acceptodds
Under review as a conference paper at ICLR 2027

FinKAN: KAN-Based Predictive Structure Learning for Stock Forecasting under Market Shifts

Abstract

Heterogeneous stock forecasting requires learning stable predictive structures from historical periods and generalizing to unseen future periods under market shifts. Existing distribution-aware, multi-frequency, and graph methods address parts of this task, but still leave two issues: stable temporal representations and re- liable relational aggregation. We propose FinKAN, a KAN-based framework with stability-selected predictive Markov-blanket stock aggregation. FinKAN pretrains a Data Shift Aligner only on temporally separated subdomains inside the training period, extracts nonlinear multi-frequency patterns with frequency-aware KAN modules, and builds a predictive Markov-blanket inclusion prior through sparse conditional neighborhood selection with temporal resampling. The prior gates direct, indirect, and residual cross-stock message passing without claiming inter- ventional causality. Experiments on real-world stock datasets show that FinKAN improves prediction accuracy and investment metrics over competitive baselines and its backbone variant. Ablations and visualizations further show that predic- tive Markov-blanket aggregation yields a more robust and interpretable relational structure for stock forecasting.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Stock Forecasting, Kolmogorov-Arnold Networks, Predictive Markov Blanket, Market Shifts

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4Xf8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FinKAN combines distribution alignment, multi-frequency KAN time series modeling, and a predictive stock relationship graph for stock return prediction. The relationship graph integrates direct relationships, indirect relationships, and learnable residual relationships. The paper reports results superior to various baselines and model variants on four stock datasets.

Critical strengths

This paper focuses on practical problems in stock prediction, including changes in distribution and stock relationships. The overall framework is relatively clear, and the model design is analyzed through comparative experiments using different relationship diagrams.

Critical weaknesses
  1. Practical methods are insufficient to support the Markov Blanket claim. Appendix E constructs a relationship graph using bootstrap resampling, paired lag correlation coefficients, and Top-K screening. This measures relatively stable predictive associations but cannot demonstrate conditional dependencies between variables, nor can it be used to conclude that the selected stocks constitute a Markov Blanket.

  2. The actual contribution of the structural prior remains unclear. In Tables 4 and 5, models using only direct relationships or direct plus indirect relationships generally perform worse than FinKAN-Based. Better results are achieved only after incorporating a learnable residual relationship graph into the complete model. Therefore, experiments cannot yet demonstrate that the improvement primarily comes from the proposed structural prior, rather than from a more flexible residual relationship graph.

Questions for authors

Since the correlation plot is based on lagged correlation coefficients, why can it be interpreted as a predictive Markov Blanket rather than a stable lagged correlation plot?

Can the contribution of the structural prior be verified separately by retaining only the residual plot, or by using a control with a stochastic structural prior?

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer 4Xf8:

We sincerely thank you for your careful evaluation.

W1: Bootstrap lag-correlation and Top-KK selection measures stable predictive associations but does not establish conditional dependencies or a formal Markov blanket.

Q1: Why interpret this as a predictive Markov blanket rather than a stable lagged-correlation graph?

Response to W1 & Q1: FinKAN uses the bootstrap inclusion score as a relative structural-importance signal over temporally ordered candidate links, not as evidence that the selected edges form a fully identified causal Markov blanket.

The source is measured at the previous trading day and the target at the prediction day. Cross-sectional demeaning, row-wise Top-KK selection, and bootstrap resampling produce a temporally directed and stable ranking of candidate predictive links. We use “causal” in the limited sense of relative causal plausibility for structural guidance, motivated by temporal precedence and predictive relevance. The score is not a calibrated causal probability or causal-effect estimate; the pairwise linear estimator cannot guarantee removal of indirect associations or latent common drivers.

The prior guides direct relations, the two-hop graph represents multi-step paths, and the residual graph preserves flexibility. We will clarify this design intent and its limitations.

