acceptodds
Under review as a conference paper at ICLR 2027

Align Before Aggregation: Component-Conditioned Evidence Alignment for Multimodal Financial Forecasting

Abstract

Different temporal dynamics within a market window may require different news, yet text-enhanced forecasting typically organizes evidence at the stock or window level. While frequency-aware methods exploit multiscale structure to improve representations and fusion, we examine whether evidence should be aligned separately to temporal components before aggregation. Under dot-product matching and linear query aggregation over a shared candidate set, we show that global alignment produces a normalized weighted geometric mean of component-level alignment distributions, allowing opposing evidence preferences to cancel. Motivated by this observation, CANER follows an align-before-aggregation principle. It decomposes the observed price–volume window into and high-frequency components and calibrates their queries using shared semantic anchors constructed by deterministically verbalizing observable market states. Each component aligns with and selectively absorbs evidence from the same historical news pool before the enhanced states are aggregated for prediction. On CMIN-US, CMIN-CN, and FNSPID, CANER achieves the highest ACC in all nine dataset–horizon settings, exceeding the strongest baseline by 0.34 percentage points on average, and the highest MCC in seven settings. Comparisons with matched parameter counts and news-slot budgets, together with cross-component mismatch analysis, support the predictive value of preserving component–evidence correspondence before final aggregation.

Primary area:
learning on time series and dynamical systems
Keywords:
multimodal financial forecasting, component-conditioned alignment, align-before-aggregation, multiscale time series

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_d8xZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a specific issue in multimodal financial forecasting: The same price-volume window contains changes at different time scales, and they may require different news evidence; If the numerical states are aggregated first and then the news is uniformly matched, will the correspondence between the components and the evidence be lost? The author proposes CANER: Firstly, low, medium and high-frequency components are obtained through wavelet decomposition, and the observable technical states are tranwritten into text semantic anchors to calibrate component queries. Subsequently, each component selects news from the same historical news pool respectively, absorbs information through gating, and finally aggregates the enhanced status to predict the rise and fall. The evaluation is conducted on three datasets and three horizons,showing that CANER outperforms the baseline methods.

Critical strengths

CANER has an inspiring motivation: Market changes at different time scales may rely on different news, and a unified news context may not be able to preserve these differences. How should research evidence correspond to numerical components to have reference value for financial forecasting and broader multimodal time series modeling is very important.

Critical weaknesses

SpecTF has used representations of different frequencies as queries to generate frequency-specific attention outputs for the same text set, then updated the frequency representations and used them for prediction. Thus,SpecTF is an impotant precedent work which inspire many designs of this work (like Align Before Aggregation). But the performance of the method is not significantly better than the SpecTF baseline.

Questions for authors
  1. Could the author provide the efficiency of the method, including the time and space complexity of the model, and the time cost of training and inference? In addition, could the author provide efficiency comparisons with other baselines like SpecTF?
  2. Could the authors provide more details about the implementation of the method, including hyperparameter settings, training strategies, and the datasets details? This will help to better understand the method and its performance.
  3. The comparision of Global versus component-level alignment is not clear. They used different Top-K and news selection. It is hard to where the gains are from. Please provide more details on global alignment setting.
2 comments
AuthorsOpenReview ↗

Response to Reviewer d8xZ

Thank you for the questions on efficiency and the comparison with SpecTF.

Q1. Efficiency. For lookback LL, KK branches, NN articles and up to MM selections per branch, cached inference with fixed widths takes O(KTenc(L)+KTsel(N,M)+L+KL+KN+KM)O(KT_{enc}(L)+KT_{sel}(N,M)+L+KL+KN+KM). Each branch runs the shared encoder (TencT_{enc}) and top-MM selection (TselT_{sel}). The last four terms cover decomposition, anchors, scoring and absorption. Keeping all branch intermediates, with at most linear selection workspace, requires O(KSenc(L)+KL+KN+KM)O(KS_{enc}(L)+KL+KN+KM) space beyond parameters; SencS_{enc} is one encoder's workspace. Uncached inputs add FinBERT encoding time and memory. Here L=30L=30, K=3K=3, N≤64N≤64, M=8M=8: at most 90 anchor weights, 192 article scores and 24 selection slots per prediction.

