acceptodds
Under review as a conference paper at ICLR 2027

VARIS: Relational Market-State Learning for Financial Time-Series Foundation Models

30 views

Abstract

Time-series foundation models have transformed forecasting from task-specific fitting into transferable temporal representation learning, yet this progress remains centered on individual sequences. Financial time-series foundation models further learn domain-specific asset dynamics from large-scale market data, but still largely treat each asset as an independent forecasting unit, leaving cross-asset structure implicit at prediction time. We introduce VARIS, a relational framework for financial time-series foundation models that represents each prediction time as a synchronized market cross-section and explicitly infers its time-varying cross-asset structure. VARIS separates latent common market components from asset-specific residual evidence, and performs reliability-aware lead–lag inference to jointly determine whether a useful cross-asset relation exists and, conditionally, its source asset and temporal lag. A state-conditioned dynamic layer router further assigns input-dependent weights to hierarchical temporal representations for individual, common-component, and relational reasoning. Across six markets and three forecasting horizons, VARIS improves mean RankIC by 46.8% and RankICIR by 52.3% over the strongest financial time-series foundation model baseline, with consistent gains in 16 of 18 evaluation settings. Controlled ablations, temporal negative controls, and unseen-asset evaluations show that these improvements arise from common-component residualization, relational abstention, and dynamic layer routing rather than increased parameterization or broad market co-movement. VARIS advances financial time-series foundation modeling from asset-level temporal representation toward prediction-time market-structure inference.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Foundation Models,Multivariate Time-Series Forecasting,Relational Representation Learning

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_Tj1iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VARIS, a framework for cross-asset financial forecasting built on a frozen Kronos backbone. It combines dynamic layer selection, common market factor modeling, and selective use of lagged residual information from other stocks. Experiments across six equity markets and three forecast horizons report improvements in return prediction and stock ranking, supported by ablation and generalization studies.

Critical strengths

Cross-asset forecasting better reflects real financial markets and provides a much-needed extension to Kronos by incorporating dependencies between stocks.

The experimental design is comprehensive, covering multiple markets, forecast horizons, ablations, and generalization tests.

Critical weaknesses

Several mathematical symbols are not clearly defined when first introduced, making the method difficult to follow. Please explain their meanings and clarify the dimensions of key tensors.

There is a potential overlap between the evaluation data and Kronos’s pretraining data. Kronos’s published pretraining range extends through June 2024, whereas VARIS’s main test period precedes this cutoff and its OOT evaluation begins in May 2024. The authors should verify the overlap for the specific checkpoint used and separately report results after its pretraining cutoff. Freezing the backbone does not eliminate this concern.

The choice of Kronos and its replaceability are insufficiently discussed. An alternative backbone such as univariate TSFM (PatchTST-FM, Tabby) and multivariate TSFM (Chronos-2, TimesFM-3, Toto-2) could help test whether VARIS’s gains transfer beyond Kronos. Native multivariate models would provide relevant forecasting baselines. TabPFN(-TS) and KumoRFM also merit discussion as alternative tabular and relational prediction approaches. These comparisons would clarify whether the improvements depend on Kronos’s financial pretraining or demonstrate a more general benefit of the proposed relational adaptation.

Questions for authors

Please define mathematical symbols when they first appear.

Please verify whether the evaluation data overlap with the pretraining data of the specific Kronos checkpoint used. Disclose its pretraining cutoff and report results on a non-overlapping test period to address potential data leakage.

Please justify the choice of Kronos and discuss whether it can be replaced by other foundation models.

1 comment
AuthorsOpenReview ↗

Why the Relation Path Is the Incremental Contribution

We thank the reviewer for asking which part of VARIS is responsible for the improvement over existing cross-asset methods.

The relation-path ablation gives a direct answer. In Table 3, removing the relation path reduces Test RankIC from 0.0948 to 0.0764 in A-shares and from 0.0607 to 0.0447 in S&P 500. The corresponding gains from the relation path are therefore 0.0184 and 0.0160. The no-relation variant keeps the individual and common-component paths unchanged, so this comparison isolates the contribution of prediction-time residual relation inference.

The similarity between the no-relation variant and MASTER or DeltaLag is also informative. These baselines already capture shared or lagged information. VARIS adds a further decision: after removing the common market component, it determines which source--lag evidence is relevant to each target asset and whether that evidence should be used.