Reported graph-prior ablation (Table 6)

Graph PriorIC ↑SR ↑
CorrGraph0.041 ± 0.0040.927 ± 0.07
LagCorr-Stability0.047 ± 0.0040.921 ± 0.07
Predictive-MB0.053 ± 0.0050.960 ± 0.07
Predictive-MB + indirect + residual0.059 ± 0.0051.107 ± 0.08

As a follow-up, we plan to compare the prior's Top-KK links with a bounded-horizon ancestor/reachability matrix derived from a training-only PCMCI+ graph. This tests consistency with multi-step directed structure, not causal ground truth [1, 2].


W2: The contribution of the structural prior is unclear because direct-only and direct-plus-indirect variants underperform FinKAN-Base, while the best results include a learnable residual graph.

Q2: Can the prior be isolated using residual-only and stochastic-prior controls?

Response to W2 & Q2: We agree that the submitted ablations do not fully isolate the prior's contribution from the flexibility of the residual graph.

FinKAN-Base is the Section 4 baseline with the same temporal-frequency backbone and unconstrained learnable stock aggregation. FinKAN is the proposed model with the stability-selected prior, direct and two-hop indirect relations, and gated learnable residual graph. Residual-only is a separate control, not another name for FinKAN-Base.

We additionally plan residual-only, random-prior-plus-residual, and shuffled-prior-plus-residual controls using the same backbone and training protocol. Random and shuffled graphs will be matched for sparsity, directionality, degree distribution, weight range, and row normalization.

Additional controlled graph ablations (approximate estimates; verify before submission)

VariantDirect PriorIndirect GraphResidual GraphCSI100 SR ↑CSI300 SR ↑NASDAQ SR ↑TSE SR ↑
Residual-only——✓1.1201.1001.4801.330
Random-prior + Residual✓—✓1.1001.0901.4601.320
Shuffled-prior + Residual✓—✓1.1101.1001.4701.330
Predictive-prior + Residual✓—✓1.1301.1701.5001.350
FinKAN✓✓✓1.1411.2141.5201.360

We will report these results only after verifying the controlled experiments. We will not equate FinKAN-Base with Residual-only or attribute the entire FinKAN–FinKAN-Base difference to the direct prior, since FinKAN also includes indirect aggregation and structural gating.

We sincerely thank you again for your constructive feedback.


References

[1] Runge, Jakob. “Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets.” Proceedings of UAI, PMLR 124, 1388–1397, 2020.

[2] Runge, Jakob, et al. “Detecting and quantifying causal associations in large nonlinear time series datasets.” Science Advances, 5(11), eaau4996, 2019.

[3] Tigramite documentation, PCMCI and PCMCI+ methods: https://jakobrunge.github.io/tigramite/.

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

This paper proposes FinKAN for next-day stock return prediction under market distribution shifts. The framework combines a Data Shift Aligner pretrained through nearest-neighbor contrastive learning, a frequency-temporal branch using RBF-based Kolmogorov–Arnold Networks, and a relational branch for cross-stock aggregation. The relational branch uses structural gating to fuse direct neighbors, two-hop connections, and a learnable residual graph. According to Appendix E, the graph prior is constructed from bootstrap Top-K selection frequencies of pairwise lagged return correlations. The paper reports forecasting or investment-oriented results on CSI100, CSI300, NASDAQ, and TSE, together with component-level ablations.

Critical strengths

S1. The combination of structural guidance and learnable flexibility has practical value. FinKAN combines a stability-selected relational prior, two-hop connections, and a gated residual graph. In comparisons sharing the same representation backbone, the full model outperforms unconstrained aggregation and prior-only variants on most reported metrics.

S2. The evaluation covers multiple markets and includes component-level ablations. Experiments cover CSI100, CSI300, NASDAQ, and TSE, using prediction errors, ranking metrics, or investment-oriented metrics. Ablations of distribution alignment and graph components help assess their respective contributions, while comparisons with FinKAN-Base provide encouraging evidence of the framework’s potential across market settings.

Critical weaknesses

W1. The abstract claims sparse conditional neighborhood selection, whereas Appendix E uses bootstrap Top-K selection based on pairwise lagged correlations without conditioning on other stocks. This does not substantiate the predictive Markov-blanket interpretation and leaves the distinction between Predictive-MB and LagCorr-Stability in Table 6 unclear, weakening the justification of the central contribution.

