pith. sign in

arxiv: 2606.18049 · v1 · pith:PLZG7E7Anew · submitted 2026-06-16 · 💻 cs.LG

ConTex: Reformulating Counterfactual Generation For Time Series Forecasting

Jan Voets , Hasan Tercan , Tobias Meisen , Sebastian Baum This is my paper

Pith reviewed 2026-06-27 01:27 UTC · model grok-4.3

classification 💻 cs.LG
keywords counterfactual explanationstime series forecastingintervention strategymodel-agnostic architecturereal-time inferencesparse counterfactualsdeep learning explanations
0
0 comments X

The pith

ConTex reformulates counterfactual generation for time series forecasting as learning one shared intervention function instead of per-instance optimization.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that counterfactual generation for time series forecasting works better when treated as learning one global intervention strategy instead of optimizing changes separately for every data point. A sympathetic reader would care because this makes it feasible to get actionable advice from forecasts in real time, without the high cost and inconsistency of older methods. The new approach uses a model-agnostic setup with encoders that process temporal context and conditions, then heads that decide which times and features to change and by how much. This produces sparse, valid counterfactuals in a single pass across different forecasting models and datasets.

Core claim

We reformulate counterfactual generation for time series forecasting as the problem of learning a globally consistent intervention strategy, allowing counterfactuals to be generated through a single shared function. We propose ConTex, a model-agnostic, decomposed architecture comprising a temporal context encoder and a conditional encoder, followed by two heads that capture interventions in terms of temporal relevance and modification strength. This structure overcomes the instability and inconsistency of instance-based approaches by producing targeted, interpretable interventions across time and feature dimensions in a single forward pass.

What carries the argument

The central mechanism is the globally consistent intervention strategy implemented as a single shared function in the ConTex architecture, which decomposes into temporal context and conditional encoders plus intervention heads for temporal relevance and modification strength.

If this is right

  • Counterfactuals achieve state-of-the-art validity on multiple datasets and forecasting architectures.
  • Generated counterfactuals are sparse, minimizing the number of necessary interventions.
  • Computational cost is reduced by at least 12-36x compared to instance-wise generation.
  • Real-time inference is supported at approximately 0.007 seconds per example.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The shared function's consistency across instances could make repeated explanations more reliable in ongoing monitoring applications.
  • This approach might allow the intervention function to be trained jointly with the base forecaster for tighter alignment.
  • The single-pass design opens use in streaming settings where forecasts update continuously and explanations must keep pace.

Load-bearing premise

A single shared function learned across instances can produce valid, minimal counterfactual interventions that match or exceed the quality of per-instance optimization without introducing systematic bias or missing instance-specific nuances.

What would settle it

If testing on new time series data shows that the shared function's counterfactuals change the prediction less often or require more modifications than individually optimized ones, the reformulation would not hold.

Figures

Figures reproduced from arXiv: 2606.18049 by Hasan Tercan, Jan Voets, Sebastian Baum, Tobias Meisen.

Figure 1
Figure 1. Figure 1: ConTex architecture. A target-conditioned model using feature-wise linear modulation [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of generated counterfactuals on M4 and NN5 using N-HiTS. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Interventional (target-conditioned) attribution: Temporal relevance masks for a smooth [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Metric behavior across increasing target difficulty on NN5 with N-HiTS. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative examples on realistic target trajectories. ConTex remains capable of reproducing [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
read the original abstract