Same A100 80 GB, CMIN-US next-day samples, BF16, training/inference batches 32/1:

MetricCANERSpecTF
Training (s/epoch)40.239.0
Time to best validation checkpoint (min)13.512.0
Cached inference (ms/sample)7.314.89
Inference with encoding (ms/sample)35.3614.26
Peak training tensor memory (GiB)1.090.65
Peak inference tensor memory with encoding (GiB)0.610.67

Timing including encoding starts from the same tokenized inputs. Epoch times are close; CANER uses more training memory and has higher inference latency.

Q2. Implementation and data. Two-level db2 wavelets yield three frequency branches. One frozen FinBERT encoder handles anchors and news. For 1/3/5-day direction prediction we use binary cross-entropy, AdamW (LR 1e-4, weight decay 1e-5), batch size 64 and up to 100 epochs. We stop after 10 epochs without improvement in validation MCC, select checkpoints by validation MCC and report five-seed test means ± SD.

Train: 2018-01-01–2021-04-30; validation: 2021-05-01–2021-08-31; test: 2021-09-01–2021-12-31. CMIN-US has 110 US stocks; CMIN-CN has 300 CSI 300 constituents. We rank FNSPID stocks by article count over the full period, so the top-465 universe reflects test-period coverage. We split by target date, drop boundary-crossing samples and use per-stock training statistics for numerical features and missing values. We deduplicate news, include only articles available by the cutoff and exclude same-day news if timestamps are unreliable.

Q3. Global control. Price branches, parameter counts, prediction architecture, pool, cutoff and selection slots are matched. Global uses one query for up to 24 articles and shares the context. CANER selects up to eight per branch. Duplicate selections occupy separate slots, so distinct article counts may differ. Jointly, branch-specific selection and absorption raise next-day ACC by 0.64–1.97 percentage points (pp) and MCC by 0.0088–0.0277 over Global. Removing anchors or fixing the gate at 1 lowers ACC on all three datasets; the fixed gate raises CMIN-CN MCC from 0.0846 to 0.0853.

Critical weakness. SpecTF already uses frequency-specific attention over text and updates each frequency representation with its corresponding context. For a given stock, the news most relevant to a short-term price move may differ from the news that informs a more persistent trend. Even when the same article is relevant to both, its predictive value may differ across timescales. CANER explicitly models which articles each component selects and how much it absorbs from each one. Each component uses observable price and volume to calibrate its query, selects articles by content and time since publication, and gates their contributions before aggregation. Against SpecTF, five-seed mean ACC increases by 0.16–0.49 pp (nine-setting average 0.34 pp) and MCC by 0.0012–0.0137. Under Appendix E's daily equal-weight long-only top-three strategy without transaction costs, cumulative returns increase by 3/5/9 pp and Sharpe by 0.04/0.03/0.04 on CMIN-US/CMIN-CN/FNSPID, respectively.

Daily differences (CANER − SpecTF). On each test day, we calculated paired differences on the same stocks within each seed and averaged the five differences. The table gives the 2.5th and 97.5th percentiles of those averages across trading days. All nine ΔACC ranges lie above zero. For ΔMCC, the CMIN-US three-day range crosses zero [−0.0004, +0.0031], while the other eight ranges lie above zero. We do not report a paired confidence interval for the full-test mean gain, so the daily ranges do not establish its statistical significance.

Dataset, daysΔACC (pp)ΔMCC
CMIN-US 1[+0.03,+0.31][+0.0118,+0.0151]
CMIN-US 3[+0.14,+0.48][−0.0004,+0.0031]
CMIN-US 5[+0.33,+0.57][+0.0008,+0.0061]
CMIN-CN 1[+0.14,+0.41][+0.0039,+0.0082]
CMIN-CN 3[+0.22,+0.50][+0.0023,+0.0066]
CMIN-CN 5[+0.30,+0.57][+0.0062,+0.0098]
FNSPID 1[+0.12,+0.44][+0.0052,+0.0097]
FNSPID 3[+0.26,+0.45][+0.0068,+0.0105]
FNSPID 5[+0.33,+0.68][+0.0044,+0.0100]

Thank you for the detailed response and additional experiments. The authors have adequately addressed my questions regarding efficiency, implementation details, and the global alignment comparison.