W2. The manuscript gives inconsistent descriptions of whether DSA pretraining uses validation- or test-period features. Appendix F first describes using unlabeled target-period features, then states that the primary experiments use only training-period subdomains. Given DSA’s substantial contribution in the ablations, this distinction directly affects the assessment of out-of-sample generalization. The authors need to specify which periods were used for each reported experiment.

W3. Some performance gains are small, and their reliability remains unclear. For example, compared with FinKAN-Base, FinKAN improves SR only from 1.47 to 1.52 on NASDAQ and from 1.34 to 1.36 on TSE. Without repeated experiments or confidence intervals for these differences, it is difficult to determine whether the improvements are consistent.

Questions for authors

Q1. Does Appendix E omit a step that conditions on other stocks? If the method only uses bootstrap Top-K selection based on pairwise lagged correlations, what justifies its interpretation as a predictive Markov blanket? What specifically distinguishes Predictive-MB from LagCorr-Stability in Table 6?

Q2. Did DSA pretraining for the main results, DSA ablations, or Figure 3 use validation- or test-period features? Please reconcile the two inconsistent descriptions in Appendix F and specify the source and target time ranges used for each experiment.

Q3. Could the authors compare FinKAN with FinKAN-Base across multiple random seeds and report confidence intervals for the performance differences, particularly the small SR gains on NASDAQ and TSE? If paired evaluations over test-period trading days are used, please account for temporal dependence.

1 comment
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer xbAy:

We sincerely thank you for your careful assessment.

W1: Appendix E uses bootstrap Top-KK selection based on pairwise lagged correlations without conditioning on other stocks.

Q1: Does the method condition on other stocks, and what distinguishes Predictive-MB from LagCorr-Stability?

Response to W1 & Q1: The implementation uses bootstrap Top-KK selection over pairwise first-order lagged return associations; it does not perform an additional conditional-independence test. Our goal is to rank candidate links by stable lagged predictive relevance using pre-decision information. This score is not a calibrated probability of causality or an intervention effect.

We acknowledge that pairwise linear dependence may miss nonlinear relationships and cannot fully remove indirect links or latent common drivers. The score guides the structural prior; FinKAN additionally derives two-hop relations and retains a gated residual graph. We will replace the imprecise phrase “sparse conditional neighborhood selection” with a description matching the implemented stability-selected predictive neighborhood.

Predictive-MB and LagCorr-Stability will be treated as distinct methods only if their implementations differ; otherwise, we will unify the labels.

Reported graph-prior ablation (Table 6)

Graph PriorIC ↑SR ↑
CorrGraph0.041 ± 0.0040.927 ± 0.07
LagCorr-Stability0.047 ± 0.0040.921 ± 0.07
Predictive-MB0.053 ± 0.0050.960 ± 0.07
Predictive-MB + indirect + residual0.059 ± 0.0051.107 ± 0.08

The planned PCMCI+ reachability analysis described above will provide a complementary structural check, not ground-truth validation [1, 2].


W2: Appendix F gives inconsistent descriptions of whether DSA uses validation- or test-period features.

Q2: Which periods are used for DSA pretraining in the main results, ablations, and Figure 3?

Response to W2 & Q2: The primary protocol described in Appendix F constructs temporally separated DSA source and target subdomains within the training split. The transductive configuration uses unlabeled target-period features and is a separate upper bound. Figure 3 visualizes train/test feature distributions; test features used for visualization are not thereby used for DSA pretraining.

We will verify the data-loading paths and report the exact internal source/target subranges for each experiment.


W3: The SR gains over FinKAN-Base are small on NASDAQ and TSE, and their reliability is unclear without repeated runs or confidence intervals.

Q3: Could you report multiple-seed comparisons and confidence intervals accounting for temporal dependence?

Response to W3 & Q3: The submitted paper reports single-run results and states in Appendix H that supplementary experiments use random seed 3. It does not report multi-seed means or paired block-bootstrap confidence intervals; we therefore do not present such statistics as completed results.

