Explaining Time Series Forecasting with Horizon-Resolved Attribution
Seunghan Lee, Jun Seo, Jaehoon Lee, Junhyeok Kang, Sangjun Han, Sungdong Yoo, Minjae Kim, Tae Yoon Lim, et al.
cs.LG · cs.AI
Quick verdict
Proposes HRX, per-step explanations for TS forecasting; extensive 56K-run evidence supports the horizon axis, though authors concede no downstream utility yet.
Summary
The paper argues that time-series explanation methods wrongly assume every forecast step depends on the same past values, since nearly all return one importance vector shared across the horizon. It proposes HRX (Horizon-Resolved eXplanation), a plug-in framework returning an H×L explanation matrix, one row per forecast step, computed for any differentiable forecaster without backbone changes; the default estimator is the input gradient of a single step, obtained efficiently via batched backward calls. A deletion-based protocol scores each step separately, comparing a step's own row against the shared vector (g_own) and against another step's row (g_shuffled), with their difference (g_margin) as the decisive quantity; an effective-rank criterion (participation ratio of the singular values) gauges whether the axis is worth resolving. On synthetic generators with planted ground truth — including a rank-1 negative control — the matrix beats the shared vector on AUPRC/AUROC/AUP/AUR. Across roughly 56K runs spanning 26 backbones, 15 benchmarks and 8 estimators, g_margin is consistently positive (median +0.142, 80.6% of runs positive), and zero-shot foundation models show the same sign though smaller (+0.023 vs +0.143 trained). Limitations the authors concede: no series statistic predicts effect size, no downstream benefit was found, some backbones (e.g., Pyraformer, SegRNN, FEDformer) show no significant gain, and 15 of 66 zero-shot datasets are significantly negative.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Different forecast steps of a trained TS forecaster depend on different lookback positions, which the single shared importance vector of prior methods cannot express.
- HRX returns a per-step explanation matrix (H×L) for any differentiable forecaster without backbone changes or auxiliary training, and any explanation estimator can fill its rows.
- A deletion-based per-step protocol shows genuine step-matched gains: median g_margin +0.1418 with 80.6% positive over 1,126 scored runs (Table 4), and g_own/g_shuffled in the ideal direction.
- Step-wise dependence is low-dimensional: rank-1 truncation recovers none of the gain, a few directions recover most of it (Table 9), and the effective rank of E predicts the gain (Table 10).
- The gain holds across eight estimators and eight zero-shot foundation models, though it is markedly smaller zero-shot (+0.023 vs +0.143, Table 21).
Strengths
Controls that actually test the instrument
g_shuffled (Eq. 4) scores each step with another step's row to rule out the 'concentrated maps break more' artifact, and the sparsenull rank-1 negative control (Appendix G, Table 14) shows the protocol manufactures no gain where none exists — rare rigor in xAI evaluation.
Estimator-agnostic evidence at scale
About 56K runs over 26 backbones and 15 benchmarks; the horizon axis is ported to seven other estimators by changing only the preserved scalar to the single step (E.3), and all show positive gains (Table 6, e.g. Dynamask g_own +0.275), isolating the axis as the cause.
Statistical hygiene
Honest negative results
Ground-truth anchoring and rank calibration
Weaknesses
No demonstrated practical utility
Advance-prediction claim overstated
Understated heterogeneity
Tail instability of the normalized quantity
Faithfulness-to-model only on real data
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
HRX (LG AI Research; arXiv 2609.12639) is a method paper about explaining time-series forecasters, not improving them. The central claim is that the single importance vector returned by prior TS explanation methods (Table 1: TimeX++, Dynamask, ExtremalMask, ContraLSP, gradient/SHAP family) hides per-step structure: different forecast steps read different lookback positions. The proposed fix is an H×L explanation matrix E, one row per step, computable for any differentiable forecaster without touching the backbone; the authors are explicit that the axis, not the estimator, is the contribution. Three supporting claims: a deletion-based per-step evaluation protocol (g_own, g_shuffled, g_margin); a rank criterion (participation ratio of E's singular values) that flags when the axis is worth resolving; and evidence that step-wise dependence is low-dimensional — a few shared maps suffice and their number does not grow with H.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is HRX (Horizon-Resolved eXplanation)?
HRX is a plug-in framework that gives each forecast step of a time-series model its own attribution map: instead of one importance vector for the whole forecast horizon, it returns an H-by-L matrix (one row per step), read out via gradients from any differentiable forecaster without retraining the model.
Does HRX make forecasts more accurate?
No. The authors tested explanation-guided input selection and horizon-specific pruning and found no measurable benefit, so HRX currently serves as a diagnostic and measurement tool rather than a way to improve forecasting performance.
Which models and datasets does the paper evaluate on?
26 forecasting backbones (linear, CNN, recurrent and Transformer families from the Time-Series-Library) on 15 benchmarks (ETT, Electricity, Traffic, Weather, PEMS, Solar, Exchange plus a new 100-company US stock dataset in log-price and log-return forms), and 8 zero-shot foundation models (Chronos, TimesFM, Moirai, Timer, TimeMoE, TTM, VisionTS) on GIFT-Eval — roughly 56K runs in total.
Is the code available?
Yes. The paper links github.com/seunghan96/HRX and states that the full source code — estimators, evaluation protocol, data preparation, and the scripts producing every figure and table — is submitted as supplementary material.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with '[...sampling gap...]' markers: mid-paper material (the end of Section 4.1, part of the references, and part of Appendix Q's Table 36) was skipped; judgments rest on the visible head, tail, and appendix tables.
- The provided text ends at Table 41 (Weather, p.43); any content after that page is not visible in the excerpt.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.