acceptodds
Under review as a conference paper at ICLR 2027

Temporal Distributional Values for Probabilistic Forecasts

Abstract

Probabilistic time-series forecasts describe a range of possible futures. When these forecasts inform health decisions, patients, clinicians and researchers need to understand how patient history shapes the predictions and their uncertainty. For example, in diabetes management individuals may wish to understand how behaviors (e.g., eating, exercise) shift predicted glucose trajectories and how they influence the range of future glycemic states. Specific patient behaviors may have similar effects on a predicted future glucose trajectory but induce substantially different levels of uncertainty in the range of glycemic outcomes predicted by the model. Scalar attribution methods explain a chosen summary of a forecast, such as its mean. However, a single aggregate statistic can conceal probabilistic information: an input may change the predicted variance without changing the mean, or shift the forecast from a uniform to a bimodal distribution. We introduce Temporal Distributional Values (TDV), a model-agnostic method that extends Shapley attribution to probabilistic time-series forecasting. TDV computes distributions of input contributions directly from the model's native probabilistic output, including quantiles and sampled trajectories. We evaluate TDV on four state of the art forecasting models using long-term glucose, meal, and activity data from type 1 diabetes, type 2 diabetes, and health and wellness populations. Experiments demonstrate accurate forecast reconstruction and close agreement between TDV average contributions and exact scalar Shapley values; stability under trajectory and order sampling; and practical computational cost. Case studies reveal previously hidden distributional effects of meals, activity, and glucose history, demonstrating the value of TDV for uncovering a richer understanding of probabilistic forecasts. These explanations give practitioners a new way to examine how patient history shapes the range of outcomes a forecasting model predicts.

Primary area:
interpretability and explainable AI
Keywords:
Explainable AI, Probabilistic Forecasting, Time Series, Glucose Modeling, Healthcare, Shapley values, Bioinformatics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SWjJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Temporal Distributional Values, which treat groups of input history as players and record paired changes in forecast quantiles or sampled trajectories as each group is added under shared randomness, yielding contribution distributions that reconstruct the full forecast. The authors validate reconstruction accuracy, mean consistency and computational efficiency on four models and three glucose cohorts, and provide case studies.

Critical strengths

This paper systematically extends distributional Shapley values to probabilistic time-series forecasting. It clearly distinguishes the explanatory scope supported by marginal quantiles and by joint trajectories, and it retains paired contributions under the same input order and shared randomness, so the baseline plus contributions exactly reconstructs the forecast distribution while the mean agrees with scalar Shapley values. Experiments cover four model types and three glucose cohorts, the numerical checks are thorough, the limitations are discussed candidly, and the time-in-range case studies have a clear application motivation.

Critical weaknesses

The core construction closely follows the Distributional Values of Franceschi et al. 2024. Reconstruction and mean recovery follow directly from a telescoping sum of successive differences, and most propositions restate existing results in the forecasting setting. We suggest clarifying the new theoretical contributions beyond that framework. The spread of the contribution distributions mixes three sources, input orders, reference histories and forecast draws, and variation across orders does not reflect predictive uncertainty. Although the appendix mentions this, the main results do not separate them. We suggest quantifying the variance from each source separately.

Questions for authors

Please refer to weaknesses.

4 comments
AuthorsOpenReview ↗

Response to weaknesses

Thank you for your comments. In order to distinguish the added theory and separate the sources of explanation spread, we have revised the contribution statement and added the quantitative decomposition below.

Comment 1: "We suggest clarifying the new theoretical contributions beyond that framework."

Response: The added result concerns what a forecasting explanation can conclude when the pairing of forecast outcomes is unspecified. For a binary event, we determine the full range of possible contribution distributions for each input. We also prove that every combination of those individual distributions can be realized together by one consistent pairing construction (Corollary 2).