However, my main concern about the novelty relative to SpecTF remains. SpecTF already performs frequency-specific text alignment before aggregation, and the improvements over SpecTF are relatively modest despite the additional complexity and inference cost. Therefore, I will keep my original score unchanged.

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

The paper presents a machine learning model for multi-modal financial forecasting which combines financial news with historic prices. Their main idea is to match different news to different frequencies in the data (for example some news might cause some high frequency patterns and vice versa).

Critical strengths

S1: The problem seems relevant and the paper provides a clear motivation for considering different temporal components separately.

S2: The proposed architecture is clearly described, and the individual modules are easy to follow.

S3: The evaluation covers three datasets and multiple prediction horizons, with comparisons against several recent methods.

S4: The paper includes ablation studies and analyses of the retrieved news topics and time lags, providing some insight into the model's behavior.

Critical weaknesses

W1: SpecTF already uses frequency-specific queries to attend to textual information and incorporates the resulting contexts into individual frequency components before forecasting. Thus, the proposed align-before-aggregation principle seems already present in closely related work. CANER introduces wavelet decomposition, semantic anchors, and gated news absorption, but it remains unclear whether these constitute a sufficiently substantial methodological advance.

W2: I would like to see some simple baselines like majority-class prediction, logistic regression or XGBoost with technical indicators and straightforward FinBERT sentiment aggregation. This would make it easier to assess how much advantage those more involved architectures bring.

W3: CANER outperforms SpecTF by only 0.16-0.49 percentage points in accuracy, while the reported standard deviations are often larger than these differences. The results are averaged over five seeds without paired significance tests or confidence intervals for the improvements so it remains unclear to me whether the gains are statistically reliable.

W4: The authors reassign news sets between frequency components at inference time without retraining. The resulting performance degradation may reflect a train–test distribution shift rather than the loss of meaningful component–news correspondence. Training with randomized or mismatched assignments would provide a stronger test of the claimed mechanism.

W5: Proposition 1 assumes linear query aggregation, dot-product attention, and a shared candidate set. However, CANER employs nonlinear query–news interactions, component-specific lag penalties, independent Top-K selection, and gated news absorption. Consequently, the derived geometric-mean relationship does not generally characterize the alignment distributions used by the proposed model. Moreover, the experiments do not directly demonstrate that preference cancellation occurs in practice or accounts for the observed improvements.

Questions for authors

See my weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer tCYv

Thank you for suggesting these comparisons.

W1. Contribution. SpecTF already uses frequency-specific queries to attend to text and updates each frequency representation with the resulting context. For a given stock, the news most relevant to a short-term price move may differ from the news that informs a more persistent trend. Even when the same article is relevant to both, its predictive value may differ across timescales. In CANER, each component decides both which articles to select and how much each contributes to its update before aggregation. Observable momentum, volatility and abnormal volume calibrate each query. Each branch selects up to eight articles from the shared pool by content and time since publication, then gates each article's contribution. With three price branches, the same pool, cutoff and 24 selection slots, and matched parameter counts, separate selection and absorption jointly raise next-day ACC over shared context by 1.40, 0.64 and 1.97 percentage points (pp) on CMIN-US, CMIN-CN and FNSPID. Removing the anchors lowers ACC by 0.75, 1.57 and 2.82 pp (paper Table 4).

W2. Simple baselines. Next-day ACC (%)/MCC is shown below. Logistic regression (LogReg) and XGBoost use technical indicators. Adding mean FinBERT sentiment to LogReg improves both metrics on all datasets.

