raises RankIC by 2.02 and 3.67 percentage points and net Sharpe by 0.41 and 0.53 on CN ALL and US ALL.The comparison uses the matched MLP readout ablation in Table 5.Source: Table 5 · p. 20, I MATCHED MLP READOUT ABLATION · p. 20
  • R2 exceeds R3 on all four metrics in both markets, including RankIC gains of 1.13 and 2.27 percentage points and net Sharpe gains of 0.08 and 0.30 on CN ALL and US ALL.R2 uses a predicted revision; R3 supplies the context representation directly, with identical input dimensionality and MLP capacity.Source: I MATCHED MLP READOUT ABLATION · p. 20, Table 5 · p. 20

J CONTROLLED ABLATION OF REPRESENTATION-LEARNING MECHANISMS

The controlled ablation compares representation-learning objectives and anchor choices under matched continuation and readout conditions. Separate-EMA has strong portfolio metrics and outperforms Ordinary-JEPA across both markets.

J CONTROLLED ABLATION OF REPRESENTATION-LEARNING MECHANISMS

  • Separate-EMA separates anchor supervision from revision prediction using an EMA anchor, while Separate-online uses the online anchor and Joint-online removes separate supervision and gradient blocking.Mechanisms continue from the same checkpoint under matched initialization, data splits, masking, and budget; frozen representations feed the same readout architecture.Source: J CONTROLLED ABLATION OF… · p. 21
  • Separate-EMA leads on three of four metrics in each market, including both portfolio metrics; versus Joint-online, net AR rises by 11.18 and 6.53 percentage points and net Sharpe by 0.26 and 0.18 on CN ALL and US ALL.Joint-online still leads CN ALL RankIC and US ALL IC; Table 6 reports the controlled mechanism ablation.Source: J CONTROLLED ABLATION OF… · p. 21
  • Holding separate supervision and gradient routing fixed, the EMA anchor makes small changes on CN ALL and improves all four US ALL means, including net AR by 5.24 percentage points and net Sharpe by 0.14.The comparison is between Separate-EMA and Separate-online.Source: J CONTROLLED ABLATION OF… · p. 21
  • Separate-EMA outperforms Ordinary-JEPA on all four mean metrics in both markets, with RankIC gains of 1.94 and 3.40 percentage points on CN ALL and US ALL.All readouts use the same state and prior inputs and equally sized predicted representations.Source: J CONTROLLED ABLATION OF… · p. 21

K SIGNAL RECOVERY IN A CONTROLLED DYNAMIC-FACTOR ENVIRONMENT

In a controlled dynamic-factor environment, the anchor and revision recover complementary predictable signal components. Diagnostics describe revision structure and market responses, while state-based evaluation finds positive RankIC across evaluated groups.

K SIGNAL RECOVERY IN A CONTROLLED DYNAMIC-FACTOR ENVIRONMENT

  • Recovery measures linear accessibility, not additive branch scores; results aggregate five worlds, and fast-component recovery varies substantially across worlds.The synthetic environment’s slow–fast components differ from the theoretical prior-conditional anchor–revision decomposition.Source: K SIGNAL RECOVERY IN A CONTROLLED… · p. 23
  • Separate-EMA’s anchor recovers 66.14% of the slow component, while its revision recovers 49.25% of the fast component; concatenation recovers 62.22% slow and 55.24% fast.The concatenated representation retains substantial recovery of both components but does not improve both scores over their respective individual branches.Source: Table 7 · p. 23, K SIGNAL RECOVERY IN A CONTROLLED… · p. 23
  • Ordinary-JEPA and Direct supervised recover the fast component but have no positive mean slow-component recovery under linear probes; Separate-EMA’s fast-recovery ordering is uncertain across worlds.All configurations also show negative mean recovery of independent future return noise, which is unavailable at the forecast origin.Source: K SIGNAL RECOVERY IN A CONTROLLED… · pp. 23–24

L REVISION STRUCTURE DIAGNOSTICS

  • US ALL revisions have mean off-diagonal absolute correlation 0.304 and average daily effective rank 21.1; the temporal diagnostic compares directional distances between market-average revisions.The cross-sectional summary spans January 2024–November 2025; the temporal display spans October 2024–November 2025.Source: L REVISION STRUCTURE DIAGNOSTICS · p. 24

M REVISION RESPONSE DIAGNOSTIC

  • Revision-displacement peaks follow selected policy, tariff, monetary-policy, and trade-tension events, reflecting broad stock-level adjustments rather than a few extreme observations.The full-test 99th-percentile reference is retrospective; event correspondence does not establish causal effects or predictive anticipation.Source: M REVISION RESPONSE DIAGNOSTIC · pp. 24–25, Figure 6 · p. 25

N MARKET-STATE EVALUATION DETAILS

  • STOCK-JEPA has positive mean RankIC in every evaluated market-state group, ranking first in all six US ALL groups and five of six CN ALL groups; some baselines do better on extreme-down CN days.Trend groups use trailing 63-day cumulative market returns, while shock groups use current-day market returns.Source: N MARKET-STATE EVALUATION DETAILS · pp. 25–26, Figure 7 · p. 25
Loading 2610.07006v1…