Decision-making with deep learning-based time series forecasting requires not only accurate predictions but also actionable insights. However, current architectures do not inherently provide such information. Specifically, guidance is needed on how current conditions must be modified to shift from a predicted outcome to a desired future scenario. Counterfactual explanations provide a natural framework for this task, as they represent minimal input changes that alter the model's prediction, indicating when and how intervention is required. Existing approaches rely on instance-wise optimization, leading to inconsistency across instances, high computational costs, and limited applicability in real-time settings. To address these limitations, we reformulate counterfactual generation for time series forecasting as the problem of learning a globally consistent intervention strategy, allowing counterfactuals to be generated through a single shared function. We propose Counterfactual Time Series Explanations (ConTex), a model-agnostic, decomposed architecture comprising a temporal context encoder and a conditional encoder, followed by two heads that capture interventions in terms of temporal relevance and modification strength. This structure overcomes the instability and inconsistency of instance-based approaches by producing targeted, interpretable interventions across time and feature dimensions in a single forward pass, making it suitable for real-time applications. Across multiple forecasting architectures and benchmark datasets, ConTex achieves state-of-the-art validity while generating sparse counterfactuals that minimize the number of necessary interventions. Additionally, our approach reduces computational cost by at least 12-36x compared to instance-wise generation and supports real-time inference at approximately 0.007 seconds.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims to reformulate counterfactual generation for time series forecasting as learning a single globally consistent intervention strategy, implemented via the ConTex architecture (temporal context encoder + conditional encoder + two heads for temporal relevance and modification strength). This model-agnostic approach is asserted to deliver SOTA validity and sparsity while achieving 12-36x speedup and real-time inference (~0.007s) over instance-wise optimization methods across forecasting architectures and benchmarks.

Significance. If the empirical claims hold, the work would meaningfully advance practical interpretability for deep time-series forecasters by replacing slow, inconsistent per-instance optimization with a fast, consistent shared function. The decomposed encoder-head design and emphasis on global consistency are constructive contributions that could influence real-time decision-support systems.

major comments (2)
  1. [Abstract] Abstract: the central claim that a single shared function produces valid, minimal interventions matching or exceeding instance-wise optimization rests on the unverified assumption that the conditional encoder fully captures instance-specific dynamics (e.g., varying seasonality or trend shifts); without explicit loss terms or constraints enforcing per-instance minimality, averaged outputs risk violating validity or increasing intervention count for some instances.
  2. [Abstract] Abstract: the reported 12-36x speedup and SOTA validity are load-bearing for the reformulation's practical value, yet the abstract supplies no quantitative tables, ablation results, or error analysis comparing ConTex to per-instance baselines; this absence prevents verification that the shared architecture does not introduce systematic bias.
minor comments (1)
  1. [Abstract] The abstract would benefit from a brief statement of the training objective or regularization used to promote sparsity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful comments on our work. We address each major comment below, providing clarifications based on the manuscript content and indicating where revisions may be appropriate.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that a single shared function produces valid, minimal interventions matching or exceeding instance-wise optimization rests on the unverified assumption that the conditional encoder fully captures instance-specific dynamics (e.g., varying seasonality or trend shifts); without explicit loss terms or constraints enforcing per-instance minimality, averaged outputs risk violating validity or increasing intervention count for some instances.

    Authors: The conditional encoder processes instance-specific inputs (current time series values and desired target) alongside the temporal context encoder, enabling capture of varying dynamics such as seasonality without relying on averaging. Section 3.2 details the training objective, which includes an explicit validity loss (ensuring prediction shift for each instance) and a sparsity regularization term that penalizes intervention count per sample. These are not post-hoc averages but optimized end-to-end for the shared function. We disagree that this is unverified, as empirical results in Section 4 confirm per-instance validity and sparsity matching or exceeding baselines; however, we will add a clarifying sentence on the per-instance nature of the losses in the revised introduction. revision: partial

  2. Referee: [Abstract] Abstract: the reported 12-36x speedup and SOTA validity are load-bearing for the reformulation's practical value, yet the abstract supplies no quantitative tables, ablation results, or error analysis comparing ConTex to per-instance baselines; this absence prevents verification that the shared architecture does not introduce systematic bias.

