acceptodds
Under review as a conference paper at ICLR 2027

SHIR: Evidence-Guided Explanations for Time-Series Predictions

Abstract

Explaining a time-series prediction requires identifying which parts of the input affect it. Perturbation responses recorded on reference inputs can guide query selection, but do not establish the effects on a new input. We introduce SHIR for frozen time-series predictors, separating unit scoring from query selection through evidence-guided experimental design. A language model proposes unit-effect comparisons from recorded evidence and an explicit account of which endpoint effects remain unmeasured. Under a fixed perturbation rule, SHIR selects the query batch whose endpoints would complete the most candidate comparisons within the budget. An evidence graph enables many-to-many evidence reuse: one execution record serves multiple applicable registered comparisons while retaining its input and configuration identity. Reference construction explores inputs in separate episodes and freezes the merged graph and reference scores. For each new input, SHIR preselects one batch of at most three perturbation queries and evaluates the separately perturbed inputs in parallel. Each perturbation starts from the same original input. Valid current effects replace the corresponding reference scores, while unmeasured units retain their defaults. On forecasting tasks built from WeatherBench2 and VitalDB, SHIR reduces mean sufficiency error by approximately 11.2% and 18.7%, respectively, relative to each task's strongest baseline for this metric among seven external methods. Our code is available anonymously.

Primary area:
interpretability and explainable AI
Keywords:
Model interpretability, Time-series forecasting, Large language models

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

In this research the authors present a new method (SHIR) for leveraging time series explanations to improve the quality of explanations for forecasting predictors. They combine reference perturbation scores with a small set of measurements on each new input. Then they have an LLM propose comparisons between predefined input units while an evidence graph shares compatible execution records across comparisons.They provide empirical results on forecasting tasks using WeatherBench2 and VitalDB datasets and show performance gains against other common methods of feature importance selection.

Critical strengths

S1. Good empirical results. The performance shown against the external baselines is strong. The authors put together a thorough evaluation with carefully constructed tasks. The results support the proposed combination of methods that make up SHIR including the graph-only comparison showing that current measurements improve explanation quality and the additional ablations focused on evidence organization and query selection.

S2. I think the core idea of adapting reference information to a particular prediction is interesting - as well as the combination of frameworks. The authors considered efficiency and system design in constructing the final method.

S3. The idea of handling evidence by having the LLM proposes comparisons, while the predictor’s measured responses determine the effects used to update scores is also interesting. The authors explicitly distinguish current measurements from reference defaults and avoid counting reused observations multiple times. This grounds score updates in observed model behavior and makes the resulting explanation easier to audit.

Critical weaknesses

W1. It is unclear exactly how much the LLM step adds independently. The results show that taking three new measurements helps, but a simpler heuristic rule or even a smaller model/less reasoning might choose equally useful measurements. Table 2 also shows a tradeoff in that the complementary selection improves sufficiency error but the reference-score Top-3 has higher comprehensiveness on both tasks. It would be helpful to have a clearer explanation of when comparison completion helps.

W2. It’s also a bit unclear how reliable the improvements are in the empirical setting. The main results seem to only be run once with no variation because there doesn’t seem to have been multiple runs with different seeds. It would be great to see more averages over 5 runs in the final results.

Questions for authors

Q1. How often do the three current measurements change the selected units compared with graph-only?

Q2. What’s the authors intuition for why SHIR can better preserve the prediction when its chosen inputs are kept, while T3 causes slightly larger changes when its choices are removed? Since T3 also uses LLM suggestions, have you already tested a simpler rule for choosing three queries with the same reference scores and score update?

Q3. What’s the e2e latency/cost and scalability in particular the measured latency of llm calls relative to the rest of the computation and how things scale with longer time series?

1 comment

We sincerely thank the reviewer for the effort in reviewing our submission. Our detailed replies are as follows:

Response to hhax-W1/Q2:

We tested simpler rules on 50 WeatherBench2 test inputs with the same reference, perturbation, three-query budget, R0, and evaluation. Global Top 3 and Global Random 3 select from the full catalog without an online LLM; LLM Direct 3 asks the same proposer for three units; T3 selects within SHIR's LLM-derived pool.

