- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper extends the scalar operator-based feature explanation framework of Fumagalli et al. (2025) to Hilbert-valued (i.e., time-dependent or functional, outputs). The central new component is an output behavior operator , a temporal aggregation and smoothing mechanism parameterized by a kernel , which yields explanations at three levels of temporal granularity: time-specific, time-resolved, and time-aggregated. The paper also gives guidance on how to select kernels. The work addresses a real shortcoming of current practice: applying scalar explanations pointwise at each output location ignores dependencies across the output domain and provides no principled mechanism for aggregating feature effects over time. The framework is relevant to applications such as time-series forecasting and dose-response analysis.
Critical strengths
- The synthetic recovery study (App. D.1) and the temporal localization experiment with its anti-causal and placebo controls (App. D.5) are carefully designed, and reproducibility is strong.
- The kernel-to-problem mapping in Table 5, along with the recommendation to assign kernels per feature, has practical value.
- The paper is clearly written and well organized.
- Theorem 2 links Shapley-style XAI with time-dependent Sobol indices.
Critical weaknesses
- Limited novelty. Relative to Fumagalli et al. (2025), the main addition is a new operator, , which is a kernel integral operator (Eq. 6) and mathematically a linear smoother over the output domain. The premise of Theorem 1, that for all s T, is a strong pointwise-dominance condition that is rarely verifiable in practice.
- The real-data experiments (SPY, IHEPC, NESO) are largely qualitative case studies based on a single instance or a median profile. The regime-shift narrative in the SPY analysis rests on (a) an interaction effect, precisely the component that the authors' own synthetic study finds hardest to recover; (b) a single trajectory with no confidence intervals; and (c) results that are sensitive to user-specified kernel parameters. I suggest measuring the divergence between kernel-before-interaction and post-hoc smoothing for the sensitivity/risk games on real data, reporting the magnitude of ranking changes and trajectory differences.
- The claim of "matching traders' interpretation of market events" (Sec. 5.2) is not supported by evidence from traders, and the statement that hourly reserve adjustments "mis-target the structure of forecast variance" (Sec. 5.2, Fig. 4) is not supported by any downstream reserve simulation.
- The Figure 9 caption says "two synthetic examples" but lists three (ICU, price pulse, 3-day medication) and contains a duplicated "and and"; the 1.47×/1.61× partial/pure ratios in Figure 3, col. 3 should be disambiguated.
- The special-case equivalences would be more convincing if verified empirically, for instance by numerically comparing the identity-kernel implementation against the published TsSHAP/ShapTime outputs on a shared dataset.
Questions for authors
Please refer to the 'critical weakness'.
2 comments
Additional Response to Reviewer kdTZ
W2: Our real-world experiments serve complementary purposes: the local SPY example deliberately explains a particular model prediction to illustrate how different kernels reveal its temporal feature effects, while other examples investigate global explanations. These case studies are complemented by synthetic recovery experiments and quantitative temporal-localization analysis (App. D.5). The SPY example is therefore not intended to establish a population-level pattern.
(a) Interaction reliability. Feature-based explanations characterize the fitted model rather than the underlying data-generating process. Different model classes may learn different interaction structures, so variability in interaction recovery across models does not necessarily indicate unreliable attribution for a particular fitted model. Moreover, the SPY interaction illustrates how model-specific interactions can be represented across time and is not the main finding of the example.
(b) Confidence intervals. For a fixed model and instance, the exact explanation is deterministic. Confidence intervals across observations would address population-level variability rather than uncertainty in the explanation of this particular instance. Numerical approximation error is a separate concern, investigated in our synthetic recovery experiments, and can be reduced by increasing the number of masking samples.
(c) Kernel sensitivity. Dependence on kernel choice is intentional: different kernels define different notions of temporal contribution and therefore answer different explanatory questions (Sec. 5.1, Sec. 6, Tab. 5). Our framework provides guidance for selecting kernels accordingly. Theorem 1 establishes sufficient conditions for preserving feature rankings across kernels, while App. D.1 (Table 6) illustrates when rankings remain stable or change.