    Authors: Abstracts conventionally provide high-level summaries without tables or detailed ablations due to space constraints. The full manuscript includes comprehensive quantitative comparisons, tables, and ablations in Sections 4.1-4.3 and 5, covering validity, sparsity, speedup (12-36x), and error analysis across multiple forecasting models and datasets, with no evidence of systematic bias from the shared architecture. We will consider a minor expansion of key result highlights in the abstract if permitted by the venue format. revision: no

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper's central contribution is a modeling reformulation and a new decomposed architecture (temporal context encoder + conditional encoder + two heads) for generating counterfactuals via a single shared function. This is presented as an independent design choice rather than a derivation that reduces to fitted parameters or self-referential definitions. No equations are supplied that equate a 'prediction' to an input by construction, and no load-bearing self-citations or uniqueness theorems are invoked. Empirical claims of SOTA validity and speedup rest on external benchmarks, not internal redefinitions, making the work self-contained against the listed circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that a globally consistent intervention strategy exists and can be captured by the proposed decomposed architecture. No free parameters or invented entities are named in the abstract.

axioms (1)
  • domain assumption A globally consistent intervention strategy can be learned from data to produce valid counterfactuals across instances without per-instance optimization.
    This is the explicit reformulation stated in the abstract as the solution to inconsistency and computational cost.

pith-pipeline@v0.9.1-grok · 5801 in / 1227 out tokens · 36737 ms · 2026-06-27T01:27:18.056655+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 13 canonical work pages

  1. [1]

    doi: https://doi.org/10.1016/j.ijforecast.2022.06.001

    ISSN 0169-2070. doi: https://doi.org/10.1016/j.ijforecast.2022.06.001. URL https: //www.sciencedirect.com/science/article/pii/S0169207022000929. Robert R. Andrawis, Amir F. Atiya, and Hisham El-Shishiny. Forecast combinations of computa- tional intelligence and linear models for the nn5 time series forecasting competition.International Journal of Forecast...

    work page doi:10.1016/j.ijforecast.2022.06.001 2070

  2. [2]

    doi: https://doi.org/10.1016/ j.ijforecast.2010.09.005

    ISSN 0169-2070. doi: https://doi.org/10.1016/ j.ijforecast.2010.09.005. URL https://www.sciencedirect.com/science/article/pii/ S0169207010001445. Special Section 1: Forecasting with Artificial Neural Networks and Com- putational Intelligence Special Section 2: Tourism Forecasting. Emre Ate¸ s, Burak Aksar, Vitus Leung, and Ayse Coskun. Counterfactual expl...

    2070

  3. [3]

    Omar Bahri, Soukaina Filali Boubrahimi, and Shah Muhammad Hamdi

    doi: 10.1109/ICAPAI49758.2021.9462056. Omar Bahri, Soukaina Filali Boubrahimi, and Shah Muhammad Hamdi. Shapelet-based counter- factual explanations for multivariate time series,

    work page doi:10.1109/icapai49758.2021.9462056 2021

  4. [4]

    João Bento, Pedro Saleiro, André F

    URL https://arxiv.org/abs/1803.01271. João Bento, Pedro Saleiro, André F. Cruz, Mário A.T. Figueiredo, and Pedro Bizarro. Timeshap: Explaining recurrent models through sequence perturbations. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD ’21, page 2565–2573. ACM, August

    Pith/arXiv arXiv

  5. [5]

    URL http://dx.doi.org/10.1145/3447548

    doi: 10.1145/3447548.3467166. URL http://dx.doi.org/10.1145/3447548. 3467166. Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza, Max Mergenthaler-Canseco, and Artur Dubrawski. N-hits: Neural hierarchical interpolation for time series forecasting,

    work page doi:10.1145/3447548.3467166

  6. [6]

    Shuang Dai, Fanlin Meng, Hongsheng Dai, Qian Wang, and Xizhong Chen

    URLhttps://arxiv.org/abs/2201.12886. Shuang Dai, Fanlin Meng, Hongsheng Dai, Qian Wang, and Xizhong Chen. Electrical peak demand forecasting- a review,

    arXiv

  7. [7]

    Abhimanyu Das, Weihao Kong, Andrew Leach, Shaan Mathur, Rajat Sen, and Rose Yu

    URLhttps://arxiv.org/abs/2108.01393. Abhimanyu Das, Weihao Kong, Andrew Leach, Shaan Mathur, Rajat Sen, and Rose Yu. Long-term forecasting with tide: Time-series dense encoder,

    arXiv

  8. [8]

