acceptodds
Under review as a conference paper at ICLR 2027

When Time Series Foundation Models Fail Under Cross-Channel Dependence

Abstract

Multivariate time series foundation models (TSFMs) can access related series and covariates, but whether their forecasts reliably reflect the predictive value of cross-channel information remains unclear. We examine how Chronos-2, TimesFM-3, and TiRex-2 use cross-channel information through controlled synthetic experiments and real-world forecasting tasks. Our analysis covers cross-channel temporal relationships, dependence structures among observed channels, and the effects of known-future conditioning and covariate reliability. Useful cross-channel relationships are not consistently reflected in forecasts, while redundant or conditionally irrelevant inputs can still influence predictions. Statistical baselines fitted to the same observed contexts recover the relationships that the models underuse, indicating that these shortfalls reflect limited use of available information rather than its absence. On real-world tasks, conditioning on each model’s own median covariate forecasts increases quantile loss and reduces interval coverage. Propagating covariate uncertainty recovers much of this loss, but not all. Together, these findings reveal model-dependent limitations in exploiting temporal dependencies, combining information across channels, and producing reliable forecasts under covariate uncertainty.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Foundation Models, Multivariate Time Series Forecasting

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

This paper presents a diagnostic study of how time series foundation models (TSFMs) use cross-channel information, evaluating frozen Chronos-2, TimesFM-3, and TiRex-2 under controlled synthetic processes with known data-generating mechanisms and on real-world forecasting tasks from fev-bench. The analysis is organized around four research questions addressing temporal relationships (event-response patterns with varying delays), dependence structures (a shared latent factor observed through noisy, duplicate, and distractor channels), conditioning interfaces (multivariate, past-covariate, and known-future), and covariate reliability (degraded and predicted covariates). The core contribution is an alignment-based diagnostic that distinguishes underuse of informative inputs from sensitivity to inputs that add no predictive information. The central finding is that cross-channel information can be available without being effectively used: the models underuse delayed or complementary signals while remaining sensitive to redundant or irrelevant inputs, and statistical baselines fitted to the same observed contexts recover the relationships the models miss, indicating that the shortfalls reflect limited use of information rather than its absence. On real-world tasks, conditioning on each model's own predicted covariate medians increases quantile loss and reduces interval coverage, and propagating covariate uncertainty recovers much but not all of this loss.

Critical strengths
  1. A clean, well-controlled diagnostic methodology. By using synthetic data with known generating processes, together with oracle references and statistical baselines, the paper can precisely separate "no information available" from "information available but unused" — a distinction aggregate benchmark accuracy cannot make.

  2. Findings that are concrete, actionable, and real-world validated. The result that conditioning on a model's own predicted covariate medians degrades probabilistic forecasts is directly relevant to common forecasting practice, and it is confirmed on three real-world task pairs.

Critical weaknesses
  1. A possible home-field advantage in one key control. The fitted state-space model used to show the information is "recoverable" is close to the factor-model class that generated the synthetic data, so its strong performance may partly reflect a lucky model-class match rather than a general deficiency of the TSFMs.
Questions for authors
  1. Since the three models are very recent and some are not peer-reviewed, could you add a short background section describing their architectures and training, so that readers less familiar with these models can interpret the results?

  2. The findings are described as model-dependent and cover only three models. Do you have evidence that the same qualitative patterns, such as underuse of delayed signals and sensitivity to irrelevant channels, extend to other TSFMs?

4 comments
AuthorsOpenReview ↗

Response to Reviewer NdnY (1/2)

We sincerely thank the reviewer for the thoughtful feedback. Below we address the concerns on the statistical reference and the evidence beyond the original three models.

[W1 : Model-class assumptions and statistical references]

We agree that inferring cross-channel structure from a finite, noisy context is challenging, and that the fitted SSM benefits from assumptions close to our shared-latent generator. Its performance should therefore be interpreted as an achievable reference under a specified model class. We will revise the wording accordingly.

To address this concern, we added SparseDirect, a direct sparse linear regression that selects its input channels, lag length, and regularization using rolling validation within the observed context. It selects between target-only and all-channel inputs, retains raw duplicates, and receives no latent state, generating parameters, informative-channel labels, or true lag. It therefore provides an additional statistical reference without the SSM’s shared-state and common-persistence assumptions, while retaining its own linear sparsity assumptions.

As summarized in Tables 1–3, the comparison changes our interpretation. At K = 16, every tested TSFM has lower absolute MAE and smaller distractor deterioration than SparseDirect, so the distractor reaction is not excessive by this standard. TimesFM-3 (+0.01 ± 0.16%) shows that it can also be avoided zero-shot. Two observations remain.

  • Independent channels. SparseDirect's gain is itself small and uncertain (+1.07 ± 0.93%), so we do not claim underuse relative to it. However, TimesFM-3 reaches +2.60% zero-shot, close to the SSM (+2.84%), which shows that this gain is attainable without knowing the model class. The other four TSFMs reach +0.3% to +0.9%.
  • Exact duplicates. SparseDirect stays at zero (+0.09 ± 0.12%), but two models are not invariant: Chronos-2's error rises (+1.01 ± 0.17%) and CITRAS-FM's falls (−0.52 ± 0.08%).

The same reference leaves the main result unchanged. In the lagged-copy task, SparseDirect is not told the lag and still recovers 97% of the attainable improvement, while four of the five TSFMs recover at most 2% (details in our response to Reviewer Es7u, Q10).

We will report these narrower observations together with both absolute accuracy and sensitivity.

The structural comparisons use ρ=0.95, λ=0.75, context 512, horizon 8, and five generation seeds with 256 paired windows per seed. Entries are relative MAE changes (%) reported as mean ± SEM across seed-level ratios, with each method’s own Noisy K1 MAE as the denominator.

