acceptodds
Under review as a conference paper at ICLR 2027

COIN: Coordinating Imputation and Adaptation for Financial Forecasting with Time-Series Foundation Models

Abstract

Missing observations and changing statistical properties in financial time series affect the forecasting performance of pretrained time series foundation models. Selecting imputation methods based on reconstruction error does not guarantee better downstream forecasts. We propose COIN, a hierarchical framework that coordinates imputation and adaptation for rolling financial forecasting. On the data side, COIN feeds histories completed by multiple lightweight imputers separately into the backbone and dynamically weights their forecasts using prediction losses from completed windows. On the adaptation side, it selects a branch for each window from unadapted predictions, residual corrections, and LoRA outputs at different strengths. Both decision layers use only forecasting targets that have become available, without requiring ground-truth values at missing historical positions. The original backbone parameters remain frozen, and LoRA is trained once during initialization; rolling forecasting updates only weights, residuals, and scoring states. Experiments cover 11 financial data sources, four backbone models, and three initial missingness rates. On the development set, COIN reduces average mean absolute scaled error (MASE) by 2.87%–3.67% compared with forecasting pipelines using NuwaTS, MOMENT, Timer, and Moirai. In confirmation experiments on 22 additional series with fixed configurations, COIN achieves lower average error for all four backbones than an STL-Kalman pipeline with matched LoRA adaptation capabilities. Ablations and analyses under different conditions further show that gains vary across data sources, missingness locations, and feedback timeliness. These results support coordinating imputation candidates and output adaptation to improve existing foundation models for financial forecasting when subsequent target feedback is available.

Primary area:
learning on time series and dynamical systems
Keywords:
Financial forecasting, Time series foundation models, Missing value imputation, Online adaptation, Hierarchical decision making

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_YXDQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper deals with an interesting problem of adapting time series foundation models for improved financial forecasting under missing historical input data and changing statistical properties. To address these challenges, the authors propose COIN, an ensemble-based framework that incorporates eight conventional imputation strategies with feedback-based forecast weighting and LoRA-based adaptation of time series foundation models. The proposed framework dynamically combines forecasts from different imputation methods and selects suitable adaptation mechanisms based on previously observed forecasting errors. The authors evaluate their framework using four time series foundation models across multiple financial time series datasets and demonstrate improved forecasting performance compared to several existing baselines.

Critical strengths
  1. The paper addresses an interesting and relevant problem of handling missing historical observations in financial time series forecasting using foundation models. The problem is reasonably well motivated, particularly considering the challenges associated with missing input data and changing statistical properties in financial time series.

  2. The authors evaluate the proposed framework across several financial time series datasets using four different foundation models. The paper also includes some additional ablation studies and computational cost analyses to investigate the effects of different components of the proposed framework.