StrategyComp ↑Suff err ↓CPD@10% ↑CPP err@10% ↓
Graph-only (0 queries)0.15310.04060.15400.0410
Global Top 30.16280.03280.15860.0304
Global Random 30.15900.03300.15570.0280
LLM Direct 30.14920.03320.15410.0299
T30.16600.03620.15740.0287
SHIR0.16340.03100.15960.0254

SHIR improves all four metrics over both no-LLM rules, reducing CPP error by 16.4% relative to Global Top 3. LLM Direct 3 improves three metrics over Graph-only but lowers Comp, and SHIR beats it on all four with the same proposer and budget, supporting the use of the LLM to propose comparisons for joint batch selection rather than to pick units directly.

To show when completion helps, we compare SHIR with T3 at each size: paired differences (SHIR − T3) over the 50 inputs, Wilcoxon signed-rank tests, Holm-corrected over eight tests.

SizeΔ deletion ↑ (Holm pp)Δ retention err ↓ (Holm pp)
5%−0.0070 (0.033)−0.0015 (0.44)
7.5%−0.0046 (0.033)−0.0088 (0.010)
10%+0.0022 (0.033)−0.0033 (0.44)
15%−0.0010 (0.44)−0.0072 (0.004)

SHIR has lower retention error at every size (significant at 7.5% and 15%) and a significantly larger deletion effect at 10%; T3's higher Comp comes from the two smallest sizes. Our intuition: T3 spends the budget on the highest-scoring candidates, refining the top of the ranking, which matters most when only the first few units are deleted. SHIR instead completes comparisons, prioritizing those that cross top-kk boundaries; this corrects set membership, which retention evaluates at every size. Completion therefore helps most when set membership, rather than the first few units, is uncertain.

Regarding a smaller model or less reasoning, we replaced only the online proposer, keeping Grounding, selector, R0, and budget fixed. All proposers, including the original, run in non-thinking mode. A local Qwen3.5 4B proposer improves all four metrics over the original (e.g., Comp 0.1733 vs. 0.1634; CPP error 0.0238 vs. 0.0254), and Qwen3.8 27B attains the lowest retention errors (Suff 0.0291, CPP 0.0231) at 1.40 s per input. SHIR thus works with a small, non-thinking proposer, since the proposer only nominates comparisons while predictor responses set every score.

Response to hhax-W2:

The ablation variants already average three repetitions, and SHIR remains best on Suff error and both 10% endpoints on both tasks. The per-input paired tests above add input-level evidence: against the closest comparator, SHIR's retention gains at 7.5% and 15% are significant after Holm correction. We will report five-run averages for the main results in the revision.

Response to hhax-Q1:

We compared SHIR's top-kk set with Graph-only under the same reference. Replacements count k−∣SkSHIR∩SkGraph∣k-|S_k^{\mathrm{SHIR}}\cap S_k^{\mathrm{Graph}}|, averaged over all 50 inputs; Δ is SHIR − Graph-only.

SizeInputs changedMean replacementsΔ deletion ↑Δ retention err ↓
5% (k=5k=5)29/500.76+0.0050−0.0030
7.5% (k=7k=7)36/501.00+0.0149−0.0106
10% (k=9k=9)27/500.54+0.0056−0.0156
15% (k=14k=14)25/500.50+0.0157−0.0092

Three measurements change the explanation for 50%–72% of inputs, and SHIR improves both mean deletion and mean retention error over Graph-only at every size.

Response to hhax-Q3:

Online explanation takes 3.33 s per input, 3.12 s (93.7%) of it in LLM calls, and uses 16,088 input and 353 output tokens in non-thinking mode. Reference construction runs once per task and is reused: 23.4 min, 1,300 predictor calls (200 original predictions, 1,100 perturbations), and 1.97M/2.16M input/output tokens including reasoning at maximum effort, which we will clarify in §4.1 and A.6.4 applies only to this stage.