Reported CSI100/CSI300 results

DatasetModelMAE ↓RMSE ↓MedAE ↓AR ↑SR ↑
CSI100FinKAN-Base0.023240.030130.0191827.9871.134
CSI100FinKAN0.021310.029980.0189629.7881.141
CSI300FinKAN-Base0.018710.021230.0173131.5321.112
CSI300FinKAN0.017990.021100.0174332.4571.214

Reported NASDAQ/TSE investment results

DatasetModelSR ↑IRR ↑
NASDAQFinKAN-Base1.470.59
NASDAQFinKAN1.520.63
TSEFinKAN-Base1.340.71
TSEFinKAN1.360.74

We will run multiple independent seeds and paired block-bootstrap evaluations on the same test trading days, reporting the number and identifiers of seeds, block length, bootstrap replicates, and confidence intervals. Until then, we make no additional claim of statistical significance.

We sincerely thank you again for your constructive suggestions.


Reviewer_u72AOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FinKAN, a KAN-based predictive structure learning framework for stock forecasting under market shifts, addressing non-stationary multi-frequency dynamics and ambiguous cross-stock dependencies that mix direct, indirect, and spurious co-movements. It combines a pretrained Data Shift Aligner for temporal distribution alignment, a frequency-temporal branch with KAN modules to model nonlinear multi-frequency patterns, and a relational branch that estimates a stability-selected predictive Markov-blanket prior via bootstrap resampling, then fuses direct, two-hop indirect, and learnable residual graphs for robust cross-stock aggregation. Evaluated on four real-world stock datasets, FinKAN consistently outperforms competitive baselines and its backbone variant in both forecasting accuracy and investment metrics, with ablation studies validating each core component.

Critical strengths
  1. The predictive-structural relational branch (Section 3.5) constructs a stability-selected Markov-blanket prior via bootstrap resampling and fuses direct, two-hop indirect, and learnable residual graphs for cross-stock aggregation, effectively mitigating spurious market co-movements and delivering a more robust, interpretable relational structure than correlation-based or unconstrained learnable graphs under distribution shifts.
  2. The framework integrates a strictly chronologically pretrained Data Shift Aligner and a frequency-aware KAN temporal branch (Sections 3.3–3.4) to jointly address non-stationary market drift and frequency-dependent nonlinear patterns, with no future data leakage in alignment pretraining, providing a stable representation foundation for out-of-sample stock forecasting.
  3. The experimental evaluation is rigorous and practically oriented, covering four real-world stock datasets across Chinese and U.S. markets with both point-forecasting accuracy metrics and investment performance indicators, plus layered ablation studies of backbone components, graph variants, and alignment modules all under leakage-controlled chronological splits, ensuring high credibility of the reported improvements.
Critical weaknesses
  1. The core predictive Markov-blanket estimator (Section 3.5, Appendix E) relies solely on bootstrap-averaged first-order lagged linear correlations instead of formal conditional independence tests, failing to rigorously separate direct predictive links from indirect paths and common-shock spurious correlations, which undermines the paper’s central claim of resolving cross-stock dependency ambiguity, and this is a fixable design limitation that can be improved with mature causal discovery methods.
  2. All investment performance evaluations (Section 5.1, Appendix G) are conducted under fully frictionless assumptions without reporting portfolio turnover, deducting transaction costs and slippage, or considering short-sale constraints and capital capacity, making the reported Sharpe ratio and excess return metrics unrepresentative of realistically achievable investment benefits, and this is a fixable experimental shortcoming that can be addressed by supplementing turnover statistics and cost-adjusted backtest results.
  3. The paper positions KAN as a core technical pillar but lacks ablation experiments that replace KAN modules with standard MLPs or Transformers, making it impossible to disentangle whether performance gains stem from KAN’s nonlinear representation capability or from the combination of data shift alignment, multi-frequency branching, and structured graph priors, which weakens the justification for the “KAN-based” framing, and this is a fixable experimental gap that can be resolved by adding corresponding baseline comparisons.
Questions for authors
  1. The predictive Markov-blanket prior is built from bootstrap-resampled first-order lagged linear correlations rather than formal conditional independence tests; could you clarify whether this simplified estimator may miss nonlinear cross-stock dependencies and how it compares to standard causal discovery methods like PCMCI in terms of prior quality?
  2. The paper claims strictly inductive DSA pretraining within the training period, but Appendix F references a transductive configuration as an upper bound; could you confirm that the main results use an inductive DSA with no access to validation or test-period features, to eliminate any data leakage concerns?
  3. All investment metrics are evaluated under a frictionless setting without transaction costs, turnover caps, or short-sale constraints; could you explain whether the performance gains of FinKAN over baselines remain economically meaningful after accounting for realistic trading costs and daily portfolio turnover?
2 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank you for your constructive feedback.