Critical weaknesses
  1. The proposed framework appears to be a complicated ensemble of several existing techniques rather than a fundamentally novel methodological contribution. The technical justification for integrating multiple imputation, forecast-weighting, and adaptation mechanisms into one framework is not sufficiently convincing.
  2. The proposed framework introduces substantial computational overhead due to multiple imputation-conditioned forecasting evaluations and adaptation mechanisms, while the improvements in forecasting performance appear relatively marginal. The necessity of such a complicated framework is not sufficiently justified, particularly when simpler configurations demonstrate comparable average forecasting performance.
  3. The experimental evaluation lacks systematic comparisons across different forecasting horizons and look-back windows within the same datasets, which limits the assessment of the method's robustness. Furthermore, direct comparisons with established deep learning-based imputation methods and explicitly missingness-aware forecasting approaches are missing, making it difficult to fully establish the effectiveness of the proposed framework.
  4. The proposed adaptation mechanism is not sufficiently clear or convincingly justified from a technical perspective. It appears to primarily rely on conventional LoRA-based fine-tuning followed by selecting among multiple parallel forecasting and residual correction branches based on historical prediction errors. It is not clear what the fundamental technical novelty of this adaptation strategy is, or why such a complicated mechanism is necessary compared to existing simpler adaptation approaches.
Questions for authors
  1. Given that the proposed framework primarily combines existing imputation, ensemble weighting, and adaptation techniques, and considering that much simpler configurations achieve comparable average forecasting accuracy, what is the fundamental technical advantage of the proposed hierarchical design that justifies its significantly higher computational complexity?
  2. Could the authors clarify the main technical novelty of COIN beyond an ensemble of multiple imputation-conditioned forecasting outputs followed by adaptation selection? What makes the proposed hierarchical coordination fundamentally different from existing ensemble learning or online expert-weighting approaches?
  3. How was the number of eight imputation methods decided? Why were these particular methods selected, and why not a smaller or larger number of imputation strategies? Given that the pruning experiments demonstrate comparable average forecasting performance with fewer candidates, how do the authors justify the necessity of the full framework?
  4. Considering the substantial additional computational overhead and relatively marginal forecasting improvements, what is the practical advantage of using COIN over simpler adaptation or ensemble-based forecasting approaches?
  5. Since the primary motivation is to address missingness in financial time series, could comparable improvements be achieved through missingness-aware fine-tuning or domain adaptation of the foundation models themselves, rather than introducing a computationally expensive ensemble framework?
  6. Why were different forecasting horizons and look-back window lengths not systematically explored within the same datasets? How can the authors establish that the improvements remain consistent across different short-term and long-term forecasting configurations?
  7. What was the rationale behind selecting the four particular foundation models used in the experiments? How representative are these models of existing time series foundation models, particularly in terms of their capabilities to handle missing input observations?
  8. Why were established deep learning-based imputation methods such as BRITS, SAITS, and CSDI not included as direct experimental baselines? Additionally, how would COIN compare against forecasting models that explicitly incorporate observation masks or are adapted to handle missing historical inputs?
  9. Could the authors provide a clearer technical explanation of how the proposed adaptation mechanism fundamentally differs from conventional LoRA-based fine-tuning combined with ensemble or output-selection strategies? Since the LoRA parameters remain fixed during online forecasting, in what sense does the proposed framework achieve model adaptation beyond dynamically selecting or correcting forecasting outputs based on historical prediction errors?
11 comments
AuthorsOpenReview ↗

Response to Reviewer YXDQ(Part 1 of 7)

Overview. Thank you for asking how the design holds up against simpler alternatives. Safeguards lower positive excess in the matched comparison. Pruning cuts latency but raises error in the worst scenes. The CSDI gap narrows with longer training. rMAE sets same-scene Linear to 100.

The online feedback changes the forecast output through candidate weights, residuals, and branch choice. LoRA is fitted once on the initial history. We compare native-mask LoRA at a common training budget and report the lookback and horizon grid by backbone. In the exploratory grid, three backbones favor COIN by point estimate in all four cells, while Sundial favors the control (Tables 37, 39 and 41).

To W1, W2, Q1 and Q4. The hierarchy separates candidate weighting from output adaptation. In a fixed flat policy, the safeguards reduce mean positive excess from 0.508 to 0.318 (Holm p=0.020p=0.020), while rMAE changes from 98.353 to 98.511. With safeguards fixed, flat and hierarchical rMAE are 98.511 and 98.478 (CI for the difference [-0.114, 0.225]). The isolated runtime cost of the hierarchy remains open (Table 30).

Preparation plus 20 windows takes 31.336 s for full COIN and 8.870 s for STL with dedicated LoRA. Pruning cuts rolling latency 38.9% (862.74 to 527.17 ms/window) with rMAE 98.407 versus 98.478, and raises the worst-three mean per source by 0.262. Static input aggregation uses two evaluations per window. It is faster on three backbones and close on TimesFM-2.5. Full COIN lowers mean rMAE by 0.709 versus equal-weight input aggregation (Holm p=0.109p=0.109). A three-window feedback delay raises the development composite from 98.720 to 99.202 (Section 5.4). See Sections 5.6–5.7 and Tables 23–25, 32–33.