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

This paper proposes SHIR, a framework for explaining time-series predictions with a limited number of perturbation queries. The main idea is to use perturbation results from reference inputs to decide which input units should be examined for a new prediction. SHIR uses an LLM to propose comparisons between input units based on existing evidence, then selects a small set of perturbation queries that completes as many comparisons as possible. It also introduces an evidence graph to reuse perturbation results across multiple comparisons. The final explanation is obtained by ranking input units using their measured effects when available and reference scores otherwise.

The authors evaluate SHIR on WeatherBench2 and VitalDB against seven existing explanation methods. SHIR achieves the best average results across four explanation metrics on both datasets. The ablation studies also show the benefits of the proposed components, including context construction, query selection, evidence reuse, and updating scores with current perturbation results.

Critical strengths
  • S1. SHIR addresses the practical challenge of explaining time-series predictions when perturbation queries are expensive.
  • S2. SHIR achieves the best reported performance across four explanation metrics on both WeatherBench2 and VitalDB, outperforming seven existing methods.
Critical weaknesses
  • W1. It is unclear why completing more pairwise comparisons leads to better explanations, since the final ranking is based on individual importance scores rather than comparison results.
  • W2. The paper does not show why LLM-generated comparisons are better than simpler ways of selecting which input units to perturb.
  • W3. Although SHIR achieves better results, the experiments do not clearly show which proposed components are responsible for the improvements.
  • W4. It is unclear why an evidence graph is necessary, since a simple table or cache might be sufficient to store and reuse perturbation results.
  • W5. The experiments cover only two datasets and 50 test inputs per dataset.
  • W6. The main results are based on a single run, without sufficient analysis of whether the improvements are statistically reliable.
  • W7. Although SHIR uses only three perturbation queries for each new input, it also requires many offline perturbations and LLM calls. The overall computational cost is not sufficiently evaluated.
  • W8. The experiments mainly focus on a fixed query budget of three, without showing how performance changes when more or fewer queries are available.
  • W9. It remains unclear whether the improvements come from the proposed comparison-based approach or simply from combining historical importance scores with a few new perturbation measurements.
Questions for authors
  • Q1. Why is maximizing the number of completed pairwise comparisons expected to improve explanation quality when the final ranking is based on individual importance scores? Could the authors provide a more direct justification for this objective?
  • Q2. How does LLM-based comparison generation perform against simpler query selection strategies?
  • Q3. What advantages does the evidence graph provide over a simpler table or cache of perturbation results? Could the authors clarify which operations specifically require the graph structure?
  • Q4. Could the authors provide confidence intervals or statistical significance tests?
  • Q5. What is the total computational cost of SHIR, including reference construction, LLM inference, and online perturbations?
  • Q6. How does SHIR perform under different perturbation query budgets?
1 comment

We sincerely thank the reviewer for the effort in reviewing our submission. Our detailed replies are as follows:

Response to AWYK-W1/Q1:

An explanation is a top-kk set, so what matters is how units near the cutoff are ordered. If uu is inside the reference top-kk and vv just outside, measuring only vv leaves membership hinging on uu's reference score, a population default whose error on this input is unknown. Measuring both gives s^x(u)−s^x(v)=ax(u)−ax(v)\widehat{s}_x(u)-\widehat{s}_x(v)=a_x(u)-a_x(v) under R0, so the decision rests on current evidence. Context construction prioritizes such boundary-crossing comparisons, and the completion count exploits shared endpoints (querying {u,v,w}\{u,v,w\} settles both u/vu/v and v/wv/w). The controls below confirm the benefit of this criterion over simpler selection rules.

Response to AWYK-W2/W9/Q2:

Controls on the 50 WeatherBench2 test inputs share the reference, three-query budget, perturbation, and R0. Global Top 3 and Global Random 3 select from the full catalog without an online LLM; LLM Direct 3 asks the same proposer for three units; Ref-score Top-3 selects within SHIR's pool.