MethodCMIN-USCMIN-CNFNSPID
LogReg52.38 / 0.031653.92 / 0.043457.77 / 0.0522
XGBoost52.91 / 0.038554.31 / 0.048459.01 / 0.0597
LogReg + FinBERT sentiment53.66 / 0.042554.95 / 0.059560.12 / 0.0728
CANER54.82 / 0.057456.88 / 0.084662.69 / 0.1089

W3. Gain over SpecTF. CANER's five-seed mean ACC on the full test set exceeds SpecTF's by 0.16–0.49 pp across the nine settings (unweighted mean 0.34 pp). On CMIN-US at one day, for example, the full-test mean gain is +0.16 pp, while the central 95% range of daily ΔACC is [+0.03, +0.31] pp. For each test day, we computed the difference on the same stocks within each seed, averaged the five paired differences, and took the 2.5th and 97.5th percentiles across trading days. All nine ΔACC ranges and eight ΔMCC ranges lie above zero. The CMIN-US three-day ΔMCC range crosses zero (see the final table). These ranges describe variation across days. We do not report a paired confidence interval or significance test for the full-test mean gain, so its statistical significance remains undetermined.

Under the Appendix E backtest, both models use the same test dates and stock universe. Each day we equally weight the three stocks with the highest predicted next-day rise probability, excluding transaction costs. CANER's cumulative return exceeds SpecTF's by 3, 5 and 9 pp on CMIN-US, CMIN-CN and FNSPID. Sharpe ratios rise from 0.37 to 0.41, 0.33 to 0.36 and 0.38 to 0.42, respectively.

W4. News reassignment. We randomly reassigned news sets among the three branches per sample in both training and testing, holding architecture and budget fixed. On CMIN-US next-day forecasts, the original assignment achieves 54.82% ACC / 0.0574 MCC versus 54.32% / 0.0521 for this control. Using the same random assignment rule during training and testing mitigates the concern that reassigning news only at inference introduces a train–test distribution shift. The original assignment still gains 0.50 pp in ACC and 0.0053 in MCC over this control, providing further evidence for the predictive value of component–news correspondence.

W5. Preference cancellation. With shared candidates and keys, dot-product softmax and a linear average of queries, Eq. 4 gives the global log attention ratio for an article pair as a weighted average of the component log ratios. Opposing rankings can therefore offset. The identity does not extend directly to CANER's nonlinear scoring, component-specific lags, top-8 selection and gates. Across branches, scores for the same article differed by at least 1 in 20.0%, 27.7% and 21.5% of sampled records on the three datasets. This measures score differences for individual articles. Detecting opposite rankings requires a pair of articles. W1 and W4 support the predictive value of keeping correspondence, while the frequency and contribution of cancellation in the full model remain unmeasured.

Daily differences (CANER − SpecTF), 2.5th–97.5th percentiles across test days

Dataset, daysΔACC (pp)ΔMCC
CMIN-US 1[+0.03,+0.31][+0.0118,+0.0151]
CMIN-US 3[+0.14,+0.48][−0.0004,+0.0031]
CMIN-US 5[+0.33,+0.57][+0.0008,+0.0061]
CMIN-CN 1[+0.14,+0.41][+0.0039,+0.0082]
CMIN-CN 3[+0.22,+0.50][+0.0023,+0.0066]
CMIN-CN 5[+0.30,+0.57][+0.0062,+0.0098]
FNSPID 1[+0.12,+0.44][+0.0052,+0.0097]
FNSPID 3[+0.26,+0.45][+0.0068,+0.0105]
FNSPID 5[+0.33,+0.68][+0.0044,+0.0100]

I would like to thank the authors for their response. While several concerns have been addressed, I remain unconvinced by the methodological novelty relative to SpecTF and the statistical reliability of the marginal performance gains. I therefore decided to maintain my original score.

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

This paper proposes CANER, a multimodal financial forecasting framework based on the principle of aligning evidence before aggregation. The method decomposes historical price-volume sequences into different frequency components and independently aligns each component with relevant financial news before combining the resulting representations for stock movement prediction.

Critical strengths

Performance better than the selected baselines.

Critical weaknesses

1. Benchmark configuration sensitivity and questionable fairness of the comparisons.

My primary concern is the fairness and reliability of the experimental comparisons.

Financial forecasting performance is highly sensitive to experimental configurations, including the historical lookback window, prediction horizon, feature selection, temporal data splits, news preprocessing, and other training choices. In my view, these configurations effectively act as experimental hyperparameters, and changing them can substantially affect the relative ranking of competing methods.