    URLhttps://arxiv.org/abs/2009.13211. A. Ferchichi, A.B. Abbes, V . Barra, and I.R. Farah. Trustworthy AI for Spatio- Temporal Forecasting via Counterfactual Causality. pages 10–17,

    arXiv 2009

  9. [9]

    URL https://www.scopus.com/inward/record.uri? eid=2-s2.0-105031887993&doi=10.1109%2FICTAI66417.2025.00010&partnerID=40& md5=d6de56c36f3abd80813868f0f54b221b

    doi: 10.1109/ICTAI66417.2025.00010. URL https://www.scopus.com/inward/record.uri? eid=2-s2.0-105031887993&doi=10.1109%2FICTAI66417.2025.00010&partnerID=40& md5=d6de56c36f3abd80813868f0f54b221b. Rakshitha Godahewa, Christoph Bergmeir, Geoff Webb, Rob Hyndman, and Pablo Montero-Manso. Electricity Hourly Dataset, June 2020a. URLhttps://zenodo.org/record/3889...

    work page doi:10.1109/ictai66417.2025.00010 2025

  10. [10]

    Sepp Hochreiter and Jürgen Schmidhuber

    URL https://arxiv.org/abs/ 2602.13087. Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory.Neural Comput., 9(8): 1735–1780, November

    arXiv

  11. [11]

    Long short-term memory.Neural Computation, 9(8): 1735–1780, 1997

    ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL https://doi.org/10.1162/neco.1997.9.8.1735. 17 Jongseon Kim, Hyungjoon Kim, HyunGi Kim, Dongjun Lee, and Sungroh Yoon. A compre- hensive survey of deep learning for time series forecasting: architectural diversity and open challenges.Artificial Intelligence Review, 58(7):216, April

    work page doi:10.1162/neco.1997.9.8.1735 1997

  12. [12]

    doi: 10.1007/ s10462-025-11223-9

    ISSN 1573-7462. doi: 10.1007/ s10462-025-11223-9. URL https://link.springer.com/10.1007/s10462-025-11223-9 . Ryoung-Eun Ko, Zero Kim, Bomi Jeon, Migyeong Ji, Chi Ryang Chung, Gee Young Suh, Myung Jin Chung, and Baek Hwan Cho. Deep Learning-Based Early Warning Score for Predicting Clinical Deterioration in General Ward Cancer Patients.Cancers, 15(21):5145, October

    work page doi:10.1007/s10462-025-11223-9

  13. [13]

    doi: 10.3390/cancers15215145

    ISSN 2072-6694. doi: 10.3390/cancers15215145. URL https://www.mdpi.com/2072-6694/15/ 21/5145. Bryan Lim, Sercan O. Arik, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for interpretable multi-horizon time series forecasting,

    work page doi:10.3390/cancers15215145 2072

  14. [14]

    URL https://arxiv.org/abs/1912. 09363. X. Luo and W. Yin. Counterfactual Explanation-Based Cryptocurrency Price Prediction.En- tropy, 28(1),

    1912

  15. [15]

    URL https://www.scopus.com/inward/ record.uri?eid=2-s2.0-105028503537&doi=10.3390%2Fe28010065&partnerID=40& md5=76f11583c4b6477488a22a29f85fa8c4

    doi: 10.3390/e28010065. URL https://www.scopus.com/inward/ record.uri?eid=2-s2.0-105028503537&doi=10.3390%2Fe28010065&partnerID=40& md5=76f11583c4b6477488a22a29f85fa8c4. Spyros Makridakis, Evangelos Spiliotis, and Vassilios Assimakopoulos. The m4 competition: 100,000 time series and 61 forecasting methods.International Journal of Forecasting, 36(1):54–74,

    work page doi:10.3390/e28010065

  16. [16]

    International Journal of Forecasting 36, 54–74

    ISSN 0169-2070. doi: https://doi.org/10.1016/j.ijforecast.2019.04.014. URL https://www. sciencedirect.com/science/article/pii/S0169207019301128. M4 Competition. Amin Nayebi, Sindhu Tipirneni, Chandan K Reddy, Brandon Foreman, and Vignesh Subbian. Windowshap: An efficient framework for explaining time-series classifiers based on shapley values,

    work page doi:10.1016/j.ijforecast.2019.04.014 2070

  17. [17]