StrategyComp ↑Suff err ↓CPD@10% ↑CPP err@10% ↓
Global Top 30.16280.03280.15860.0304
Global Random 30.15900.03300.15570.0280
LLM Direct 30.14920.03320.15410.0299
Ref-score Top-30.16600.03620.15740.0287
SHIR0.16340.03100.15960.0254

SHIR improves all four metrics over Global Top 3, Global Random 3, and LLM Direct 3; relative to Global Top 3, CPP error falls by 16.4% and Suff error by 5.5%. The gain thus depends on which effects are measured, not only on adding measurements. SHIR's advantage over LLM Direct 3, which uses the same proposer, supports the comparison-guided design.

Response to AWYK-W3:

Each ablation in Table B2 changes one stage. Relative to Flat presentation, which changes only evidence organization, SHIR reduces Suff error by 19.9%/9.2% (WeatherBench2/VitalDB). Relative to same-pool Ref-score Top-3, which changes only the selector, it reduces Suff error by 14.4%/4.7% and improves both 10% endpoints. Relative to Graph-only, which removes current measurements and R0, it reduces CPP error by 38.0%/27.8%.

Response to AWYK-W4/Q3:

A cache keyed by (x,u)(x,u) records whether uu was measured; SHIR keeps this as the execution table that determines MM. The evidence graph adds the comparison-level view used by proposal and planning: comparison patterns shared across inputs, their per-input instances, and execution records, with applicability edges linking e(x,u)e(x,u) to every instance containing uu (Eq. 4). One record thus supplies an endpoint to every applicable comparison yet counts once in aggregation, linking updates all affected comparisons for the next proposal round, and patterns connect a pair on a new input to matched reference evidence. These relations could equally be stored in relational tables. Disabling linking during reference construction raises VitalDB CPP error from 0.1402 to 0.1677.

Response to AWYK-W5/W6/Q4:

The two tasks differ in domain, predictor, input structure, and temporal scale. Weather test windows have nonoverlapping input-to-target intervals, and VitalDB windows come from 50 distinct held-out patients. SHIR is best on all four metrics on both tasks against seven baselines.

Per-input paired tests against the closest comparator, Ref-score Top-3 (Wilcoxon signed-rank, Holm-corrected over eight tests), show significantly lower retention error for SHIR at 7.5% and 15% (Holm pp = 0.010, 0.004) and larger deletion at 10% (pp = 0.033); Top-3 has larger deletion at 5% and 7.5% (pp = 0.033). The ablation variants average three repetitions, and SHIR remains best on Suff error and both 10% endpoints on both tasks. The revision will add repeated online runs, these tests for all comparisons, and a Limitations section.

Response to AWYK-W7/Q5:

StageTimePredictor callsLLM tokens (in / out)
Reference (once per task)23.4 min1,3001.97M / 2.16M
Online (per input)3.33 s316,088 / 353

Reference construction comprises 200 original predictions and 1,100 perturbations, and its output tokens include reasoning at maximum effort; online proposals use non-thinking mode, as we will clarify in §4.1 and A.6.4. LLM calls take 3.12 s (93.7%) of online time. As the reference is reused, NN inputs cost Cref+N⋅ConlineC_{\mathrm{ref}}+N\cdot C_{\mathrm{online}}, with three predictor queries per input versus 90 (WeatherBench2) or 20 (VitalDB) for Occlusion and 512 samples for KernelSHAP/TimeSHAP.

Reviewer_MLX1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents SHIR, an LLM-based approach for identifying the input units that influence a particular time-series prediction the most: Initially, the input and reference scores of units are given. SHIR asks an LLM to propose comparisons between the perturbation effects of units that look important for explaining the prediction, and then selects a budget-constrained subset of units, perturb those units to get their actual effects (measured perturbation effects). Finally, SHIR picks the top-k units based on the actual effect scores of those measured units and estimated reference scores of unmeasured units.

To improve the quality of reference scores, the authors propose an evidence graph mechanism to reuse results of existing perturbations of reference inputs.

Critical strengths

None.