Table 1. Independent observations: relative MAE reduction (%). Positive values indicate lower MAE.

ModelK2K4K8K16
SSM+0.976 ± 0.216+1.674 ± 0.211+1.969 ± 0.143+2.838 ± 0.313
Chronos-2+0.902 ± 0.403+1.672 ± 0.481+1.464 ± 0.643+0.735 ± 0.709
TimesFM-3+0.651 ± 0.279+1.402 ± 0.290+2.220 ± 0.360+2.598 ± 0.461
TiRex-2+0.000 ± 0.222+0.230 ± 0.231+0.301 ± 0.246+0.317 ± 0.269
t0+0.250 ± 0.301+0.364 ± 0.341+0.706 ± 0.251+0.934 ± 0.329
CITRAS-FM+0.132 ± 0.032+0.226 ± 0.041+0.454 ± 0.064+0.557 ± 0.065
SparseDirect+0.052 ± 0.392+0.780 ± 0.364+0.313 ± 0.944+1.073 ± 0.933

Table 2. Exact duplicates: relative MAE increase (%). Positive values indicate higher MAE.

ModelK2K4K8K16
SSM0.000 ± 0.0000.000 ± 0.0000.000 ± 0.0000.000 ± 0.000
Chronos-2+0.078 ± 0.042+0.260 ± 0.094+0.602 ± 0.140+1.014 ± 0.171
TimesFM-3−0.101 ± 0.077−0.022 ± 0.131+0.132 ± 0.155+0.288 ± 0.172
TiRex-2−0.062 ± 0.023−0.097 ± 0.044−0.102 ± 0.062−0.100 ± 0.079
t0−0.094 ± 0.057−0.118 ± 0.099−0.081 ± 0.110+0.008 ± 0.108
CITRAS-FM−0.125 ± 0.024−0.264 ± 0.048−0.400 ± 0.069−0.518 ± 0.084
SparseDirect+0.074 ± 0.086+0.116 ± 0.123+0.075 ± 0.139+0.092 ± 0.122

Table 3. Distractors: relative MAE increase (%). Positive values indicate higher MAE.

ModelK2K4K8K16
SSM+0.013 ± 0.040+0.048 ± 0.104+0.041 ± 0.105+0.295 ± 0.256
Chronos-2−0.049 ± 0.271+0.001 ± 0.533+0.674 ± 0.746+1.452 ± 0.737
TimesFM-3+0.060 ± 0.106+0.033 ± 0.144+0.050 ± 0.143+0.013 ± 0.164
TiRex-2+0.791 ± 0.152+1.065 ± 0.165+1.063 ± 0.174+0.997 ± 0.173
t0−0.078 ± 0.306+0.009 ± 0.433+0.229 ± 0.394+0.628 ± 0.267
CITRAS-FM+0.189 ± 0.117+0.269 ± 0.153+0.275 ± 0.230+0.131 ± 0.162
SparseDirect+0.637 ± 0.609+1.009 ± 0.505+1.863 ± 1.044+2.624 ± 0.626
AuthorsOpenReview ↗

Response to Reviewer NdnY (2/2)

[Q1 : Architecture and pretraining background]

We added a short background section to the revised paper, uploaded with this response (Appendix A.1). It describes the architecture, cross-channel mechanism, and pretraining of each released checkpoint, for the original three models (Chronos-2, TimesFM-3, and TiRex-2) and the two added models. In brief, for the added models: CITRAS-FM [1] is a small decoder-only patch Transformer with cross-time and shifted cross-variate attention, and t0 [2] is a decoder-style patch Transformer alternating temporal and group attention. All five checkpoints natively accept known-future covariates.

[Q2 : Evidence from additional TSFMs]

We agree that additional models are important for assessing the scope of the findings. To address this, we added CITRAS-FM [1] and t0 [2], two recent TSFMs with native support for known-future covariates. We evaluated their frozen checkpoints using the same temporal, structural, and irrelevant-future protocols.

As summarized in Tables 4–5, both models have small response components along the true response template, already at zero lag. For conditionally irrelevant future covariates, t0 reacts more than SparseDirect and CITRAS-FM reacts less. These results establish under-response rather than a common pattern of lag-induced collapse. Their structural behavior remains different: t0 has a mean distractor penalty, while CITRAS-FM’s distractor change is small relative to its SEM and its MAE decreases with duplicated inputs (Tables 2–3). Your question led us to test two more models, and the answer differs by pattern. None of the five tested models uses the delayed relation well: Chronos-2 and TimesFM-3 lose it as the lag grows, while TiRex-2, t0, and CITRAS-FM respond weakly even at zero lag. Sensitivity to irrelevant inputs extends to t0 but not to CITRAS-FM.

Table 4. Temporal response gain: original models, additional TSFMs, and the context-fitted baseline (corresponding to Figure 2 of the paper).

Model / referenceGain: lag 0Gain: lag 16Gain: lag 64
EventResponseFit0.686 ± 0.0030.713 ± 0.0040.670 ± 0.004
Chronos-20.468 ± 0.0030.342 ± 0.0020.002 ± 0.000
TimesFM-30.533 ± 0.0050.035 ± 0.0010.005 ± 0.000
TiRex-20.018 ± 0.001-0.005 ± 0.001-0.001 ± 0.000
t00.047 ± 0.0010.001 ± 0.0000.000 ± 0.000
CITRAS-FM0.002 ± 0.0000.001 ± 0.0000.000 ± 0.000

The setup is the same as in Sec. 2 of the paper. Entries are mean ± SEM across generation seeds. Gain is the signed projection of the mean forecast response onto the true injected response, normalized so that the exact response has gain 1. Zero gain means zero projection onto that template.