    Thach Le Nguyen and Georgiana Ifrim

    URLhttps://arxiv.org/abs/2211.06507. Thach Le Nguyen and Georgiana Ifrim. Tshap: Fast and exact shap for explaining time series classification and regression. In Rita P. Ribeiro, Bernhard Pfahringer, Nathalie Japkowicz, Pedro Larrañaga, Alípio M. Jorge, Carlos Soares, Pedro H. Abreu, and João Gama, editors,Machine Learning and Knowledge Discovery in Datab...

    arXiv

  18. [18]

    David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski

    URLhttps://arxiv.org/abs/1709.07871. David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. Deepar: Probabilistic forecasting with autoregressive recurrent networks.International Journal of Forecasting, 36(3): 1181–1191,

    Pith/arXiv arXiv

  19. [19]

    International Journal of Forecasting 36, 1181–1191

    ISSN 0169-2070. doi: https://doi.org/10.1016/j.ijforecast.2019.07.001. URL https://www.sciencedirect.com/science/article/pii/S0169207019301888. Udo Schlegel and Thomas Seidl. What-if explanations over time: Counterfactuals for time series classification,

    work page doi:10.1016/j.ijforecast.2019.07.001 2070

  20. [20]

    Andreas Theissler, Francesco Spinnato, Udo Schlegel, and Riccardo Guidotti

    URLhttps://arxiv.org/abs/2603.27792. Andreas Theissler, Francesco Spinnato, Udo Schlegel, and Riccardo Guidotti. Explainable ai for time series classification: a review, taxonomy and research directions.Ieee Access, 10:100700–100724,

    arXiv

  21. [21]

    URL https: //www.scopus.com/inward/record.uri?eid=2-s2.0-85185401353&doi=10.1109% 2FICDM58522.2023.00180&partnerID=40&md5=206d3e362e2089be0f73b3aeca43966c

    doi: 10.1109/ICDM58522.2023.00180. URL https: //www.scopus.com/inward/record.uri?eid=2-s2.0-85185401353&doi=10.1109% 2FICDM58522.2023.00180&partnerID=40&md5=206d3e362e2089be0f73b3aeca43966c. 18 Z. Wang, I. Samsten, I. Miliou, and P. Papapetrou. COMET: Constrained Counterfactual Explanations for Patient Glucose Multivariate Forecasting. pages 502–507,

    work page doi:10.1109/icdm58522.2023.00180 2023

  22. [22]

    URL https://www.scopus.com/inward/record.uri? eid=2-s2.0-85200437241&doi=10.1109%2FCBMS61543.2024.00089&partnerID=40& md5=a224e4dfcecfaf00ad2130609bd91307

    doi: 10.1109/CBMS61543.2024.00089. URL https://www.scopus.com/inward/record.uri? eid=2-s2.0-85200437241&doi=10.1109%2FCBMS61543.2024.00089&partnerID=40& md5=a224e4dfcecfaf00ad2130609bd91307. Jingquan Yan and Hao Wang. Self-interpretable time series prediction with counterfactual explana- tions,

    work page doi:10.1109/cbms61543.2024.00089 2024

  23. [23]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu

    URLhttps://arxiv.org/abs/2306.06024. Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting?,

    arXiv

  24. [24]

    URLhttps://arxiv.org/abs/2205.13504. G. Zuin and A. Veloso. Navigating Time’s Possibilities: Plausible Counterfactual Explanations for Multivariate Time-Series Forecast through Genetic Algorithms. Number 2024, pages 2575–2582,

    arXiv 2024

  25. [25]

    URL https://www.scopus

    doi: 10.1109/TrustCom63139.2024.00359. URL https://www.scopus. com/inward/record.uri?eid=2-s2.0-105006506164&doi=10.1109%2FTrustCom63139. 2024.00359&partnerID=40&md5=6a1191858aa7346ef48d19159a09185f. 19

    work page doi:10.1109/trustcom63139.2024.00359 2024