Critical weaknesses
  1. The writing of this paper is difficult to follow in several places.
  • An earlier illustrative example would make the workflow easier to follow. The authors provide formal definitions of their approach, with only a short query-selection example in §\S 3.3. However, the more detailed example showing the workflow appears at the beginning of the Experiments section. Moving it earlier could help readers understand the approach's intuition and workflow better.
  • The advantage of comparison-based proposals over direct unit selection is unclear. The approach asks the LLM to propose comparisons rather than candidate units, but this design choice is not fully motivated. Do comparison-based proposals support settings with different budgets? Current experiments do not show this. And if there are different budgets, the selected units seem to be unstable: Given pairs {(a1,a2),(a1,a3),(a2,a3),(b1,b2),(b1,b3),…,(b3,b4)}∖{(b1,b2),(b3,b4)}\{(a_1, a_2), (a_1, a_3), (a_2, a_3), (b_1, b_2), (b_1, b_3), \dots, (b_3, b_4)\} \setminus \{(b_1, b_2), (b_3, b_4)\}: under budget 3, the selected subset will be {a1,a2,a3}\{a_1, a_2, a_3\} (because there is no 3 bib_i elements relate to each other). Under budget 4, the selected subset changes to {b1,b2,b3,b4}\{b_1, b_2, b_3, b_4\}.
  • The evidence graph lacks a definition, either a formal or an intuitive one. §\S 3.4 states "Let GG be the working evidence graph, and let e(x,u)e(x,u) denote a valid execution record for unit uu on input xx". This introduces the notation and describes the evidence graph, but does not define it. And the symbol GG is not used in subsequent main-text formulae. Does GG consist of records like e(x,u)e(x, u) only? Why is GG a graph? Its descriptions of stored records and their intended uses do not substitute for a definition of the graph structure.
  • The "many-to-many" nature of evidence reuse is not clearly illustrated. What does this mechanism add beyond storing each perturbation result and reusing it for future inputs?

These presentation issues limit my ability to assess the method’s technical novelty with confidence.

  1. The time-series-specific contribution is unclear. Although the paper focuses on time-series predictions, the proposed approach operates on predefined input units and does not appear to explicitly model temporal dependencies or interactions between units. Its central mechanisms, including comparison-based query selection and evidence reuse, seem applicable to general structured inputs rather than being specific to time series. While the authors incorporate temporal information through unit definitions and input descriptions, it remains unclear which parts of the proposed methodology fundamentally exploit time-series characteristics. The paper should better justify its emphasis on time-series problems or clarify the generality of the approach.
Questions for authors
  1. Why ask the LLM to propose comparisons and then perform the query selection, instead of simply asking the LLM to propose top-bb candidate units/a batch of bb candidate units to perturb?

  2. Could the authors provide a formal definition of the evidence graph, specifying its node and edge types, and how they are used?

1 comment

We sincerely thank the reviewer for the effort in reviewing our submission. Our detailed replies are as follows:

Response to MLX1-W1:

SHIR ranks predefined input units (e.g., a variable group in a time block) by how much perturbing them changes the forecast; the top-kk units form the explanation. Offline, it aggregates perturbation responses on 200 reference inputs into a reference score per unit. Online, it spends three perturbation queries on the new input, and each measured effect replaces that unit's reference score. In Figure 1, the reference top-2 is {u,v}\{u,v\}; querying {u,v,w}\{u,v,w\} settles the proposed comparisons u/vu/v and v/wv/w, since vv serves both; the measured effects (u=0.2u=0.2, v=0.9v=0.9, w=0.7w=0.7) change the top-2 to {v,w}\{v,w\}. We will move Figure 1 and a two-stage overview to the start of §3.

Response to MLX1-W1/Q1:

An explanation is a top-kk set, so what matters is how units near the cutoff are ordered. If uu is inside the reference top-kk and vv just outside, measuring only vv leaves membership hinging on uu's reference score, a population default whose error on this input is unknown. Measuring both gives s^x(u)−s^x(v)=ax(u)−ax(v)\widehat{s}_x(u)-\widehat{s}_x(v)=a_x(u)-a_x(v) under R0, so the decision rests on current evidence. A comparison thus names a decision and the two measurements that settle it, exposing shared measurements; a list of units encodes neither. The LLM nominates comparisons, the selector allocates the budget jointly, and proposed orderings never enter the scores.