Table 5. Sensitivity to conditionally irrelevant future covariates: original models, additional TSFMs, and statistical baselines (corresponding to Table 1 of the paper).

Model / referenceS_future^50: lag 8S_future^50: lag 16
SelectedLagOLS0.000 ± 0.0000.000 ± 0.000
SparseDirect0.030 ± 0.0010.021 ± 0.001
Chronos-20.162 ± 0.0010.154 ± 0.000
TimesFM-30.114 ± 0.0010.053 ± 0.001
TiRex-20.128 ± 0.0010.061 ± 0.001
t00.074 ± 0.0010.077 ± 0.001
CITRAS-FM0.020 ± 0.0010.007 ± 0.000

The setup and the sensitivity measure are the same as in Table 1 of the paper (noisy lagged copy, β = 0.8, context 512, horizon 8). Entries are mean ± SEM across five generation seeds.

Thank you again for the careful review. Your comments helped us clarify the role of the statistical references and the scope of our findings.

[1] CITRAS-FM: Tiny Time Series Foundation Model for Covariate-Informed Zero-Shot Forecasting, arXiv 2026

[2] t0: A Time-Series Foundation Model for Forecasting with Context, arXiv 2026

AuthorsOpenReview ↗

Follow-up on our rebuttal

Thank you again for your suggestions. The uploaded revision includes the requested model background in Appendix A.1. Our rebuttal also includes a statistical baseline for the structural and lagged-copy experiments and evaluations of two additional models. These additions clarify the structural comparisons and extend the evidence for underuse of delayed relations across all five tested models. We would be grateful if you could review our responses when you have a chance.

I thank the authors for their detailed response. The rebuttal addressed some of my concerns; however, after considering the other reviewers' feedback as well, I believe the current score is consistent with my overall assessment of the paper, and I therefore maintain my score.

Reviewer_5Bk5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how time series foundation models use information from different channels.

The authors evaluate three models: Chronos-2, TimesFM-3, and TiRex-2.

They use synthetic experiments to study temporal dependencies, shared information, redundant channels, and future covariates.

The results show that these models do not always use useful information. Some models are also sensitive to irrelevant or repeated inputs.

The authors compare them with simple statistical models. These baselines often use the available information more effectively.

Finally, experiments on three real-world datasets show that using predicted covariates can reduce forecast quality. Accounting for covariate uncertainty improves the results but does not fully solve the problem.

Critical strengths

The synthetic experiments allow the authors to control which information is useful. They can then measure how each model reacts to this information.

The comparison with statistical baselines is also valuable. It shows that some failures are not due to missing information.

The paper identifies clear limits of three recent foundation models. The results show that adding more channels does not always improve predictions.

The experiments on predicted future covariates are particularly useful. They show that ignoring covariate uncertainty can lead to poor prediction intervals. This is an important issue for real-world forecasting.

Critical weaknesses

Most experiments use simple synthetic processes, such as linear dependencies and Gaussian noise. The models are also evaluated without fine-tuning.

These settings are useful for understanding model behavior. However, they may not reflect more complex real-world relationships.

It is therefore unclear how broadly the reported failures apply to time series foundation models.

Most experiments use simple synthetic processes, such as linear dependencies and Gaussian noise. The models are also evaluated without fine-tuning.

These settings are useful for understanding model behavior. However, they may not reflect more complex real-world relationships.

It is therefore unclear how broadly the reported failures apply to time series foundation models.

Questions for authors

Do the reported failures persist when the models are adapted to the task, for example through fine-tuning? This would help clarify whether the failures are specific to zero-shot forecasting.

How much of the poor interval coverage comes from the foundation models, and how much comes from the method used to provide future covariates? Could a more complete treatment of covariate uncertainty change the conclusions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 5Bk5 (1/3)

We thank the reviewer for the constructive comments. We conducted additional experiments to examine the scope of the failures, their persistence after adaptation, and the effect of propagating covariate uncertainty. We will incorporate these results and the corresponding qualifications into the revision.

[W : Simple synthetic processes and breadth of the failures] We agree that simple synthetic tests alone cannot show how broadly the failures apply. We therefore tested the same relations on real series and on two more models, and we fine-tuned three models (Q1). The underuse of delayed relations appears on real series and in all five models, and fine-tuning removes it only partly.

Why the tests are simple. Several of these models learn their cross-channel behavior mainly from synthetic multivariate data [1, 2, 4], built for example with same-time, lagged, and latent-factor dependencies [1, 2]. Our tasks are simple examples of these dependency types, so they test whether the trained ability holds in its simplest form. Not every task is linear: in the event–response task, a sign-reversed trigger must produce no response.

Testing on real series. We injected the event–response relations into real hourly electricity-load (GFC17) and half-hourly wind-power (KDD) series and evaluated five models at lags 0, 4, 16, 32, and 64. The same pattern appears on both. TimesFM-3's projection gain drops between lag 0 and lag 4 (0.176 → 0.029 on GFC17), and Chronos-2's gain is near zero by lag 32 (0.350 → 0.013 on GFC17), while the fitted event–response reference keeps gains of 0.33–0.47 at every lag. The relation itself is still planted, so this shows that the failure survives real inputs, not how often such relations occur in practice.

Adding more models. We added CITRAS-FM [3] and t0 [4]. Both have small projection gains even at zero lag (0.002 and 0.047). Chronos-2 and TimesFM-3 have substantial gains at short lags, which fade as the lag grows. All five models therefore underuse the delayed relation, each in its own way. These small gains do not mean that CITRAS-FM and t0 cannot use covariates: on real KDD data, supplying the true future covariates lowers their MASE from 0.955 to 0.460 and from 0.888 to 0.177, respectively.