The paper adopts its own experimental protocol, including a 30-day historical window and prediction horizons of 1, 3, and 5 days. However, the original methods used as baselines were developed and optimized under different experimental configurations. A model optimized for one forecasting setup is not necessarily optimal under another.

Although evaluating all models under a common protocol is necessary, this alone does not establish a fair comparison. It is essential to ensure that the competing methods have been properly adapted and independently tuned for the new setting. The paper does not provide sufficient details to establish that comparable tuning efforts and optimization budgets were applied to all external baselines.

Moreover, the manuscript does not clearly report the exact temporal boundaries of the training, validation, and test splits for each dataset, making it difficult to verify comparability with previous benchmarks.

I therefore find it difficult to interpret the reported rankings as convincing evidence of methodological superiority. They may partly reflect the choice of experimental configuration rather than the intrinsic advantage of the proposed architecture.

A more convincing evaluation would compare methods under their established benchmark protocols and examine whether the relative rankings remain consistent across different lookback windows, temporal splits, and forecasting settings.

2. Marginal improvements over recent baselines and insufficient evidence of robust superiority.

The financial forecasting baselines are relatively dated, with the most recent financial-specific method included being CausalStock (2024). Although the paper includes more recent multimodal time-series methods from 2026, these models were originally designed and optimized for different forecasting settings, which raises additional questions about the fairness of their adaptation.

More importantly, the improvements over the strongest recent baseline, SpecTF, are very small. Across the nine dataset-horizon configurations, the average ACC improvement is only 0.34 percentage points, with individual improvements ranging from 0.16 to 0.49 points.

These differences are small relative to the reported run-to-run variability. The paper does not provide sufficient paired statistical analysis or robustness evaluations to establish whether these improvements are consistent and statistically meaningful.

Given the sensitivity of financial forecasting benchmarks to experimental configurations, such marginal improvements are not sufficient to support a strong claim of superiority without further validation.

Questions for authors
  1. Baseline tuning fairness: Were all external baselines independently re-tuned under the proposed experimental protocol? Could the authors provide the hyperparameter search spaces, tuning budgets, and implementation details for each baseline?

  2. Benchmark compatibility: Could the authors report the exact training, validation, and test date ranges, stock universes, and preprocessing configurations? How do these settings differ from the original benchmark protocols used by CMIN and CausalStock?

  3. Configuration sensitivity: Do the improvements remain consistent across different historical lookback windows and temporal data splits? In particular, does CANER still outperform the baselines when evaluated under their original benchmark settings?

  4. Error analysis and complementarity with existing methods: Could the authors provide a more detailed error analysis comparing CANER with strong recent baselines such as SpecTF? In particular, does CANER correctly predict cases where existing models fail, suggesting that it captures information overlooked by previous methods, or do the models largely succeed and fail on the same samples? An analysis of prediction disagreements, error overlap, and potential complementarity would help clarify whether the proposed alignment mechanism provides genuinely additional information beyond existing approaches.

1 comment
AuthorsOpenReview ↗

Response to Reviewer PhJE

Thank you for the questions on fairness and reliability.

Q1. Tuning. All methods share dates, labels, information cutoffs, 30-day lookbacks and 1/3/5-day horizons. StockNet, CMIN and CausalStock retain their classifiers. iTransformer, TaTS, VoT and SpecTF use a single logit and BCEWithLogitsLoss, retaining auxiliary objectives. Per dataset and horizon we searched each external baseline with one fixed seed, chose its configuration by validation MCC and ran five seeds. The same metric selects checkpoints. Paper Table 1 reports test means ± SD. LR denotes learning rate.

MethodSearch spaceConfigs
StockNetα {0.1,0.5,1.0}3
CMINmemory layers {1,3,5} × dropout {0.1,0.2,0.3}9
CausalStockLR {1e-3,1e-4,1e-5,1e-6} × price hidden size {4,8,16}, lag fixed at 512
iTransformerLR {1e-3,5e-4,1e-4} × layers {2,3,4}9
TaTSmain LR {1e-4,5e-5,1e-5} × prior weight {0,0.1,0.5}9
VoTjoint-training LR {5e-4,1e-5}2
SpecTFLR {1e-5,5e-5,1e-4}3
CANERfixed AdamW, LR 1e-4, weight decay 1e-5, batch 641

VoT retained its staged training. Other methods trained for up to 100 epochs, stopping after 10 epochs without improved validation MCC.