LLM Direct 3 asks the same proposer for three units (50 WeatherBench2 inputs; same reference, budget, R0):

StrategyComp ↑Suff err ↓CPD@10% ↑CPP err@10% ↓
LLM Direct 30.14920.03320.15410.0299
SHIR0.16340.03100.15960.0254

SHIR outperforms direct selection by the same proposer on all four metrics, supporting the comparison-guided design. SHIR also beats Global Top 3 and Global Random 3 on all four metrics, and same-pool Ref-score Top-3 on all but Comp (0.1660).

Regarding different budgets, your example is correct, and non-nestedness follows from optimizing the whole batch: at b=4b=4, Eq. (3) settles four comparisons, while adding one unit to the triangle settles three. This changes which units are measured, not what a measurement means: each measured score is ax(u)a_x(u) whichever batch obtained it. The explanation size kk is separate from bb, and top-kk prefixes are nested within a run. At b=2,3,4b=2,3,4, SHIR beats Global Top-bb on all four metrics, e.g., CPP error 0.0233/0.0254/0.0267 vs. 0.0299/0.0304/0.0309.

Response to MLX1-W1/Q2:

For a fixed configuration, G=(P∪D∪E, Ainst∪Aapp)G=(\mathcal P\cup\mathcal D\cup\mathcal E,\ \mathcal A_{\mathrm{inst}}\cup\mathcal A_{\mathrm{app}}) (Figure A1):

  • Patterns p={u,v}∈Pp=\{u,v\}\in\mathcal P: shared across inputs, with proposal records.
  • Instances dx∈Dd_x\in\mathcal D: pp on input xx, with missing endpoints and measured ordering.
  • Executions e(x,u)∈Ee(x,u)\in\mathcal E: unique valid perturbation records storing ax(u)a_x(u).
  • Edges: (dx,p)∈Ainst(d_x,p)\in\mathcal A_{\mathrm{inst}} (dashed); (e(x,u),dx)∈Aapp(e(x,u),d_x)\in\mathcal A_{\mathrm{app}} iff u∈pu\in p (solid).

The method reads off GG: M={u:e(x,u)∈E}M=\{u:e(x,u)\in\mathcal E\} comes from the complete execution table, and a pair is pending (Eq. 2) exactly when p⊈Mp\nsubseteq M; Ex(d)\mathcal E_x(d) (Eq. 4) is the applicability neighborhood of dxd_x, which automatic linking materializes; Eq. (5) counts each execution once regardless of degree; episodes merge on patterns. We will add this to §3.4 and express Eqs. (2)–(5) via GG.

The relation is many-to-many: one record serves every comparison containing its unit on that input, and each comparison needs two records. This goes beyond store-and-reuse. Within an input, e(x,v)e(x,v) supplies the shared endpoint of {u,v}x\{u,v\}_x and {v,w}x\{v,w\}_x, so three executions on u,v,wu,v,w complete both, which Eq. (3) counts. Across inputs, comparison patterns, their proposal records, and reference scores and summaries guide proposals on a new input, whereas its endpoints come only from executions on that input. Relative to Flat presentation (same records, per unit), SHIR reduces Suff error by 19.9%/9.2% (WeatherBench2/VitalDB).

Response to MLX1-W2:

Temporal structure enters every step touching the predictor: units are contiguous time blocks × variable groups (× regions); perturbations follow temporal semantics (seasonal climatology; out-of-block medians that keep observation masks and recompute derived features); effects span three forecast horizons; the proposer reasons over per-unit summaries of the current window. SHIR measures the frozen forecaster's responses directly rather than assuming a temporal model, so it applies to any queryable forecaster. The principle also extends to other structured inputs, and time series is a natural fit given its rich unit semantics (time × variable × region). We will make this explicit in §1 and add a Limitations section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.