This gives an exact ranking rule. Define an input's switching probability as how often adding it changes whether the event occurs. One input has a larger switching probability under every allowed pairing if and only if its smallest possible value exceeds the other's largest (Corollary 4). The result distinguishes a ranking supported by the forecasts from one that depends on the selected pairing. We further separate changes in event probability from additional spread introduced by pairing (Corollary 3). These results are new to the theoretical basis of DV and specific to time series data.

A supporting temporal analysis computes tighter bounds when early forecast outcomes cannot be paired using later information (Appendix B.2). We credit the established probability and sequential-transport tools used. DV supplies the attribution operator, mean-recovery and efficiency results; the added contribution is this identification and interpretation analysis for forecasting. The characterization concerns each input's contribution distribution, rather than every possible joint dependence among inputs.

Comment 2: "The spread of the contribution distributions mixes three sources, input orders, reference histories and forecast draws, and variation across orders does not reflect predictive uncertainty. Although the appendix mentions this, the main results do not separate them. We suggest quantifying the variance from each source separately."

Response: We now quantify all three sources and their interactions. Intuitively, an explanation can change because we use a different reference history, reveal the inputs in a different order, or draw different forecast outcomes. Their combinations can matter too. These are sources of variation in the explanation, not interchangeable measures of uncertainty about the future.

For the duration target in our controlled dependence-stress setting with sampled empirical references, the shares of contribution variance are:

Source: Share Input order: 1.27% Reference history: 11.17% Forecast draw: 25.99% Order × reference: 2.73% Order × forecast: 7.11% Reference × forecast: 31.47% All three together: 20.26%

The interaction terms account for 61.57%, so reporting only three separate main effects would leave out substantial variation. We decompose each history/input combination first and then aggregate the components. The primary calculation uses 20 fixed histories, all six input groups, exact averaging over orders, 16 references and 256 forecast draws. With a fixed reference, all reference-related components are zero.

We also repeated the calculation across the full roster with 64 references and 1,024 forecast draws; substantial interactions persist. The added explanation in Section 4.1 points to these analyses, with numerical results in Table 16 and the complete decomposition in Appendices B.3 and E. Both budgets and their sensitivity are reported. These measurements concern the controlled synthetic study; sampling error is distinguished from contribution spread.

We have uploaded a revised manuscript and a supplementary version with the changes highlighted in yellow.

After reading Reviewer jZQ4's comments, I noticed additional weaknesses in the paper. In particular, the contribution distributions depend on the chosen coupling, meaning different couplings can produce different explanations while preserving the same marginals. Without a principled way to justify a specific coupling or bound this sensitivity, the resulting explanations risk being arbitrary and unreliable. Consequently, I am lowering my score to 2 to better align with jZQ4' evaluations.

AuthorsOpenReview ↗

Information about the choice of coupling

Thank you for raising this. We agree that two distinct pairing choices can preserve every forecast and still give different distributions of contributions. You ask whether this makes the explanations “arbitrary and unreliable.” However, the revised paper gives both a reason to choose a particular pairing and an exact account of what follows if one declines to choose it.

What is the theoretical reason for this choice of coupling? Franceschi et al. derive their Bernoulli Distributional Value using one uniform draw across coalitions. Their null-player example explains the choice. When adding a block leaves event risk at 50%, the shared draw gives zero switches, whereas independent draws disagree half the time. We use their construction and ask how far their switching score can vary across all pairings consistent with the forecasts. Franceschi et al., ICML 2024, §3.1 and Remark 3.6

Suppose adding a history block raises forecast event risk from 20% to 30%. Every pairing must switch at least 10% of outcomes. Applying both probabilities to the same uniform draw switches exactly 10%, all from absence to occurrence. Independent draws would switch 38%, despite the same ten-point change in risk. The shared threshold therefore counts only switches required by the change in risk, and it achieves that minimum simultaneously for every addition of a block. Matching percentiles at a single forecast horizon answers a similarly explicit question about how a given forecast percentile moves.