The relation path first estimates a common market state and uses it to construct residual representations. It then scores source--lag candidates according to their current compatibility and their consistency in previous windows. The retained candidates compete with a NULL option in the same softmax. The weights q_t distribute information across accepted candidates, while rho_t controls the amount of relational information entering the update. State-conditioned routing selects the backbone representations used for individual prediction, common-state estimation, and relation inference.

The routing--selection matrix is:

RoutingRelation selectionA-sharesS&P 500
StaticForced0.07760.0462
StaticSelective0.08470.0528
DynamicForced0.08310.0499
DynamicSelective (VARIS)0.09480.0607

With selective relation updates, dynamic routing improves RankIC over static routing by 0.0101 in A-shares and 0.0079 in S&P 500. The corresponding interaction contrasts are 0.0046 and 0.0042. These results show that the choice of representation and the decision to accept relational information both contribute to the final prediction.

The negative controls show what information the relation path uses. Source permutation lowers A-shares RankIC to 0.0773, lag randomization lowers it to 0.0842, and past-window displacement lowers it to 0.0798, compared with 0.0948 for VARIS. The improvement therefore depends on the source identity, the temporal alignment of the lag, and the consistency of the relation in the preceding window.

Overall, the relation path contributes more than an additional lagged feature. It removes the shared market component, selects relevant residual source--lag candidates, and decides whether their information should enter the prediction. This conditional use of relational evidence is the main incremental contribution of VARIS over the existing cross-asset baselines.

Reviewer_TbH1OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

VARIS sits on top of a frozen Kronos-base backbone and adds a cross-sectional layer for stock ranking. Per asset and per path (individual, common, relational), it mixes four layer snapshots with a router conditioned on a market summary. Learned factor queries estimate a common market state, and relations are scored on residuals after that state is removed. SLRA picks the top 2 source-lag pairs (lags 1, 2, 5), adjusts their scores using two past windows, and puts them in one softmax with a learned NULL logit, so the model can decline to use any relation. Training uses pairwise ranking loss plus Huber. Evaluation covers six equity universes at a 5-day horizon against Kronos (probe and adapted), HIST, MASTER, LIFT and DeltaLag, with ablations, an OOT period, held-out stocks and a long-only backtest in the appendix.

Critical strengths
  1. The SLRA design makes sense. Residualizing on a common state before lead-lag scoring, then letting NULL compete in the same softmax, is a neat way to avoid forced retrieval when no source is informative. The 2x2 in Table 3 (static/dynamic routing x forced/selective) is the right experiment to run, and the gap from forced to selective is visible in both focal markets.

(2) Breadth. Six markets, paired date-block bootstrap for the main gain (Table 6), and a real OOT split that the authors report even where it looks bad (Nifty CI crosses zero, Nikkei point estimate negative).

Critical weaknesses

(1) Survivorship. The S&P 500 universe is a fixed 472-stock snapshot of current constituents, applied back to 2000 (Appendix A). The other four index universes are listed as "index constituents", which I assume means the same thing. This is look-ahead in the universe itself. It inflates IC levels and backtest returns, and it matters for a relational model in particular, since co-movement among surviving firms is not the same as co-movement in the investable universe at that time. The conclusion mentions point-in-time constituents as future work. For a paper whose headline is about cross-sectional structure, I think it is a prerequisite.

(2) The backtest is too thin to support any economic claim. Appendix C.3 covers the OOT period only, long-only top decile, and compares against two baselines. There is no excess return over the equal-weight benchmark, no Sharpe, no drawdown, and no turnover number. 10 bps one-way is optimistic for A-shares once stamp duty is included. No A-share tradability filters are described (limit-up/down, suspensions, ST). The CN panel goes down to 1 eligible stock on some dates. The label is close-to-close from t, but execution is at the t+1 open, and I could not tell which return the backtest actually uses.

(3) The headline numbers are oversold. The 46.8% and 52.3% figures are against adapted Kronos, which is the weakest active baseline in Table 1. Against DeltaLag on A-shares the gain is about 20%. The 1-day and 20-day results behind "16 of 18" are not shown anywhere I could find. Neither are the seed variance, the "held-out-market transfer", or any parameter-matched control, even though the abstract says the gains are not due to parameterization. Ablations and negative controls are single numbers with no intervals, and the interaction contrast (0.0046) has no CI. It is also unclear whether HIST, MASTER, LIFT and DeltaLag get Kronos features or raw OHLCV. If they get raw OHLCV, Table 1 confounds backbone with method.