AuthorsOpenReview ↗

Response to Reviewer YXDQ(Part 2 of 7)

To W4, Q2 and Q9. Discounted weighting scores completions by base forecast losses on arriving observed targets. The shared mixture weights produce PP and QQ for the base and LoRA models. LoRA is fitted once. Online feedback updates candidate weights, stepwise residuals and branch scores. The selector chooses among PP, two residual corrections of PP, (P+Q)/2(P+Q)/2, and QQ. This is output-level adaptation with fixed model parameters. Matched-warm-up ML-Poly scores 98.467 versus COIN's 98.478, showing that alternative weighting can be used within the same design (Sections 3–4, Table 29).

The branch selector compares outputs built from the same mixture of completed histories. For the dynamic residual branch, each forecast step gets its own smoothed error state, updated from the base mixture when observed targets arrive. That state changes the correction applied to the next forecast. The selector scores all five branch outputs on arriving targets and uses those scores for the next choice. The online updates therefore cover both the correction itself and the choice of output.

AuthorsOpenReview ↗

Response to Reviewer YXDQ(Part 3 of 7)

To Q3. Linear and nearest neighbor cover local structure, mean and median cover levels, directional fills carry observations forward or backward, and seasonal median and blend cover seasonality. Across 57 balanced configurations, one, three and eight candidates score 98.884, 98.613 and 98.478. Eight versus one or two gives Holm p=0.048p=0.048 each, and eight versus three gives p=0.063p=0.063. The 1–8 sweep suggests diminishing gains within that range (Table 31).

To vary pool size, we tested eight balanced compositions at each size from one to seven, plus the full pool. The separate three-candidate configuration selected on development cuts rolling latency by 38.9%, with a 0.262-point increase in the worst-three mean. The sweep measures the trade-off through eight candidates. Larger pools remain untested (Tables 24–25 and 31).

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

The paper proposes COIN, a framework that coordinates imputation and adaptation for financial forecasting with time-series foundation models. It combines forecasts from multiple imputations and uses arriving targets to select among unadapted, residual-corrected, and LoRA-adapted predictions. Experiments across 11 financial data sources and four foundation models report modest average improvements, supported by ablations and analyses of feedback dependence and computational cost.

Critical strengths

Forecast-oriented handling of missing data. COIN uses arriving prediction targets to assess the downstream usefulness of different imputations and guide adaptation, without requiring ground-truth values for historical gaps. This addresses a practical limitation of reconstruction-based imputation evaluation.

Critical weaknesses
  1. Insufficient justification for shared imputation weights. COIN determines imputation weights using only the base model’s forecasting errors and reuses them for the LoRA-adapted model. However, adaptation may change which imputations are most useful, and this assumption is insufficiently validated. The hierarchical design therefore limits interactions between imputation selection and model adaptation.

  2. Incomplete characterization of missingness. The experiments introduce artificial gaps into datasets that retain native missing values. Although the artificial masking procedure is specified, dataset-level characteristics of native missingness and the resulting combined missingness—such as missing rates and gap-length distributions—are insufficiently reported. This makes it difficult to assess the severity and representativeness of the evaluated missingness conditions.

Questions for authors
  1. Could you justify sharing imputation weights between the base and LoRA-adapted models, given that adaptation may change which imputations are most useful?
  2. Could you report the native missing rates and gap-length distributions for each dataset?
5 comments
AuthorsOpenReview ↗

Response to Reviewer JENf(Part 1 of 2)

Overview. Thank you for raising candidate ranking and native gaps. On the evaluated windows, shared weights give similar mean rMAE with one score per candidate. We also report missing rates and gap lengths for all 11 sources. rMAE sets same-scene Linear to 100.

We changed only the LoRA candidate scores in the weighting ablation. The missingness table distinguishes native gaps before artificial masking from combined gaps in the confirmation runs’ initial histories. For example, Monash Bitcoin and PortBench Market have similar native rates in the confirmation set, but their longest gaps are 1,437 and 3 positions (Tables 29 and 36).