For a given history, reference and event, Proposition 11 and Corollaries 2–4 give the full attainable interval for each block’s switching probability. Every combination of values in those intervals is attainable under one coherent pairing across all input subsets. Consequently, block A switches the event more often than B under every pairing exactly when A’s lower bound exceeds B’s upper bound. This certified 750 of 2,250 comparisons in the strong-ranking controls (Appendix F.1). The temporal analysis also shows when to withhold a ranking. Requiring each forecast to preserve its next-step law given both observed pasts narrowed 576 of 768 pairwise intervals in four synthetic cases, yet the resulting bounds certified no strict block ranking there (Appendix B.2).

How does the coupling choice influence detection of faults? In our learned fault-localization study, the shared-threshold explanation can be computed directly from the forecast probabilities. For each block, average its risk increases and its risk decreases separately across input orders. These two quantities depend only on the coalition risks. Corollary 3 proves that any extra switching introduced by another pairing adds equal mass to both signs. Retaining the two quantities localized faults in 62.67% of 1,200 synthetic device histories, versus 52.53% for the validation-selected policy using signed scalar summaries (paired gain 10.14 percentage points, 95% interval 8.86–11.33). Computing them directly reproduced the same inspection decisions (Section 5.8 and Appendix F.4).

We believe this merits publication because it gives a forecaster a way to determine which conclusions require a pairing assumption and to test whether the information that survives can improve a decision. The revised paper supplies an exact test for the first in the binary-event setting and an independently graded example of the second. We respectfully ask you to reconsider the rating in light of those results.

The authors' have addressed some of my concerns, and thus I decide to raise my score.

Reviewer_jZQ4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This work addresses the problem of explaining probabilistic time-series forecasts. The authors introduce Temporal Distributional Values (TDV), a method that extends distributional Shapley attribution to temporal inputs. TDV targets the distribution of marginal contributions across probabilistic forecast samples and input permutations. The approach makes use of shared randomness to couple predictions across coalitions. The approach is evaluated empirically on several probabilistic forecasting models using time-series data, showing numerical reconstruction, compatibility with standard scalar Shapley values, and also illustrative case studies.

Critical strengths

I think the problem of probabilistic forecasts explanations and understanding how different inputs alter predictive distributions is relevant and interesting. Also, the proposed formulation admits different representations, such as point forecasts, quantiles or joint trajectories. Finally, the mathematical developments look sound, and the paper discusses carefully relevant properties and limitations of the proposed method.

Critical weaknesses
  • This paper basically extends Distributional Values (Franceschi et al., 2024) to probabilistic time-series forecasting by defining temporal input groups, applying the existing framework to several probabilistic forecasting interfaces, and considering trajectory-level functionals. These applications are definitely valuable and useful, but I am not sure this contribution is substantial enough for ICLR.
  • In this work, the contribution distributions depend on the coupling used to compare different probabilistic forecasts. Sharing the same forecast randomness across coalitions is very reasonable. However, different couplings can generate different contribution distributions and still produce the same marginal distributions. This is acknowledged in the work, but I think its implications for the interpretation of the explanations require further investigation and discussion.
  • The empirical evaluation largely focuses on showing accurate reconstruction of the predictive distribution and compatibility between expected TDV contributions and scalar Shapley values. These checks are important, but they I don’t think believe are enough to demonstrate the usefulness of the proposed method. It would be very interestign to see controlled synthetic experiments with known ground truths for the effects of different input features on the predictive distribution. This would be particularly important given the problem with the randomness coupling, which undermined the theoretical guarantees of the method.
  • The framework enables to obtain the payoffs by marginalizing over samples from the reference distribution. However, the experiments rely on fixed median references. This may generate unrealistic time-series histories. What’s the justification of this choice? How do the attribution change when using marginalization over samples?
  • I think that presentation could be a bit better. Especially, it would be useful to explain a bit better how to interpret intuitively some introduces terms. For example, after defining omega, it would be worth explaining its meaning and interpretation.
Questions for authors

See the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to weakness and suggestions

Thank you for the specific suggestions. We added theoretical analysis, controlled experiments and reference comparisons, and identify the corresponding changes below.