Smaller: OOT gains shrink in all six markets, which looks like the design was tuned on Test. RankIC at 75% source retention is higher than at 100% (Fig. 4C). On Nikkei, VARIS has lower RankICIR than LIFT and MASTER.

Questions for authors
  1. Can you rerun at least S&P 500 and A-shares on point-in-time constituent universes, with delisted names included? How much of the gain over DeltaLag survives?

  2. In the backtest, please report net excess return over the benchmark, Sharpe, max drawdown and average daily turnover for all baselines, with A-share limit and suspension filters and realistic A-share costs. Is the return measured open-to-open from t+1?

  3. Do the cross-asset baselines use the same frozen Kronos representations as input? What are the trainable parameter counts?

  4. Where are the 1-day and 20-day tables and the seed variance behind the abstract claims?

Reviewer_FDU8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Paper states that financial foundation models train jointly on many stocks’ price history, but forecast each stock individually at inference time. Propose a method with additional inductive bias for leveraging market state and correlation between related stocks at inference time. Train using price-volume histories from 6 stock markets. Objective is a pairwise ranking loss to push stocks with higher realized returns to get higher ranking scores, plus a Huber loss on the predicted return itself. Use a frozen Kronos backbone to encode each stock’s history. Then train adapter layers with learned softmax mixing, which taps different layers of the network to produce separate features for individual assets, market-wise summaries, and lead-lag prediction. Features from these intermediate layers are then combined into an update rule to produce a single feature that is used to feed the two loss heads (rank and return prediction). Main task predicts five-session returns from 128 trading sessions of history. Measure RankIC, the daily cross-sectional Spearman correlation between ranking scores and future returns, and RankICR, temporal mean of RankIC divided by standard deviation.

Critical strengths

Motivation of the paper is logical, forecasting should also leverage cross-asset information; this information can be used beyond just foundation time-series model pretraining. Compare to relevant baselines leveraging cross-asset prediction. See modest gain in RankIC compared to cross-asset baselines

Critical weaknesses

Clarity is quite poor. Most of the paper is dedicated to describing the method, and required significant effort to understand the precise setup. In the end, my understanding is that a FROZEN Kronos backbone is used, adapter layers are trained to produce intermediate representations and mix features across assets, that then combine into a specific update rule to provide a single feature vector that feeds into the the two loss prediction heads (but please correct me if my understanding here is lacking).

Difficult to understand significance of results, lack of standard errors beyond Table 6, and in particular, lack them for the relevant cross-asset baselines.

Most of gain is due to lead-lag relation, without it performs similarly to DeltaLag and MASTER. Difficult to understand core contribution of the work, what are the important components. Many of the ablations show small differences with no error bars either.

Minor, but abstract claims three horizons, but only see results for five-day results?

Questions for authors

Please see weaknesses above. Happy to consider updating score based on author feedback. Main issue however is clarity of the work and an understanding of the significance of the results, advantage over cross-asset baselines, complexity of the method, and a distillation of the important contributions in the work.

3 comments
AuthorsOpenReview ↗

Clarification of VARIS's Three-Path Flow and Tensor Dimensions

We thank the reviewer for asking us to clarify the method flow and tensor roles. Your high-level reading is close, but one important distinction is that VARIS does not concatenate all asset features into one undifferentiated vector. A frozen Kronos pass first produces layer snapshots, and three path-specific routers then construct representations for individual prediction, common-component estimation, and relational inference. The relational path is residualized with respect to the common state before source--lag evidence is evaluated.

For one market-date batch, the notation and dimensions are:

SymbolShapeRole
X_t[B, N_t, T, 6]OHLCV plus amount histories for assets observed at date t
y_t^(h)[B, N_t, 1]Future-return labels for horizon h
H^(k)[B, N_t, T, D]Frozen Kronos snapshot at layer/depth k
alpha^m[B, N_t, K]Layer weights for path m: ind, com, or rel
H^m[B, N_t, T, D]Weighted representation for path m
h_ind_t[B, N_t, D]Latest individual-path representation
C_t[B, F, D]Learned common-market factor state
c_panel_t[B, N_t, D]Current cross-sectional panel context
c_com_t[B, N_t, D]Common-component context for each target asset
c_rel_s_given_t[B, N_t, T, D]Relational common context over the history window
r_s_given_t[B, N_t, T, D]Relational residual after removing common context
tilde_r_s_given_t[B, N_t, T, d_r]Residual projected into the relation space
u_ij_l_t[B, N_t, N_t, L]Compatibility scores for target i, source j, and lag l; L is the number of candidate lags
S_tset; [B, N_t, S, 2] if materializedRetained source--lag pairs; the last dimension stores source and lag indices
e_t[B, N_t, S]Historical-consistency-adjusted scores for retained candidates
p_null_t[B, N_t, 1]Probability of the NULL relation
rho_t[B, N_t, 1]Probability of accepting relational information
q_t[B, N_t, S]Conditional weights over accepted source--lag candidates
d_rel_t[B, N_t, D]Conditional relational content
Delta_rel_t[B, N_t, D]Accepted relational update, scaled by rho_t
z_t[B, N_t, D]Fused individual, common, and relational representation
rhat_t, yhat_t[B, N_t, 1]Ranking-score and return-prediction head outputs