[Q1 : Do the failures persist after fine-tuning?]

Partly. Fine-tuning largely restores Chronos-2’s response gain within the trained lag range, with only partial transfer beyond it; t0 and CITRAS-FM retain low gains. We fine-tuned Chronos-2, t0, and CITRAS-FM on event–response data with training lags 0–48, using 1,000 steps, 16,000 window exposures per training run, and three training seeds. TimesFM-3 and TiRex-2 are not included because their public releases provide inference code only, without fine-tuning code for the covariate interface.

We report two measures. Gain measures the signed response component along the true response (1 for an exact match). Profile error measures the relative mismatch of the full response curve (0 for an exact match; about 1 for no response). Gain can be near 1 even when the response shape is wrong. Variable-lag training at learning rate 10⁻⁵ gives:

ModelGain, frozen → tunedProfile error, frozen → tuned
Chronos-20.214 → 1.0170.821 → 0.333
t00.008 → 0.0861.000 → 0.954
CITRAS-FM−0.003 → 0.0991.005 → 0.976

Gains are averaged over evaluation lags within the training range (0 to 48), so the frozen values differ from the single-lag gains quoted above. For Chronos-2, the response component along the true response largely recovers within the training range, including lags 33–48 (0.006 → 1.010), and the result is stable across the three training seeds (gain SD 0.011). Its profile error stays at 0.333, so the full response curve still differs noticeably from the target.

Recovery generalizes only partially beyond the training range. With onset positions matched, Chronos-2's tuned gain falls to 0.381 at lag 56, 0.098 at lag 64, and about zero at lag 80. For t0 and CITRAS-FM, gains improve but stay near 0.1 even in range. Fine-tuning also costs general accuracy: on the 26-task fev-bench [5] subset, geometric-mean SQL increases by 10.7% for Chronos-2, 6.8% for t0, and 16.5% for CITRAS-FM. The architecture is unchanged, so Chronos-2 is able to produce stronger in-range responses; why the pretrained model does not is still open. We are currently evaluating whether the gains from fine-tuning Chronos-2 transfer to the real-background injection tasks and will report the results in a follow-up response once available.

AuthorsOpenReview ↗

Response to Reviewer 5Bk5 (2/3)

[Q2 : How much of the poor coverage comes from the foundation models, and how much from the way future covariates are provided?]

Both contribute, and we can separate the two parts. MV is direct multivariate forecasting, as in Table 3 of the paper: it uses covariate history without a predicted future path. Part of the coverage gap is already there under MV, before any predicted covariate is used. This part comes from the model. The rest appears when a predicted covariate is passed in as if it were known (Median). This second part is the larger one in 12 of the 15 model–dataset pairs, and it makes up 76% of the gap to the nominal 80% in the median pair. The SSM and the VAR also lose coverage in this step, so it is not unique to foundation models. Still, the TSFMs lose more than the VAR in 13 of 15 comparisons. Passing the covariate's uncertainty through recovers a median of 77% of this loss (Q3). All methods are compared on the same cohorts with complete future covariates, and the original results are recomputed on these subsets where needed.

Conditional forecasting with true future covariates. When the covariate's true future values are supplied, the original three TSFMs have lower SQL and MAE than the fitted joint SSM and the lag-selected VAR on every dataset.

MV versus predicted median. Supplying a predicted median path as if it were known lowers coverage relative to MV in all 15 model–dataset pairs (Table 1); for TimesFM-3 on KDD, coverage falls from 71.8% under MV to 25.5% under Median. These comparisons show that good conditional forecasting with realized covariates can coexist with deterioration when predicted covariates are supplied as known.

Conventional forecasters versus TSFMs. The SSM also loses coverage—for example, from 81.5% to 41.1% on KDD. The TSFMs show a larger coverage drop than the VAR in 13 of 15 comparisons, but larger than the SSM in only 6 of 15 (Table 4 of our response to Reviewer Es7u).

Table 1. SQL / coverage (%); nominal coverage is 80%. Median supplies one predicted path as known, and Temporal propagates 100 sampled paths (Q3). All methods are evaluated on the same complete-future-covariate subsets: 160 GFC17, 1,339 KDD, and 10,250 Rossmann item–origins. Values are reaggregated on these subsets, which are smaller than the original full cohorts where future covariates are missing.

DatasetModelMVMedianTemporal
GFC17Chronos-2.968 / 69.381.062 / 46.171.032 / 65.65
GFC17TimesFM-3.954 / 71.18.996 / 53.24.953 / 66.54
GFC17TiRex-2.930 / 71.39.942 / 60.87.930 / 66.03
GFC17t0.957 / 70.32.983 / 59.30.974 / 65.13
GFC17CITRAS-FM.990 / 46.241.021 / 38.99.989 / 43.01
KDDChronos-2.828 / 71.89.926 / 46.09.888 / 65.23
KDDTimesFM-3.807 / 71.80.931 / 25.52.799 / 72.53
KDDTiRex-2.735 / 84.05.839 / 30.48.728 / 79.82
KDDt0.732 / 79.88.808 / 30.18.720 / 68.54
KDDCITRAS-FM.802 / 53.04.849 / 34.39.798 / 43.48
RossmannChronos-2.459 / 76.51.507 / 37.43.469 / 73.93
RossmannTimesFM-3.490 / 82.97.545 / 36.92.497 / 78.23
RossmannTiRex-2.520 / 81.81.562 / 47.88.524 / 77.77
Rossmannt0.475 / 82.40.539 / 36.56.485 / 80.26
RossmannCITRAS-FM.527 / 66.51.538 / 56.46.534 / 57.08
AuthorsOpenReview ↗

Response to Reviewer 5Bk5 (3/3)

[Q3 : Could a more complete treatment of covariate uncertainty change the conclusions?]