To W1 and Q1. LoRA can reorder candidates. On the evaluated windows, the rankings largely agree. Across 5,280 feedback windows, the mean Spearman correlation is 0.929, and base and LoRA predictions share at least one best candidate in 84.7% of windows. With candidates and the five-branch rule fixed, independent and shared weights yield rMAE 98.497 and 98.478. The paired difference is 0.019 (95% source-cluster CI [-0.003, 0.050], Holm p=0.560p=0.560). Sharing keeps one score per candidate and has similar mean error in this setting (Table 29, Section G.1).

The shared rule uses base forecast losses to score all completions, keeping the input mixture the same when the branch selector compares base and LoRA forecasts. In the independent replay, the base scores stay the same and the LoRA mixture gets a second set of scores from LoRA forecast losses. The ranking statistics and paired error comparison support using one score vector in the tested setting.

AuthorsOpenReview ↗

Response to Reviewer JENf(Part 2 of 2)

To W2 and Q2. Table JENf-1 covers the selected segments of 38 development and 22 confirmation series. Native rates use each entire selected segment before masking. C1 combines native and injected gaps within the confirmation initial history. Deduplicating shared masks across backbones leaves 66 series-by-seed masks. Gap lengths are calendar positions, reported as median/P95/maximum. Adjacent gaps count as one run and a dash means no native gap. In the confirmation set, Monash Bitcoin and PortBench Market have native rates of 27.01% and 31.04%, yet their longest gaps are 1437 and 3 positions. The statistics characterize the sampled segments (Section G.4, Table 36).

Table JENf-1. Native and combined missingness (revised Table 36).

SourceDev. %Dev. gapConf. %Conf. gapC1 %C1 gap
Binance USDT0.00%—0.00%—19.71%11 / 11 / 93
FNSPID0.00%—0.00%—19.55%12 / 14 / 14
GIFT-Eval M40.00%—0.00%—18.71%2 / 9.2 / 10
Historical Futures0.00%—0.00%—19.71%11 / 39 / 73
Indian Market0.13%1 / 1 / 10.31%1 / 1 / 119.64%13 / 30 / 82
Monash Bitcoin1.86%38 / 75 / 7527.01%21.5 / 727.8 / 143745.12%13 / 79 / 1437
MTBench Finance0.00%—0.00%—19.71%11 / 11 / 88
OpenMedallion0.00%—8.47%96 / 96 / 9631.04%6 / 11.8 / 96
PortBench Market31.86%2 / 3 / 4731.04%2 / 3 / 344.64%2 / 13 / 76
Time Finance3.23%1 / 2 / 60.74%8 / 8 / 820.47%6 / 10 / 18
Twelve Data0.00%—0.00%—19.53%11 / 29.5 / 51
AuthorsOpenReview ↗

Follow-up on our response to Reviewer JENf

Thank you for highlighting the shared-weight assumption and the characterization of missingness. We have addressed these points through a direct shared-versus-independent weighting comparison, an analysis of agreement between base and LoRA candidate rankings, and source-level reporting of native and combined missing rates and gap-length distributions. These additions provide empirical support for the shared-weight choice in the evaluated setting and make the missing-data conditions more transparent. We would appreciate your feedback on whether these analyses resolve your concerns, and remain available to clarify any details.

Reviewer_N4FpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses financial time-series forecasting with incomplete historical observations and distribution shifts when using pretrained time-series foundation models (TSFMs). The authors propose COIN, a hierarchical online decision framework that combines multiple imputation strategies with lightweight forecast adaptation.

COIN first generates candidate forecasts using eight imputation methods and an additional anchor expert. These forecasts are combined through exponentially weighted aggregation based on historical forecasting losses. The framework then selects among five adaptation branches, including no adaptation, residual correction, and different strengths of LoRA-based predictions. As new observations arrive, COIN updates its ensemble weights and decision states without retraining the foundation model or LoRA parameters.

The authors evaluate COIN on 11 financial data sources using four TSFM backbones, with separate development and confirmation sets. The experiments report modest improvements in average forecasting accuracy over existing imputation pipelines and matched adaptation baselines. Additional analyses investigate the effects of delayed feedback, ongoing missingness, individual components, and computational efficiency.