Here B is batch size, N_t is the number of assets observed in the market cross-section at date t, T is the input-history length, K is the number of routed Kronos snapshots, D is the backbone representation width, F is the number of learned common factors, d_r is the relation-space width, L is the candidate lag set size, and S is the candidate budget. We leave D, F, and d_r symbolic because their values are configuration details rather than part of the conceptual definition.

The computation is therefore:

  1. The frozen Kronos backbone maps each asset history to K layer snapshots H^(k).
  2. Separate asset- and path-conditioned routers mix these snapshots into H^ind, H^com, and H^rel.
  3. The common path estimates C_t and constructs panel and common contexts.
  4. The relational path subtracts the corresponding common context, yielding residuals r.
  5. SLRA scores source--lag candidates, evaluates their consistency in preceding windows, and compares the retained candidates with a learned NULL logit in one softmax.
  6. q distributes accepted source residual content, while rho controls whether the relational update enters the forecast.
  7. The gated sum produces z_t, which is passed to the ranking and return heads.

Thus, the two prediction heads do receive one final fused representation per target asset, but that vector is formed only after the separate-path routing, common-state estimation, residualization, source--lag selection, and NULL-aware acceptance steps. The Kronos backbone is frozen; the routers, common-context and factor modules, SLRA projections and gates, fusion parameters, and prediction heads are trained under the pairwise ranking plus Huber return objective.

AuthorsOpenReview ↗

Cross-Baseline Uncertainty and Interpretation of RankIC Gains

We thank the reviewer for pointing out that the original uncertainty table only covered VARIS versus adapted Kronos. We agree that cross-asset baselines should be compared on the same dates and target masks. We have therefore added the complete paired-baseline analysis to the supplementary source and reproduce the results below.

RankIC is a bounded daily cross-sectional Spearman correlation, so its absolute numerical gaps are intrinsically modest: a difference such as 0.0159 is an absolute correlation gain, not a 1.59 percentage-point return gain. Such a small decimal gap is therefore a property of the metric scale and should not be judged as if it were a return difference. The relative comparisons are descriptive only. The relative comparisons are descriptive only. For example, VARIS improves over DeltaLag by 0.0159 in A-shares (0.0948 versus 0.0789, 20.2% relative) and by 0.0113 in S&P 500 equities (0.0607 versus 0.0494, 22.9% relative). The point difference is positive in all six markets.

For each baseline m, we compute the daily paired difference d_t^(m) = RankIC_VARIS,t - RankIC_m,t using identical Test dates and eligible-target masks. We use the same date-block procedure as the original Table 6 to estimate the paired standard error and the corresponding normal-approximation 95% confidence interval. All values below are RankIC points.