Yes. Instead of one median path, we sampled 100 future covariate trajectories and pooled the resulting target forecasts. TSFMs return a separate distribution for each future step, so the samples must be linked across time. We used a Gaussian copula with an AR(1)-style temporal correlation estimated from the observed history, following [6, 7].

Temporal propagation improves SQL and coverage over Median in all 15 pairs (Table 1) and recovers a median of 77% of the additional coverage loss from MV to Median (range 6–102%). SQL ends close to MV for most pairs; the largest remaining gaps are for Chronos-2 on GFC17 (6.6%) and KDD (7.2%). For the SSM and the VAR, propagating their own joint Gaussian trajectories brings SQL back to within 0.7% of MV.

Recalibration. Correcting the target intervals on historical backtests after propagation reduces the mean absolute deviation from 80% coverage from 8.4 to 3.4 percentage points across the nine original-model comparisons (better in eight of nine), at an SQL cost of at most 1.0%. For Rossmann, all compared methods use the same 10,070 of 10,250 item–origins with sufficient observations in the historical calibration block. Recalibrating only the covariate forecasts has mixed effects.

We will therefore revise the conclusion of Secs. 4.2–4.3. Supplying an uncertain covariate as if it were known lowers coverage relative to MV, for conventional forecasters as well as TSFMs. Propagating that uncertainty recovers much of this additional loss in most model–dataset pairs, and recalibrating the target intervals brings coverage closer to 80% in eight of the nine comparisons we tested.

We thank the reviewer for prompting these additional experiments and helping us clarify the scope of the conclusions.

[1] Chronos-2: From Univariate to Universal Forecasting, arXiv 2025

[2] TiRex-2: Generalizing TiRex to Multivariate Data and Streaming, arXiv 2026

[3] CITRAS-FM: Tiny Time Series Foundation Model for Covariate-Informed Zero-Shot Forecasting, arXiv 2026

[4] t0: A Time-Series Foundation Model for Forecasting with Context, arXiv 2026

[5] fev-bench: A Realistic Benchmark for Time Series Forecasting. NIPS 2026

[6] Efficiently Generating Correlated Sample Paths from Multi-step Time Series Foundation Models, arXiv 2025

[7] Interweaving Marginals into Multivariate Sample Paths: Training-Free Dependence Construction for Probabilistic Time Series Foundation Models, arXiv 2026

AuthorsOpenReview ↗

Follow-up on our rebuttal

Thank you again for your review. We added real-background evaluations, two additional models, and fine-tuning experiments. Chronos-2 improves substantially within the trained lag range, with limited extrapolation and a real-data accuracy trade-off. For predicted covariates, uncertainty propagation recovers a median of 77% of the additional coverage loss from median-path conditioning, and target-interval recalibration further improves coverage. We would appreciate your thoughts on these results and the revised scope of our conclusions.

Reviewer_Es7uOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper tests whether three time series foundation models - Chronos-2, TimesFM-3 and TiRex-2 - actually make good use of cross-series information when forecasting a target, and not just that adding it improves accuracy. In synthetic data, where the true relationships are known and can be controlled, the authors find that each model, to varying degrees, misses signals that arrive with a delay or that are spread across several noisy related series, yet still reacts to duplicate, irrelevant, or uninformative inputs. Simple statistical models fitted to the same data, built to assume the right kind of relationship, recover most of what the foundation models miss. On three real forecasting tasks, feeding a model its own median forecast of the related series, as if it were known, makes its prediction intervals too narrow, and passing in a range of possible covariate paths fixes some, but not all of it.

Critical strengths

I enjoy reading about failure modes, and the authors did a good job with careful diagnostic design: instead of only comparing forecast accuracy, they constructs settings where they control everything, so that the predictive value of each channel is known, and they can distinguish underusing useful information from being distracted by irrelevant information.

The references are well chosen and honestly framed, exact oracles are the ceiling, and statistical models fitted to the same data show what is achievable without knowing the true parameters. The authors state each baseline’s assumptions and when an advantage comes from construction, as with duplicate removal.

I also like careful reporting: five seeds, paired comparisons, uncertainty on most results, and explicit notes on what each experiment can and cannot show. The real-data section tests a common practical shortcut, feeding in predicted covariates as if they were known, and shows its cost on prediction intervals. It also offers a partial remedy, passing several covariate paths instead of one, checked with sensitivity analyses.

Critical weaknesses

Its biggest weakness is the yardstick. The fitted baselines that “recover what the models miss” are built to assume the right kind of relationship, while a zero-shot model must also work out what kind of relationship it faces from a single context. Some reaction to redundant or irrelevant inputs can then be reasonable, and the paper never says how much would be. Its own lag-selected regression reacts to irrelevant future values when the covariate is unrelated. The two covariate-reliability experiments test what any conditional forecaster would do when an uncertain input is treated as known, and no comparison method is run there.

The findings are also narrow. Each failure appears in only some of the three models, several effects are around one percent of MAE, and the synthetic setups are far from typical pretraining data. Yet the title presents them as general failures of foundation models.

Most importantly for a diagnostic paper, it never traces any failure to an architectural choice or a training method. It doesn’t explain why TimesFM-3 loses delayed signals, or why Chronos-2 is hurt by duplicates. Its representation probes don’t identify mechanisms. It also offers no remedies for the cross-channel failures it documents. This is essentially the first third of a solid paper, where the next two sections - what is responsible for these failures and how to fix it - don't yet exist.