Q2. Protocol. Train: 2018-01-01–2021-04-30; validation: 2021-05-01–2021-08-31; test: 2021-09-01–2021-12-31. CMIN-US (110 stocks) and CMIN-CN (300 CSI 300 constituents) follow CMIN/CausalStock dates. FNSPID uses the 465 stocks with most news over the full period, so test-period news counts enter universe selection. Each prediction uses only information available at its cutoff.

We split by target date, exclude boundary-crossing samples and use per-stock training statistics for features and missing values. Labels compare adjusted closes at prediction time and H trading days later. We exclude same-day news if timestamps are unreliable and retain up to 64 unique articles after exact deduplication. In its original protocol, CMIN used a five-day history to predict the next day's price direction. CausalStock used prices and news from the five periods preceding the target trading day to predict that day's direction. We retrained both with a common input window of 30 trading days and forecast horizons of 1, 3 and 5 trading days. CausalStock's model lag remained fixed at 5.

Q3. Window sensitivity and Weakness 1. On the same next-day samples, CANER exceeded CMIN in mean ACC and MCC with both 30- and 5-day windows. Shortening the window reduced the ACC lead from 1.43 to 0.65 percentage points (pp) on CMIN-US and from 2.55 to 1.46 pp on CMIN-CN. The 5-day MCC gaps were +0.0068 and +0.0026. Both comparisons use the same test period. Robustness across temporal splits remains untested.

Q4. Error overlap. Matching stock, date and run, the models disagreed on 13.16–18.59% of samples across the nine settings. Only CANER was correct on 6.75–9.43% and only SpecTF on 6.41–9.16%. On CMIN-US at one day these shares were 8.27% and 8.11%, giving a net ACC gain of 0.16 pp. CANER thus corrects some SpecTF errors and introduces others, indicating predictive complementarity at the sample level, although this comparison cannot attribute the differences to a particular alignment module.

Weakness 2. Reliability and economic value. CANER's five-seed mean ACC exceeds SpecTF by 0.16–0.49 pp (nine-setting mean 0.34 pp), and MCC by 0.0012–0.0137. ACC gains over CausalStock (2024), our latest finance-specific baseline, are 0.24–2.30 pp. Under the backtest protocol in Appendix E, CANER and SpecTF each form an equally weighted portfolio on the same trading dates by selecting from the same stock universe the three stocks each model ranks highest by predicted probability of a next-day price increase, with no transaction costs. CANER gains 3/5/9 pp in cumulative return and 0.04/0.03/0.04 in Sharpe on CMIN-US/CMIN-CN/FNSPID.

For each test day, we computed ACC and MCC on the same stocks, subtracted SpecTF from CANER within each seed, and averaged the five paired differences. The table below gives the 2.5th and 97.5th percentiles of these daily averages. All nine ΔACC ranges and eight ΔMCC ranges lie above zero. The three-day ΔMCC range on CMIN-US crosses zero [−0.0004, +0.0031]. These percentiles describe variation across trading days. We have not reported a paired confidence interval for the full-test mean gain, so its statistical significance remains undetermined.

Dataset, daysΔACC (pp)ΔMCC
CMIN-US 1[+0.03,+0.31][+0.0118,+0.0151]
CMIN-US 3[+0.14,+0.48][−0.0004,+0.0031]
CMIN-US 5[+0.33,+0.57][+0.0008,+0.0061]
CMIN-CN 1[+0.14,+0.41][+0.0039,+0.0082]
CMIN-CN 3[+0.22,+0.50][+0.0023,+0.0066]
CMIN-CN 5[+0.30,+0.57][+0.0062,+0.0098]
FNSPID 1[+0.12,+0.44][+0.0052,+0.0097]
FNSPID 3[+0.26,+0.45][+0.0068,+0.0105]
FNSPID 5[+0.33,+0.68][+0.0044,+0.0100]

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.