MarketBaselineDelta RankICSE95% CI
A-sharesadapted Kronos0.02810.0038[0.0207, 0.0355]
A-sharesHIST0.02630.0037[0.0190, 0.0336]
A-sharesMASTER0.02020.0036[0.0131, 0.0273]
A-sharesLIFT0.02300.0037[0.0157, 0.0303]
A-sharesDeltaLag0.01590.0034[0.0092, 0.0226]
S&P 500adapted Kronos0.02210.0045[0.0133, 0.0309]
S&P 500HIST0.02150.0044[0.0129, 0.0301]
S&P 500MASTER0.01340.0043[0.0050, 0.0218]
S&P 500LIFT0.01010.0046[0.0011, 0.0191]
S&P 500DeltaLag0.01130.0048[0.0019, 0.0207]
HSCIadapted Kronos0.03050.0056[0.0195, 0.0415]
HSCIHIST0.02530.0055[0.0145, 0.0361]
HSCIMASTER0.02160.0055[0.0108, 0.0324]
HSCILIFT0.02000.0056[0.0090, 0.0310]
HSCIDeltaLag0.01750.0054[0.0069, 0.0281]
Nifty 500adapted Kronos0.01780.0042[0.0096, 0.0260]
Nifty 500HIST0.01340.0042[0.0052, 0.0216]
Nifty 500MASTER0.00800.0042[-0.0002, 0.0162]
Nifty 500LIFT0.00470.0044[-0.0039, 0.0133]
Nifty 500DeltaLag0.00590.0043[-0.0025, 0.0143]
Nikkei 225adapted Kronos0.01410.0051[0.0041, 0.0241]
Nikkei 225HIST0.01400.0051[0.0040, 0.0240]
Nikkei 225MASTER0.00530.0050[-0.0045, 0.0151]
Nikkei 225LIFT0.00210.0051[-0.0079, 0.0121]
Nikkei 225DeltaLag0.00320.0049[-0.0064, 0.0128]
KRX 300adapted Kronos0.03700.0048[0.0276, 0.0464]
KRX 300HIST0.03420.0048[0.0248, 0.0436]
KRX 300MASTER0.02860.0048[0.0192, 0.0380]
KRX 300LIFT0.02990.0049[0.0203, 0.0395]
KRX 300DeltaLag0.02410.0047[0.0149, 0.0333]

The intervals are above zero for all five baselines in A-shares, S&P 500, HSCI, and KRX 300. In Nifty 500, the intervals for MASTER, LIFT, and DeltaLag overlap zero; the same is true for these three baselines in Nikkei 225. We therefore describe those comparisons as positive point estimates with uncertain stability. This analysis addresses the reviewer’s concern about missing uncertainty for cross-asset baselines while keeping the interpretation tied to paired evidence rather than to point estimates alone.

AuthorsOpenReview ↗

Why the Relation Path Is the Incremental Contribution

The relation-path ablation directly identifies where VARIS obtains its incremental predictive value. In Table 3, removing the relation path reduces Test RankIC from 0.0948 to 0.0764 in A-shares and from 0.0607 to 0.0447 in S&P 500, corresponding to gains of 0.0184 and 0.0160 from the relational correction. The no-relation variant retains the individual and common-component paths, so this comparison isolates the added value of prediction-time residual relation inference. Its proximity to MASTER and DeltaLag is informative: these models capture substantial shared or lagged information, while VARIS further determines which residual source--lag evidence should enter each target forecast.

The central contribution is therefore a conditional relation-inference mechanism. VARIS first estimates a common market state and uses it as the reference for residualization. It then evaluates source--lag candidates using current relevance and historical consistency, places the retained candidates and a NULL option in one softmax, and separates two decisions: q distributes content across accepted candidates, while rho controls the amount of relational content entering the update. State-conditioned routing selects the backbone depths used for individual prediction, common-state estimation, and relation inference. This sequence gives the model a target- and date-specific rule for when relational evidence is useful and how much of it should affect the forecast.

The routing--selection matrix in Table 3 further supports this interpretation:

RoutingRelation selectionA-sharesS&P 500
StaticForced0.07760.0462
StaticSelective0.08470.0528
DynamicForced0.08310.0499
DynamicSelective (VARIS)0.09480.0607

Under selective relation updates, dynamic routing adds 0.0101 and 0.0079 RankIC over static routing in A-shares and S&P 500, respectively. The corresponding interaction contrasts are 0.0046 and 0.0042. At the point-estimate level, these values show that representation selection and relation acceptance reinforce each other. Source permutation, lag randomization, and past-window displacement further reduce A-shares RankIC to 0.0773, 0.0842, and 0.0798, respectively, compared with 0.0948 for VARIS. The gain therefore tracks source identity, temporal alignment, and historical relation evidence.

The preceding paired cross-baseline table now provides SEs and 95% CIs for the main model comparisons. Table 3 reports the mechanism ablations and routing--selection interaction contrasts as point estimates; we use them to localize component contributions and keep their interpretation at the effect-localization level. The consistent directions across the two focal markets and the source/lag controls support the same mechanism interpretation, while the point estimates are not presented as significance tests. Taken together, the ablations identify the contribution of VARIS as conditional selection and acceptance of residual lead--lag evidence, with state-conditioned routing determining the representations used to make that decision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.