Questions for authors

Set a fair standard for how much a model should react to redundant or irrelevant inputs. The authors could build a reference that, like a foundation model, doesn’t know the true structure: a Bayesian average over plausible structures (which lags, which channels matter, whether any relationship exists), or generic fitted methods such as a vector autoregression with lag selection or cross-validated sparse regression over all channels and lags. Measuring how much these react to duplicates, distractors and irrelevant future values would show how much sensitivity is reasonable from one context. The models could then be judged by their excess over that amount. The context-length check from Appendix D.5 should be extended to every experiment, to show whether this sensitivity falls as evidence accumulates, as it should for a rational learner.

Give the covariate-reliability results a point of comparison. The mismatch experiment and the median-path pipeline should be run with the fitted state-space model and the lag-selected regression, to show whether the foundation models lose more than a standard conditional forecaster does. To make the remaining gap after propagation interpretable, the authors should use sampled covariate trajectories instead of fixed-rank quantile paths and recalibrate the covariate forecasts before propagating them. Otherwise that gap can’t be separated from the approximation and the covariate forecasts’ own undercoverage.

Show how general these findings are and how much they matter. Right now, each failure is seen in only one or two of three models, so it is unclear whether these are traits of foundation models or quirks of particular ones. Testing more covariate-capable models the paper already cites, such as t0 and CITRAS-FM, could settle that. The synthetic series are also unlike real data, so planting the same known relationships into real series, rather than into white noise or random walks, would show whether the failures persist on realistic inputs. Effects are reported in the paper only as raw MAE changes, which makes them hard to judge, so each should also be stated as a share of the improvement the oracle shows is possible. If the failures turn out model-specific or small, the title should say so instead of claiming foundation models fail in general.

Failures must be traced to their causes for this to be useful. To find where a delayed trigger’s information is lost, the authors could patch activations between the event and no-event runs layer by layer, or block cross-channel attention at chosen depths. For TimesFM-3’s sharp drop at a lag of four steps, they could shift the trigger relative to the model’s patch boundaries to test whether patching is responsible. Briefly fine-tuning each model on synthetic data with delayed relations and duplicate or distractor channels would show whether a failure comes from the training data or from the architecture.

Finally, propose some remedies and test them on both the synthetic and real tasks. Candidates include that fine-tuning or added synthetic pretraining data, removing duplicate and near-duplicate channels before forecasting, and choosing between target-only and multivariate input by backtesting within the observed context, as the regression baseline already does. For predicted covariates, propagating covariate uncertainty and then recalibrating the intervals on rolling backtests may also help. Full retraining of every model is of course not a reasonable expectation, but you should walk down some of these paths at least a little, perhaps at small sizes. But remedies cannot be left purely to specualtion.

10 comments
AuthorsOpenReview ↗

Response to Reviewer Es7u (1/8)

We thank the reviewer for the detailed and concrete suggestions, and we answer each of them below.

We number the weaknesses W1–W5 and the requests Q1–Q18 in the order they appear in the review. The answers are grouped by topic, and each heading states the point it answers.

[W3a / Q8 : Test more covariate-capable models, such as t0 and CITRAS-FM]

With five models, the underuse of delayed relations appears in every model; the other findings remain model-specific. We added CITRAS-FM [1] and t0 [2] to all synthetic experiments and to the real-data pipeline.

Table 1. Event–response gain (five seeds). Gain measures the signed component of the mean forecast response along the true injected response, normalized so that the exact response has gain 1. Zero gain means no component along that template; it does not imply an absent response. EventResponseFit is a regression fitted to the same context.

Model / referenceLag 0Lag 16Lag 64
EventResponseFit0.6860.7130.670
Chronos-20.4680.3420.002
TimesFM-30.5330.0350.005
TiRex-20.018−0.005−0.001
t00.0470.0010.000
CITRAS-FM0.0020.0010.000

There are two patterns. Chronos-2 and TimesFM-3 have substantial projection gains at short lags, which fade as the lag grows. TiRex-2, t0, and CITRAS-FM have small gains at every tested lag, indicating little response along the true response template. These small gains do not mean that the models cannot use covariates: on real KDD data, supplying the true future covariate lowers MASE from 0.888 to 0.177 for t0 and from 0.955 to 0.460 for CITRAS-FM.

A. A fair standard for reactions to redundant or irrelevant inputs

[W1 / Q1 : Add a reference that does not know the true structure]

We added SparseDirect, the cross-validated sparse regression over all channels and lags that the reviewer suggested. It fits one elastic-net regression per forecast step. It chooses between target-only and all-channel inputs, the lag length (up to 64), and the regularization strength by validation on two folds inside the observed context. It is not told the latent state, the generating parameters, which channels are informative, or the true lag, and it does not remove duplicate channels.

We also tried the reviewer's first suggestion, a Bayesian average over candidate structures (which lags and channels matter, and whether any relation exists). Its fits did not converge reliably on single contexts of this length in a substantial share of cases, so we do not use it as a reference. We will describe the attempt and its diagnostics in the appendix.

SparseDirect still assumes a sparse linear relation, so it is a structure-agnostic reference among linear models, not a universal standard. We designed it after seeing the original results and will label it as exploratory. A small share of fits stopped at the solver's iteration limit; we keep their outputs and will report the diagnostics in the appendix. The fitted SSM stays in the paper, described as a reference that matches the generator's model class.

[W2a / Q2 : How much does the reference react to duplicates, distractors, and irrelevant future values?]

The full shared-latent results for K = 2, 4, 8, and 16 channels, including SparseDirect, are in Tables 1–3 of our response to Reviewer NdnY (part 1/2). We refer to them here as Tables N1–N3 and describe the pattern.

SparseDirect shows little reaction to duplicates: its MAE change stays between +0.07% and +0.12% at every K, within its standard error. It does react to distractors, and more as channels are added: +0.64% at K = 2 and +2.62 ± 0.63% at K = 16 (Table N3). Irrelevant channels therefore hurt the fitted reference too.