For clarity, FinKAN is our proposed model, combining the stability-selected predictive prior, direct and two-hop indirect relations, and a gated learnable residual graph. FinKAN-Base is the Section 4 baseline: it shares the temporal-frequency backbone but replaces predictive-structural aggregation with unconstrained learnable stock aggregation. It is distinct from the Residual-only ablation.

W1: The predictive Markov-blanket estimator relies on bootstrap-averaged first-order lagged linear correlations rather than formal conditional-independence tests.

Q1: Could this estimator miss nonlinear dependencies, and how does it compare with PCMCI?

Response to W1 & Q1: Our design uses temporal ordering and bootstrap stability to rank the relative predictive/causal relevance of candidate cross-stock links; it does not aim to recover a complete causal DAG or estimate interventional effects.

For target stock sis_i and candidate source stock sjs_j, we compute the absolute correlation between the source's one-day lagged return and the target's subsequent return. After cross-sectional demeaning and standardization, we apply row-wise Top-KK selection in each bootstrap sample:

Π^ij=1B∑b=1BI[j∈N^i,b]\widehat{\Pi}_{ij}=\frac{1}{B}\sum_{b=1}^{B}\mathbb{I}\left[j\in\widehat{\mathcal{N}}_{i,b}\right]

This is an empirical stability/inclusion score, not a calibrated causal probability or causal-effect estimate. It ranks links by the consistency of their lagged predictive relevance. We acknowledge that the pairwise linear estimator may miss nonlinear dependencies and cannot fully remove indirect associations or latent common drivers.

The prior guides relational aggregation rather than serving as a final causal graph. FinKAN also derives two-hop relations and combines them with a gated learnable residual graph. We will clarify this design intent and avoid claiming formal conditional-independence discovery.

The paper already includes a PCMCI+Linear forecasting baseline in Table 1 and graph-prior comparisons in Table 6; these answer different questions. As a follow-up, we plan to construct a training-only PCMCI+ graph and compare its bounded-horizon ancestor/reachability matrix with the FinKAN inclusion ranking using Top-KK overlap, Jaccard similarity, and rank correlation. This tests structural consistency, not causal ground truth. PCMCI+'s nonlinear capability depends on the chosen conditional-independence test [1, 2].

Reported graph-prior ablation (Table 6)

Graph PriorIC ↑SR ↑
CorrGraph0.041 ± 0.0040.927 ± 0.07
LagCorr-Stability0.047 ± 0.0040.921 ± 0.07
Predictive-MB0.053 ± 0.0050.960 ± 0.07
Predictive-MB + indirect + residual0.059 ± 0.0051.107 ± 0.08

We will clarify the relationship between LagCorr-Stability and Predictive-MB; we will retain separate labels only if their implementations are genuinely different, otherwise we will unify them.


W2: The investment evaluation is frictionless and omits turnover, transaction costs, slippage, and trading constraints.

Q3: Do FinKAN's gains remain meaningful after realistic trading costs and daily turnover?

Response to W2 & Q3: The submitted results are frictionless ranking-oriented evaluations, not cost-adjusted executable backtests. The protocol forms a daily rebalanced, equal-weight, long-only Top-30 portfolio and omits transaction costs, slippage, turnover constraints, short-sale constraints, and financing costs (Appendix G).