Critical strengths

1. Well-motivated problem and practical hierarchical framework.

The paper addresses a relevant problem in time-series forecasting: accurate reconstruction of missing historical values does not necessarily lead to better downstream predictions. The authors provide empirical evidence supporting this discrepancy. The proposed hierarchical framework offers a practical approach by combining multiple imputation-conditioned forecasts and selecting among different adaptation strategies using historical predictive feedback. An appealing aspect is that the framework does not require ground-truth labels for missing historical values or online retraining of the foundation model, making it potentially applicable to pretrained TSFMs in deployment settings.

2. Comprehensive and relatively rigorous experimental evaluation.

The paper provides an extensive empirical evaluation across 11 financial data sources and four TSFM backbones. The use of a separate confirmation set with frozen configurations, matched adaptation baselines, and component ablations strengthens the experimental design. The authors also investigate feedback delays, ongoing missingness, and computational efficiency. In addition, the statistical analysis includes source-cluster bootstrap confidence intervals and multiple-testing correction. I appreciate the transparent reporting of limitations and results that do not reach statistical significance, which helps provide a balanced assessment of the method's effectiveness.

Critical weaknesses

Insufficient justification for the proposed hierarchical ensemble and adaptation framework.

Although the paper provides several component-level ablations, I believe the experimental evaluation is insufficient to establish whether the proposed hierarchical architecture offers meaningful advantages over existing online aggregation and adaptation methods. In particular, the following comparisons would help isolate the contributions of different components.

1. Comparison with established online expert aggregation methods.

In Section 4.2, COIN updates the weights of nine imputation-conditioned experts according to

\frac{\mu_j\exp(-\eta R_{k-1,j})} {\sum_i\mu_i\exp(-\eta R_{k-1,i})} +\rho\mathbf{1}\\{j=0\\},$$ where $R_{k-1,j}$ is the discounted cumulative relative loss. This formulation is closely related to classical exponential weighting in online learning. Therefore, it would be informative to compare COIN's weighting mechanism with established algorithms, such as ML-Poly implemented in *Online Prediction by Expert Aggregation* (OPERA; Gaillard and Goude, 2016). Using identical expert predictions, available feedback, and downstream adaptation mechanisms would help determine whether the proposed weighting strategy offers improvements beyond existing online aggregation approaches. **2. Necessity of the hierarchical ensemble and shared weights.** COIN first aggregates imputation-conditioned predictions separately for the base and LoRA models: $$P_k=\sum_jw_{k,j}p_{k,j}^{\mathrm{base}}, \qquad Q_k=\sum_jw_{k,j}p_{k,j}^{\mathrm{LoRA}}.$$ However, the paper does not sufficiently justify why this hierarchical aggregation is preferable to a simpler flat ensemble: $$\hat y_k^{\mathrm{flat}} =\sum_j \left( a_{k,j}p_{k,j}^{\mathrm{base}} +b_{k,j}p_{k,j}^{\mathrm{LoRA}} \right),$$ where $a_{k,j},b_{k,j}\geq 0$ and $\sum_j(a_{k,j}+b_{k,j})=1$. Allowing these weights to be updated independently based on past predictive losses would directly test whether sharing imputation weights between base and LoRA models is necessary. Furthermore, COIN performs aggregation at the prediction level: $$\sum_jw_{k,j}f_\theta(\widehat C_{k,j}),$$ whereas an alternative is to aggregate the completed histories before applying the backbone: $$f_\theta\left(\sum_jw_{k,j}\widehat C_{k,j}\right).$$ Since the backbone is nonlinear, these two formulations are generally not equivalent. The latter also requires fewer backbone evaluations. A comparison would therefore help justify the choice of output-level aggregation in terms of both forecasting accuracy and computational efficiency. Finally, directly applying online aggregation to the same five adaptation-branch outputs, rather than selecting a single branch, would help isolate the benefit of the proposed hierarchical selection mechanism. **3. Comparison with existing lightweight online TSFM adaptation methods.** Section 4.3 introduces residual correction and LoRA-based adaptation branches, followed by online branch selection based on historical forecasting losses. However, it remains unclear whether this strategy provides advantages over existing methods for adapting frozen time-series foundation models. For example, *Lightweight Online Adaption of Time Series Foundation Model Forecasts* (ELF; Lee et al., 2025) introduces an online-trained lightweight forecaster and dynamically combines its predictions with those of a frozen foundation model. Comparing COIN against ELF under matched imputation inputs, feedback availability, and appropriately reported computational budgets would help assess whether the proposed adaptation-selection mechanism is more effective than an existing lightweight online adaptation strategy. Such a comparison would require adapting ELF to the missing-history setting. Overall, while the current ablations demonstrate that individual components contribute to COIN's performance, they do not fully establish the necessity of the proposed hierarchical aggregation and adaptation strategy. Given the relatively modest improvements over strong baselines, these additional comparisons are important for determining whether the architectural complexity and computational overhead are justified.
Questions for authors