Table 2 measures how much predictions change when irrelevant future values are replaced. For SparseDirect, the change is 0.030 at lag 8 and 0.021 at lag 16. When the covariate is entirely unrelated to the target (β = 0), both OLS and SparseDirect react. These measure prediction changes, not accuracy losses.

Table 2. Sensitivity to irrelevant future covariate values. In the lagged-copy task (lags 8 and 16, horizon 8), the target depends only on past covariate values, so the supplied future values add no information. Sensitivity is the mean change in the point forecast when only those future values are replaced by other draws. The correct value is 0; for scale, the target-only MAE is about 0.80. At β = 0 the covariate is unrelated to the target.

Model / referenceβ = 0.8, lag 8β = 0.8, lag 16β = 0
Lag-selected OLS (paper)0.0000.0000.009
SparseDirect0.0300.0210.038
Chronos-20.1620.1540.015
TimesFM-30.1140.0530.022
TiRex-20.1280.0610.013
t00.0740.0770.070
CITRAS-FM0.0200.0070.007
AuthorsOpenReview ↗

Response to Reviewer Es7u (2/8)

[Q3 : Judge the models by their excess over that reference]

Judged by the excess over SparseDirect, the four findings stand as follows. The first three come from the shared-latent task (Tables N1–N3) and the fourth from the lagged-copy task (Table 2).

  • Distractors (not excessive, but avoidable). The ideal reaction is zero. SparseDirect reacts at every K, because telling an irrelevant channel from a relevant one needs evidence that a single context may not contain, and no TSFM clearly exceeds it at any K. We therefore do not call the reaction a failure. What remains is a difference between models. TimesFM-3 does not react at any K (at most +0.06%) while gaining the most from informative channels, so the reaction can be avoided zero-shot. TiRex-2 shows a penalty of 0.8% to 1.1% at every K; for Chronos-2 and t0 it grows with K and reaches +1.45 ± 0.74% and +0.63 ± 0.27% at K = 16.

  • Duplicates. An exact copy adds no information, so the correct forecast is unchanged, and SparseDirect stays at zero. Two models depart from this, and their reaction grows steadily with the number of copies. For Chronos-2 it is an accuracy penalty: its error rises from +0.08% at K = 2 to +1.01 ± 0.17% at K = 16. CITRAS-FM's error instead falls, from −0.13% to −0.52 ± 0.08%. This is not a loss of accuracy, but its forecast depends on how many times the same information is supplied, which could suggest that it underweights a single covariate (consistent with its weak covariate response in Tables 1 and 7). In practice, accidental duplicates in a data pipeline would change its forecasts in a direction that cannot be guaranteed.

  • Independent channels. TimesFM-3 stays close to the class-matched SSM at every K and reaches +2.60 ± 0.46% at K = 16 (SSM +2.84 ± 0.31%), so this gain can be reached zero-shot without knowing the model class. Chronos-2 matches the SSM up to K = 4 (+1.67%) but then gets worse as channels are added (paired MAE difference from K = 4 to K = 16: +0.0072 ± 0.0020, Table 18 of the paper). TiRex-2, t0, and CITRAS-FM stay below +1% at every K. SparseDirect's own gain is too uncertain (+1.07 ± 0.93% at K = 16) to separate any TSFM from it, so we state this as a difference between models, not as underuse relative to that reference.

  • Irrelevant future values. When the covariate is related to the target through its past (β = 0.8), Chronos-2, TimesFM-3, TiRex-2, and t0 react 2.5 to 7 times as much as SparseDirect; CITRAS-FM reacts less. When the covariate is unrelated (β = 0), only t0 reacts more than SparseDirect, so we make no excess claim there for the other models.

SparseDirect is not a better forecaster for reacting less: with one noisy channel, its MAE (0.801) is higher than that of every TSFM (0.779–0.787). We will report absolute error next to every sensitivity number.

The same reference does not narrow the main result. In the lagged-copy task, SparseDirect recovers 97% of the attainable improvement, while four of the five TSFMs recover at most 2% (Q10).

[Q4 : Extend the context-length check to every experiment] We are extending the context-length comparisons and will report the results in a follow-up response once available.

AuthorsOpenReview ↗

Response to Reviewer Es7u (3/8)

B. Covariate reliability: comparison methods and uncertainty propagation

[W2b / Q5 : Run the mismatch and median-path experiments with the fitted SSM and a lag-selected regression]

Both effects also appear in the conventional forecasters, so they are a general risk of treating an uncertain input as known, not a property of foundation models alone. Among the TSFMs, the size of the effect differs by model. We added a fitted joint Gaussian state-space model (SSM) and a lag-selected VAR to both experiments; both are fitted and selected on the observed history only.

Reliability mismatch (Sec. 4.2). Here the covariate history is clean and only the supplied future values are unreliable. Table 3 compares this with the case where history and future are equally unreliable, at the lowest reliability (λ = 0).

Table 3. Increase in MAE when only the future covariate is unreliable (λ = 0, five seeds).

ForecasterMAE increaseExtra over VAR
Fitted joint SSM0.444+0.067
Lag-selected VAR0.377–
Chronos-20.459+0.083 ± 0.013
TimesFM-30.475+0.099 ± 0.009
TiRex-20.435+0.058 ± 0.009
t00.348−0.029 ± 0.016
CITRAS-FM0.065−0.312 ± 0.019

Chronos-2, TimesFM-3, and TiRex-2 lose about as much as the SSM and more than the VAR. t0 loses about as much as the VAR. CITRAS-FM is barely affected. It also makes little use of the covariate in the other synthetic tasks (Tables 1 and 7), and a model that does not rely on the covariate cannot be misled by it. For both conventional forecasters the effect shrinks as the covariate becomes more reliable (SSM: 0.444 at λ = 0 to 0.100 at λ = 0.75).