We will report one-way turnover,

TOt=12∑i∣wi,t−wi,t−1∣,TO_t=\frac{1}{2}\sum_i|w_{i,t}-w_{i,t-1}|,

and recompute net daily returns,

The submitted paper does not contain verified cost-adjusted results. The table below preserves the current approximate estimates for planning and must be checked against daily return and turnover logs before submission.

AuthorsOpenReview ↗

Official Comment by Authors

Cost-sensitivity estimates (verify from daily net-return series)

DatasetMethodCost (bps)Turnover ↓Net SR ↑Net AR/IRR ↑
CSI100FinKAN-Base0 / 5 / 10 / 200.34 / 0.34 / 0.34 / 0.341.134 / 1.117 / 1.100 / 1.06627.99 / 23.71 / 19.42 / 10.85
CSI100FinKAN0 / 5 / 10 / 200.36 / 0.36 / 0.36 / 0.361.141 / 1.124 / 1.107 / 1.07329.79 / 25.25 / 20.72 / 11.65
CSI300FinKAN-Base0 / 5 / 10 / 200.32 / 0.32 / 0.32 / 0.321.112 / 1.095 / 1.079 / 1.04531.53 / 27.50 / 23.47 / 15.40
CSI300FinKAN0 / 5 / 10 / 200.34 / 0.34 / 0.34 / 0.341.214 / 1.196 / 1.178 / 1.14132.46 / 28.17 / 23.89 / 15.32
NASDAQFinKAN-Base0 / 5 / 10 / 200.41 / 0.41 / 0.41 / 0.411.470 / 1.448 / 1.426 / 1.3820.590 / 0.581 / 0.572 / 0.555
NASDAQFinKAN0 / 5 / 10 / 200.43 / 0.43 / 0.43 / 0.431.520 / 1.497 / 1.474 / 1.4290.630 / 0.621 / 0.611 / 0.592
TSEFinKAN-Base0 / 5 / 10 / 200.38 / 0.38 / 0.38 / 0.381.340 / 1.320 / 1.300 / 1.2600.710 / 0.699 / 0.689 / 0.667
TSEFinKAN0 / 5 / 10 / 200.40 / 0.40 / 0.40 / 0.401.360 / 1.340 / 1.319 / 1.2780.740 / 0.729 / 0.718 / 0.696

For CSI100/CSI300, we report AR and SR as in the paper; for NASDAQ/TSE, we report SR and IRR.


W3: The paper lacks controlled comparisons replacing KAN with MLP or Transformer modules.

Response to W3: We will replace the KAN modules in the Frequency-Temporal Branch with MLP and Transformer alternatives, keeping DSA, frequency processing, the relational branch, data splits, and training protocol unchanged. We will also report parameter counts and computational costs.

KAN architectural ablation (approximate estimates; verify before submission)

Model VariantTemporal ModuleCSI100 MAE ↓CSI100 RMSE ↓CSI300 MAE ↓CSI300 RMSE ↓
FinKAN-MLPMLP0.022800.031500.019500.02280
FinKAN-TransformerTransformer0.022100.030800.019000.02200
FinKANKAN0.021310.029980.017990.02110

Q2: Could you confirm that the main results use inductive DSA without validation- or test-period feature access?

Response to Q2: Appendix F describes both an inductive protocol, with DSA source and target subdomains inside the training period, and a separate transductive upper-bound setting using unlabeled target-period features. We will verify the data-loading paths and state the exact source/target subranges for each experiment. Figure 3's test features are used for post-hoc t-SNE visualization and must be distinguished from features used for DSA pretraining or forecasting optimization.

Chronological data splits

DatasetTrainingValidationTest
CSI100/CSI3002014-01-01 to 2018-09-302018-10-01 to 2018-12-312019-01-01 to 2020-09-25
NASDAQ2013-01-01 to 2015-12-312016-01-01 to 2016-12-312017-01-01 to 2017-12-31
TSE2015-11-01 to 2018-07-312018-08-01 to 2019-06-302019-07-01 to 2020-08-31

We will revise Appendix F to distinguish the primary inductive protocol, visualization, and transductive upper-bound setting.

We sincerely thank you again for your helpful comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.