Following the concerns raised in the Critical Weaknesses, I have two main questions:

  1. Hierarchical ensemble design: Could the authors provide further justification or empirical evidence for the proposed hierarchical output-level ensemble, particularly compared with simpler flat or input-level aggregation strategies? Why should base and LoRA predictions share the same imputation weights?

  2. Effectiveness and practical value: Could the authors clarify whether the proposed adaptation-selection mechanism offers meaningful advantages over existing online adaptation methods, given the relatively modest performance improvements and additional computational overhead?

5 comments
AuthorsOpenReview ↗

Response to Reviewer N4Fp(Part 1 of 2)

Overview. Thank you for suggesting the ML-Poly and ELF comparisons. With matched warm-up, ML-Poly and COIN have close mean rMAE. Shared weights also give similar mean error to independent weights. ELF is faster, while five-branch adaptation lowers positive excess on both matched inputs. rMAE sets same-scene Linear to 100.

The input and output ablation holds LoRA and the branch rule fixed. Input mixing needs two evaluations per window and was faster on three of the four tested backbones. Output mixing keeps a forecast and a loss for each completion, which the online rule uses to update candidate weights (Tables 32–33).

To W1, W2 and Q1. On the same 264 scenarios and feedback, ML-Poly scores 98.503 with uniform initialization and 98.467 with the same two-window anchor warm-up, versus 98.478 for COIN. Flat18 and Soft A5 score 98.356 and 98.689. Their paired 95% intervals versus COIN are [-0.576, 0.348] and [0.002, 0.533], with Holm p=1.000p=1.000 and p=0.449p=0.449. These comparisons leave equivalence unresolved (Table 29).

Shared weights need one candidate score per expert. Across 5,280 feedback windows, base and LoRA rankings have mean Spearman correlation 0.929 and share a best candidate in 84.7% of windows. Independent LoRA weights score 98.497 versus 98.478 with shared weights (paired difference CI [-0.003, 0.050], Table 29). Holding LoRA and the five-branch rule fixed, equal-weight input versus output aggregation scores 99.187 versus 98.920 (Holm p=0.114p=0.114). With the fixed anchor prior, the scores are 99.178 versus 98.850 (Holm p=0.109p=0.109). Input aggregation uses two evaluations per window, compared with 16 for output aggregation. On Chronos-2, Kairos-23M, Sundial and TimesFM-2.5, it takes 126, 156, 306 and 532 ms/window, versus 403, 1308, 2716 and 527 for equal-weight output aggregation. Output aggregation preserves per-candidate feedback for dynamic weighting (Tables 32–33).

The uniform comparison gives each of the eight completions the same weight. The fixed prior assigns half to the anchor and divides the rest evenly across the eight. Applying each policy at the input and output gives two checks on where to aggregate, with candidate weights held fixed in each comparison.