Median path on real data (Sec. 4.3). MV is direct multivariate forecasting, as in Table 3 of the paper: it uses covariate history without a predicted future path. Median supplies one predicted covariate path as if it were known.

Table 4. Coverage (%) of the nominal 80% interval, MV → Median. All methods are evaluated on the same complete-future-covariate subsets: 160 GFC17, 1,339 KDD, and 10,250 Rossmann item–origins. Values are reaggregated on these subsets, which are smaller than the original full cohorts where future covariates are missing.

ForecasterGFC17KDDRossmann
Fitted joint SSM77.8 → 67.181.5 → 41.174.9 → 28.1
Lag-selected VAR57.9 → 51.176.7 → 57.484.7 → 56.3
Chronos-269.4 → 46.271.9 → 46.176.5 → 37.4
TimesFM-371.2 → 53.271.8 → 25.583.0 → 36.9
TiRex-271.4 → 60.984.1 → 30.581.8 → 47.9
t070.3 → 59.379.9 → 30.282.4 → 36.6
CITRAS-FM46.2 → 39.053.0 → 34.466.5 → 56.5

Coverage falls for every forecaster on every dataset. The TSFMs' drop is larger than the VAR's in 13 of 15 comparisons and larger than the SSM's in 6 of 15; their SQL increase is likewise larger in 13 of 15 and 6 of 15. CITRAS-FM again changes least, and its coverage is far below 80% even under MV. When the true future covariate is supplied instead, the three original TSFMs and t0 have lower SQL than both conventional forecasters on every dataset. Weak conditional forecasting therefore does not explain the loss: it appears when a predicted covariate is passed in as if it were known.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks from the Authors

We thank the reviewers for their careful reading and constructive suggestions. Their feedback helped us strengthen the statistical comparisons, broaden the evaluation to five time series foundation models (TSFMs), and clarify which findings depend on the model or task. We have incorporated the analyses and experiments summarized below into the revised manuscript. Revisions are marked in blue.

Strengths

  • Controlled evaluation of cross-channel information use. The synthetic experiments specify which channels carry predictive information, allowing us to distinguish weak use of informative signals from sensitivity to redundant or irrelevant inputs. [Es7u, 5Bk5, NdnY]
  • Interpretable statistical references. Exact oracles quantify the improvement available from the supplied information, while statistical models fitted to the same observed history show how much of that improvement a forecaster can recover from a finite context. [Es7u, 5Bk5, NdnY]
  • Practical relevance for probabilistic forecasting. The real-data experiments examine how to use a covariate whose future values must themselves be predicted. All three reviewers highlighted the consequences of this common forecasting decision for target prediction intervals. [Es7u, 5Bk5, NdnY]

Discussion

Discussion topicR-Es7uR-5Bk5R-NdnYLocation in the revision
Statistical baselines and sensitivity to redundant or irrelevant inputsQ1–3W1Sections 3 and 4.1, Appendices C–D
Generality across TSFMs, effect sizes, and scope of the claimsQ8, Q10–11Generality concernQ2Sections 2–4, Appendix D.5
Delayed responses on real GFC17 and KDD backgroundsQ9Realism concernAppendix B.5
Sensitivity to context lengthQ4Appendix J
Event information in model representations, cross-channel attention, and patch alignmentQ12–13Appendices G.3 and I
Fine-tuning, extrapolation to unseen delays, and forecasting trade-offsQ14–15Q1Appendix H
Forecast-time covariate reliability and its effect on interval coverageQ5Q2Sections 4.2–4.3, Appendices E–F
Covariate trajectory propagation and interval recalibrationQ6–7, Q18Q2–3Section 4.3, Appendix F
Model architectures, cross-channel interactions, and pretrainingQ1Appendix A.1
Rating (initial → after discussion)1 → 33 → 33 → 3

Key outcomes of the discussion

  • More appropriate statistical baselines. We added SparseDirect, which selects channels and lags from the observed history, and evaluated fitted state-space and vector autoregressive models as controls for covariate conditioning. SparseDirect itself incurs a distractor penalty, and none of the five TSFMs shows a clearly larger penalty. With exact duplicates, SparseDirect’s error changes little, while additional copies increase Chronos-2’s error and decrease CITRAS-FM’s. [Es7u, NdnY]

  • A practical remedy for predicted covariates. Propagating sampled covariate trajectories improves both quantile loss and interval coverage over median-path conditioning. Trajectory propagation recovers a median of 77% of the coverage lost when switching from direct multivariate forecasting to median-path conditioning. [Es7u, 5Bk5]

  • Fine-tuning and generalization to unseen delays. Fine-tuning largely restores Chronos-2’s response gain within the trained lag range, although some error in the response shape remains and the gain falls at longer, unseen delays. Under the same training budget, t0 and CITRAS-FM show smaller improvements. All three models incur higher quantile loss on the evaluated real-world forecasting tasks after fine-tuning. [Es7u, 5Bk5]

Contribution

Recent time series foundation models give forecasters several ways to supply related series and covariates. Our study asks when these inputs improve a forecast and when they change it despite adding no predictive information. We evaluate five recent TSFMs in controlled experiments where the predictive value of each channel is known by design. Across the five models, useful signals can go underused when they arrive with a delay or are spread across noisy channels, while redundant inputs can still change forecasts.

Across three real-world tasks, all five models achieve higher target interval coverage and lower quantile loss when they propagate sampled covariate paths instead of conditioning on a single predicted median path.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.