Point 1

To address this concern, we added three results: an exact characterization of binary-event contribution distributions and switching-probability rankings under different forecast pairings (Section 4.3), a temporal coupling restriction with proved event bounds (Appendix B.2) and a concrete path-event example (Section 4.2), and controlled experiments measuring known effects and inspection/repair value (Section 5.8). The learned experiment improves localization by 10.14 percentage points and improves repair value. These add theoretical guarantees and independently graded decision evidence to the temporal grouping and interface implementation, new from DV's attribution operator and specific to time series data.

Point 2

Reconstruction and mean recovery remain valid for the specified pairing. We added theory and measured comparisons to establish what changes when that pairing varies. Corollaries 2–4 characterize the possible contribution distributions, separate effects of already-revealed inputs from pairing-induced spread, and determine which switching-probability rankings survive every allowed pairing. Section 4.3 explains these consequences.

Empirically, replacing shared pairing with independent or antithetic pairing changes duration-contributions by average distances of 1.0771/1.4458. (Table 10). We also investigate a temporal restriction that prevents early pairings from using later information: 576/768 pairwise intervals narrow, with 12.8% lower mean width across four fixed forecast conditions (Appendix B.2).

Point 3

We added synthetic forecasters with known mean-changing, spread-changing and null inputs, plus suitable scalar controls (Appendix E). The new equipment experiment then asks whether explanations help select a faulty recording to inspect; faults and repair outcomes are graded independently.

On 1,200 independent histories with three fitted finite-state forecasters, TDV localizes faults in 62.67% versus 52.53% for a validation-selected nonlinear policy using signed mean, variance, quantile and distribution-probability explanations. The paired gain is 10.14 percentage points [95% interval 8.86, 11.33]. Compared with that scalar policy, net expected Brier repair gain improves by 0.01533 [0.01390, 0.01672], including inspection cost. Section 5.8 and Appendix F.4 report the stress-linked fault mechanism, all-64-state fleet calibration and controls. Methods retaining positive/negative effects separately or using the full forecasts recover the benefit, identifying the information responsible.

Point 4

Reference choice is modular: TDV can use a mean, median, custom history or distribution of histories. Section 5.1 now makes this explicit. The median supplies a reproducible central comparison.

We added comparisons with 16 sampled whole histories and conditional completion. Duration-contribution distributions change from the median baseline by mean distances of 0.0784/0.1514 time slots, respectively (Table 11). We also report replacement-history support diagnostics (Table 12). Conditional completion answers a different question because retained inputs help infer missing ones. Section 6 now identifies realistic reference selection as further work for perturbation-based explanation techniques generally.

Point 5

We added explanations beside the definitions: W supplies replacement history; Omega selects a forecast outcome, with sharing specifying how outcomes are paired; Pi selects the order of adding input groups (Section 3). We also separate order, reference and forecast-draw variation and their interactions, so order variation is not mistaken for predictive uncertainty (Appendix B.3).

We have uploaded a revised manuscript and a supplementary version with the changes highlighted in yellow.

Thanks for your careful rebuttal. I appreciate the new theoretical additions and experiments, specially regarding the coupling problem. Some of my concerns were addressed, so I will raise my score. However, I still think the novelty of this work is moderate.

Reviewer_6VwzOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work introduces Temporal Distributional Values (TDV) by adapting distributional values to probabilistic forecasting with temporal input groups, marginal quantiles, and joint trajectories. TDV captures variation in each input group’s contribution, including opposing effects that can cancel out in scalar averages. The computed contributions match the corresponding scalar Shapley values on average and reproduce the represented forecast distribution when kept paired with one another and added to the matching reference forecast. Experiments assess scalar agreement across four glucose forecasting models, with additional reconstruction checks using TACTiS-2.