AuthorsOpenReview ↗

Response to Reviewer N4Fp(Part 2 of 2)

To W3 and Q2. Five-branch adaptation lowers the positive excess in both comparisons, while ELF costs less. We fit ELF's Fourier forecaster on complete targets and use the same target masks and rolling feedback. The two pairs hold the base forecast fixed to Linear or COIN's PP. Mean rMAE (ELF/A5) is 99.368/99.366 and 98.951/98.478, with paired intervals [-2.432, 1.981] and [-1.441, 2.081]. Positive excess is 1.817/0.380 and 1.791/0.397 (Holm p=0.038p=0.038 and p=0.008p=0.008). In each pair, 30/264 ELF scenes versus 2/264 five-branch scenes exceed rMAE 105 (Holm p=0.156p=0.156 for each threshold comparison, Tables 3 and 34). Table N4Fp-1 averages the eight workload medians from three repeats, excluding model loading. ELF is faster in both pairs on all eight workloads (Section G.3, Table 35).

Table N4Fp-1. ELF timing on matched inputs (revised Table 35). Times in seconds.

PipelinePrep20 windowsTotalEvals/window
Linear + ELF0.291.842.131
Linear + dedicated LoRA + A53.074.637.702
COIN P + ELF10.176.7116.928
COIN15.4915.9731.5516
AuthorsOpenReview ↗

Follow-up on our response to Reviewer N4Fp

We appreciate your concrete suggestions for testing the ensemble and adaptation design. Following your comments, we have added matched comparisons with ML-Poly and ELF, examined shared versus independent weights, and compared input-level aggregation, output-level aggregation, and alternative branch-combination strategies. The revised discussion distinguishes average forecasting accuracy, positive error excess, and computational cost, making the practical tradeoffs clearer. We hope these comparisons help address your questions about the design choices. Please let us know if any aspect warrants further explanation; we look forward to hearing your assessment.

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

This paper proposes COIN, a hierarchical framework for financial forecasting with time-series foundation models (TSFMs) under missing historical observations and distribution shifts. COIN combines forecasts from multiple imputation candidates using historical prediction losses and dynamically selects among different adaptation strategies, including residual correction and LoRA. Experiments across diverse financial datasets and four TSFM backbones demonstrate generally improved forecasting accuracy over existing imputation pipelines and matched adaptation baselines. However, the improvements are relatively modest, and the primary comparisons do not reach the prespecified significance threshold.

Critical strengths
  1. The paper clearly distinguishes reconstruction quality from downstream forecasting utility, supported by empirical evidence showing weak agreement between the two. Building on this motivation, COIN introduces a hierarchical framework that coordinates imputation and adaptation through forecasting feedback, without requiring missing-value labels or continuous model retraining. The overall design is intuitive and provides a practical approach to handling missing history and distribution shifts in pretrained TSFMs.

  2. The paper evaluates multiple TSFM backbones and diverse financial datasets, with matched adaptation baselines, component ablations, and analyses of feedback delays and ongoing missingness. The careful temporal protocol and transparent reporting of statistical results and limitations strengthen the credibility of the empirical findings.

Critical weaknesses
  1. COIN mainly integrates existing techniques, including online forecast weighting, residual correction, and LoRA-based adaptation, into a hierarchical decision framework. While the design is practical, the two decision layers remain largely decoupled, with imputation weights determined independently of adaptation outcomes. It is therefore unclear what distinctive methodological contribution arises from their coordination beyond combining established techniques.

  2. While COIN improves over matched adaptation baselines, the gains are relatively modest, and simpler alternatives such as Discounted Hedge and flat expert combination achieve comparable performance. It would be helpful to further clarify the advantages of the proposed hierarchical design over these simpler alternatives, particularly given its additional computational cost.

Questions for authors
  1. Could the authors further clarify the key methodological novelty of COIN beyond integrating existing imputation, ensemble weighting, and adaptation techniques? In particular, how do the two decision layers interact beyond sharing forecasting feedback?

  2. Table 9 shows that Discounted Hedge and flat expert combination achieve comparable or slightly better average accuracy. What are the main advantages of the proposed hierarchical structure over these simpler decision strategies?