Additional divergence measurements. With the extension of Theorem 1, we have clarified that kernel aggregation and linear interaction operators commute for prediction-, risk-, and sensitivity-based explanations. Consequently, applying the same fixed kernel before or after the interaction operator yields mathematically identical results. Additional divergence measurements would therefore only reflect numerical approximation error rather than a methodological difference. We have clarified this in Appendix D.6.
W3: The first statement was meant to refer to the established interpretation of scheduled macro announcements in the market microstructure literature, not to evidence collected from traders. We have rephrased it accordingly and added the corresponding references. The second statement describes a hypothetical use case motivating the sensitivity analysis; a downstream reserve simulation is outside the scope of this paper. We have reworded it to make its hypothetical nature explicit.
W4: Thank you, we have corrected the Figure 9 caption (now Figure 8 in revised manuscript) and clarified in the Figure 3 caption and corresponding main text that the two ratios refer to the identity kernel (1.47×) and the causal kernel (1.61×), respectively.
W5: The special-case relationships in Sec. 4.3 are mathematical equivalences of attribution definitions under matching coalition constructions, rather than claims that independently implemented pipelines produce identical numerical outputs. We have therefore added explicit derivations for TsSHAP and ShapTime in Appendix C, showing how their attribution formulas follow from the corresponding operator choices. These derivations establish the correspondence independently of any particular model or dataset. Comparing separate implementations would primarily introduce differences in surrogate modeling, coalition evaluation, and numerical approximation, rather than test the claimed mathematical relationship. Our identity-kernel experiments already illustrate the corresponding pointwise and aggregated explanation settings (e.g., TsSHAP in Fig. 3, row 1, col. 3).
Response to Reviewer kdTZ
We thank the reviewer for their detailed review. We appreciate that they find the paper clearly written, the synthetic and localization experiments carefully designed, and the kernel selection guide of practical value. We have uploaded a revised manuscript in which all changes are highlighted in blue, and we refer to the corresponding sections in our responses to any open points below.
W1: We thank the reviewer for these observations. Regarding novelty, we agree that the proposed output behavior operator is linear. However, the contribution lies not in this operator alone, but in the generality of the resulting construction. By introducing a Hilbert-valued functional decomposition and extending the unifying framework of Fumagalli et al. (2025), we establish a systematic framework for feature-based explanations of Hilbert-valued outputs, including time-dependent predictions.
Importantly, this connects two previously separate research directions: time-dependent XAI, which primarily explains individual time points or temporal summaries without explicitly accounting for cross-time dependencies, and functional sensitivity analysis, which incorporates such dependencies but focuses largely on global variance-based importance. Our framework extends dependency-aware explanations to local prediction, global risk, and global sensitivity, covering pure, partial, and full operators; individual, interaction, and joint effects; and different temporal granularities. This enables, for instance, cross-time-dependency-aware global risk attribution.
Rather than merely smoothing explanations, kernels emphasize different relationships between output locations, providing distinct, interpretable perspectives on feature influence. We provide guidance on when and how to select kernels, combining this flexibility with a broad class of previously unavailable dependency-aware explanations through a single, coherent mathematical construction.
We have revised the introduction and related work (Sec. 1 and 2) to emphasize these contributions and added Table 3 in Appendix B highlighting the explanation settings enabled beyond existing pointwise XAI and functional sensitivity approaches.
Regarding Theorem 1, we agree that pointwise dominance is a strong sufficient condition that may not hold in practice. However, it is not a requirement for the reliability or applicability of our explanations. Different kernels emphasize different temporal relationships and therefore answer different explanatory questions; differing rankings do not inherently indicate inconsistency or reduced reliability. The theorem establishes a sufficient condition for ranking preservation under temporal aggregation, which can be checked directly from computed time-resolved effects. Moreover, App. D.1 (Table 6) shows that rankings may remain stable even when the condition fails, as long as features are well separated in aggregate magnitude. Neither pointwise dominance nor ranking preservation is required to apply the framework, as clarified in Sec. 4.2 and App. D.1.