Critical strengths

This work clearly identifies an important limitation of mean-only explanations that can obscure meaningful variation when opposing contributions cancel out. The proposed method is carefully formulated. TDV retains distributions of input contributions, while numerical experiments verify reconstruction of represented forecasts and recovery of the corresponding scalar Shapley values on average. The paper is also clear and well written. Its explicit treatment of assumptions and limitations makes the method’s interpretation and the scope of its claims easy to assess.

Critical weaknesses

The methodological novelty appears incremental, as the core attribution formulation already comes from DV. Proposition 3 establishes that TDV uses the same contribution random variable, formed from paired output differences across input orders and shared randomness. Recovery of scalar Shapley values on average is also inherited from DV. The main contribution therefore appears to be applying this framework to temporal input groups, forecast quantiles, and joint trajectories rather than introducing a novel attribution rule. The case for acceptance would be stronger if the authors can demonstrate what TDV specifically contributes to forecasting-specific challenges beyond what is mechanically offered by DV.

The evaluation provides limited comparative evidence of the added value of retaining contribution distributions. The examples in Appendix D.6 illustrate information that mean-only explanations miss. The TimeSHAP comparison tests scalar agreement rather than whether TDV helps identify specific model effects more accurately. A controlled synthetic experiment could test whether TDV correctly distinguishes inputs that affect the forecast mean, spread, or neither, using a forecaster constructed with known effects. Comparing TDV with suitable scalar explanations on a defined task with known model effects would help clarify what the richer explanations add.

Questions for authors

Beyond Proposition 3, what does TDV add for temporal inputs on top of what is mechanically offered by DV? Could you define a concrete example of a temporal challenge that requires more than applying DV to history windows?

The TimeSHAP comparison in Section 5.4 tests scalar agreement. Can you give a concrete temporal forecasting example where TDV reveals input effects that scalar explanations of the mean, variance, or relevant event probabilities miss? As further validation, a synthetic forecaster with known input effects would help demonstrate this added value.

2 comments
AuthorsOpenReview ↗

Response to Questions:

Thank you for these questions.

Q1:

The added theory gives two results beyond history windows: what remains valid under different forecast-outcome pairings, and how temporal information narrows the explanations.

For a binary event with fixed coalition probabilities, we characterize all jointly attainable individual contribution distributions: every combination within the stated intervals is attained by one consistent pairing across all input subsets. This gives an exact ranking criterion: input i changes whether the event occurs more often than input j under every allowed pairing if and only if i's lower switching bound exceeds j's upper bound. See Section 4.3, Proposition 11 and Corollaries 2–4.

We study a rule requiring each forecast's next-step conditional law to remain unchanged when the other forecast's observed past is also available. A finite-tree backward construction computes the sharp minimum and maximum event-switching probabilities under this rule (Appendix B.2, Equation 12 and Proposition 12). The construction specializes established bicausal transport; the forecasting contribution is the resulting event-law bounds and their interpretation.

A fair-coin example illustrates the difference. In forecast A, one fair coin determines the outage state in both hours. In forecast B, the first hour is clear and the second has a fair-coin outage. Both second-hour risks are 50%. Unrestricted pairing can make the final outcomes always agree or always disagree, giving switching probabilities from 0 to 1. Under the temporal rule, B's second-hour risk must remain 50% after A's first outcome is observed; switching is then exactly 1/2. The information structure changes the explanation while marginal event risks stay fixed.

In synthetic experiments with four fixed forecast conditions, the temporal restriction narrows 576/768 pairwise intervals and reduces mean width by 12.8%. Appendix B.2 specifies this comparison assumption and the conservative aggregation of pairwise bounds. Section 4.2 additionally states which joint-forecast information path-dependent targets require. There is no support for this scenario present in DV and we develop this new theoretical work specific to time series data.