  3. The improvement of dynamic weighting over equal weighting appears relatively modest. Could the authors further clarify how much of COIN's benefit comes from feedback-driven candidate selection rather than simply combining multiple imputation-conditioned forecasts?

  4. Given the additional computational cost of evaluating multiple candidates and adaptation branches, could the authors discuss whether the observed accuracy and stability improvements justify this overhead, particularly compared with simpler alternatives?

8 comments
AuthorsOpenReview ↗

Response to Reviewer TWpf(Part 1 of 4)

Overview. Thank you for asking how the two decisions work together and what they cost. The matched ablation links lower mean positive excess to the anchor safeguards. Pruning cuts rolling latency by 38.9% with little change in mean rMAE, at the cost of higher error in the worst scenes. rMAE sets same-scene Linear to 100.

The comparison holds the flat architecture and its warm-up fixed, so the drop in excess can be assigned to the weighting rule. The timing comparison reports average error and performance in the worst scenes for both configurations (Tables 24–25 and 30).

To W1 and Q1. The candidate decision changes the predictions available to the adaptation decision. In the confirmation set, reconstruction and first-window forecast rankings have Spearman correlation 0.251, and their best sets are disjoint in 81.4% of scenes (Table 12). As targets arrive, base forecast losses update weights for the eight completions and the anchor. The same weights produce a base mixture PP and a LoRA mixture QQ. The five outputs are PP, two residual corrections of PP, (P+Q)/2(P+Q)/2, and QQ. Adaptation updates residuals and branch scores for the next window, and candidate weights continue to use base forecast loss (Sections 1, 3 and 4).

For example, increasing the weight of Seasonal Median shifts both mixtures toward that candidate’s forecasts. The residual branches then correct that new base mixture. Sharing the weights also keeps the completion mix the same when the selector compares the base and LoRA options. All candidate and branch scores use targets that have already arrived.

AuthorsOpenReview ↗

Response to Reviewer TWpf(Part 2 of 4)

To W2 and Q2. In original Table 9 (revised Table 10), discounted Hedge, the 45-expert flat rule and COIN have mean rMAE 98.313, 98.352 and 98.478. Both paired mean intervals versus COIN span zero. COIN reduces mean positive excess, max⁡(rMAE−100,0)\max(\mathrm{rMAE}-100,0), from 0.768 for the original flat rule to 0.397 (95% CI for the reduction [0.157, 0.698], Holm p=0.020p=0.020). With a conditional flat rule and the same two-window warm-up, adding the anchor safeguards reduces positive excess from 0.508 to 0.318 (Holm p=0.020p=0.020) and changes rMAE from 98.353 to 98.511. With these safeguards fixed, flat versus hierarchical rMAE is 98.511 versus 98.478 (difference CI [-0.114, 0.225], Holm p=1.000p=1.000). The matched results attribute the lower excess to the safeguards. The hierarchy lets us replace either decision separately. Its incremental latency remains to be measured under a matched protocol (Table 30, Section G.1).

The protected rule puts prior weight on the anchor, reserves a minimum share for it, and clips losses relative to it. The controlled flat comparison changes this candidate-weighting rule while retaining the same conditional branch rule and two-window fallback (Table 30).

AuthorsOpenReview ↗

Response to Reviewer TWpf(Part 3 of 4)

To Q3. We selected nearest neighbor on 456 development scenarios and compared it with nine equally weighted experts and dynamic COIN on 264 confirmation scenarios, holding LoRA, warm-up and branch selection fixed. Mean rMAE is 99.175, 98.768 and 98.478, respectively. The reductions are 0.407 for equal weighting versus one candidate (95% CI [-0.105, 0.907]) and 0.290 for dynamic versus equal weights ([-0.133, 0.825]). Both Holm pp values are 0.316 (Section G.2).

The two changes account for a combined 0.697-point reduction in this replay. The larger point estimate comes from combining candidates. Since both intervals span zero, we treat the two contributions as descriptive. Nearest is fixed for the confirmation replay, so the comparison uses the same single candidate in every scene.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.