Q2:

Consider two sensors monitoring a machine. The model assigns higher interruption risk when their readings disagree. Starting from a reference with two low readings, including sensor A’s high reading alone raises predicted risk. But if sensor B’s high reading is already included, adding A makes the readings agree and lowers risk. Averaging A’s opposing effects can cancel them and make an important input appear unimportant. TDV keeps both directions visible. In our synthetic study, one sensor in each pair is trusted, and faults are more common in components whose sensor disagreements most strongly affect interruption risk. The retained information helps choose which recording to inspect.

We test this as follows:

  • Known effects and exact information loss: we constructed a synthetic forecaster with mean-changing, spread-changing and null inputs; TDV and appropriate scalar explanations recover their intended effects (Appendix E). A separate 1,650-game study includes 300 matched pairs with identical signed duration-CDF allocations but different context effects (Appendix F.1). This isolates information discarded by taking means.
  • Learned localization and repair: three independently fitted finite-state forecasters are evaluated on 1,200 independent synthetic device histories. TDV achieves 62.67% versus 52.53% for a validation-selected nonlinear policy using signed mean, variance, quantile and event-probability explanations. Averaging models within history gives a paired 10.14-point gain [95% interval 8.86, 11.33]. Compared with that policy, net expected Brier repair gain improves by 0.01533 [0.01390, 0.01672], after inspection cost. Choices precede revelation of clean readings; faults and repair benefits are graded independently of attribution scores (Section 5.8; Appendix F.4).
  • Decision consequence in a synthetic oracle model: on 4,200 independent instances, retaining the law improves net paid-inspection benefit by 0.01979 [0.01825, 0.02138] over the Bayes-optimal policy conditional on the specified rich signed-summary interface. This isolates the cost of lost information (Appendix F.3).

The learned study tests fault-policy transfer in a fleet calibrated at all 64 input states under the stated stress/fault process. Full no-fault, cost and sensitivity results are in Appendix F.4.

We have uploaded a revised manuscript and a supplementary version with the changes highlighted in yellow.

Reviewer_6VwzOpenReview ↗

Thank you for substantial updates. Revising my review to weak acceptance

Thank you for the response and substantial revisions. The added theory clarifies the contribution beyond applying DV to history windows, and the controlled experiments address my request for validation beyond scalar agreement. I appreciate the added synthetic experiments, which directly address my request for validation with known input effects and demonstrate improvements in fault localization and repair over the selected scalar-summary baselines. However, I still have some reservations. The attribution rule remains inherited from DV, and the demonstrated gains are over the selected signed-summary baselines instead of DV or direct methods retaining full forecast information. TThe temporal restriction narrowed the bounds but did not establish input rankings valid across all allowed pairings, leaving its practical benefit unclear. Nonetheless, these additions address my main concerns leading to the previous recommendation for rejection, and I am updating my recommendation from weak rejection to weak acceptance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to all reviewers

We thank all reviewers for their thoughtful feedback, and especially those who have followed up and revised their assessments after considering our responses. The discussion identified two central priorities: demonstrating the value of contribution distributions through controlled experiments and establishing the theoretical contribution relevant to time series data beyond applying DV to history windows. Our revisions address both, alongside clearer treatment of reference selection and the sources of contribution variability.

The added synthetic experiments establish both recovery of known input effects and practical value beyond the tested signed-summary explanations. We verify mean-changing, spread-changing, and null effects, then demonstrate improved fault localization and repair decisions using fitted probabilistic forecasters. These studies directly connect information preserved in contribution distributions to independently evaluated outcomes, extending the evidence beyond reconstruction and scalar agreement (Section 5.8; Appendices E–F).

The added theory makes more clear the connection with time series data and establishes which conclusions are supported across forecast pairings and how temporal information constrains explanations. We characterize the jointly attainable binary-event contribution distributions, derive an exact criterion for rankings valid under every admissible pairing, and obtain sharp pairwise bounds under temporal information constraints (Section 4.3; Appendix B.2). These results provide a forecasting-specific analysis of identification and interpretation beyond DV’s inherited attribution operator. We welcome any remaining questions and respectfully invite reviewers to update their assessments in light of these additions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.