acceptodds
Under review as a conference paper at ICLR 2027

PathFM: Fast Generative Forecasting of Multivariate Trajectories

Abstract

Time series foundation models are becoming increasingly popular for general-purpose forecasting. However, most leading models predict marginal quantiles, leaving the dependence structure of future trajectories unspecified. This distinction matters for trajectory-dependent quantities such as cumulative demand, peak load, and event timing, yet point and marginal scores do not reveal it. To model this dependence, we introduce PathFM, a trajectory-based forecasting model that uses flow matching to generate complete future paths. We further distill the model to one or two steps and share history attention cache across draws. Together, these enable PathFM to generate multiple paths for a single series in roughly the time needed for one query to a quantile model. The distilled sampler also makes it practical to optimize a calibration loss directly. PathFM achieves competitive marginal accuracy on the established forecasting benchmarks fev-bench and TIME. To complement these evaluations, we introduce Forecast Path Score Bench (fps-bench), which augments fev-bench tasks with energy and variogram scores and downstream metrics. PathFM achieves the highest aggregate path score on fps-bench among the evaluated models. Experiments on chaotic systems further show that our flow model sustains long-run variability, while strong baselines rapidly collapse. PathFM makes generative time series forecasting both competitive and practical, enabling diverse downstream applications.

Primary area:
learning on time series and dynamical systems
Keywords:
time series foundation models, trajectory forecasting, flow matching, probabilistic calibration, chaotic systems, trajectory evaluation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_NtmGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PathFM, a time series foundation model that generates forecast trajectories with flow matching. PathFM predicts competitive marginal quantiles while also producing realistic sample paths with dependence across time steps. The authors introduce a new benchmark (FPS) to evaluate sample paths and show that PathFM sets the state-of-the-art on their benchmark at a reasonable latency (achieved via parallel decoding distillation).

Critical strengths
  1. FPS fills an important gap in previous literature around evaluating sample paths (rather than just marginal forecasts)
  2. Flow matching is a well-motivated solution to sample accurately from the joint distribution over future time steps
  3. PathFM performs competitively to other time series foundation models on marginal accuracy, while providing improved sample paths
Critical weaknesses
  1. Missing key details: PathFM is supposedly a near state-of-the-art time series foundation model (even when considering marginal accuracy alone). This is an impressive feat! However, the only information given on the training data is "a 83M-row dataset of real-world and synthetic time series." Additionally, the submission did not include code or model weights (though the authors say these will be released later). This lack of information makes it impossible to reproduce PathFM or decouple the contributions of the training data vs architecture and training objective. I would have quite liked to test out the model as part of this review.

  2. Several parts of the paper are written in a way that is quite dense and hard to read (likely AI-generated), which is a shame as I think there are many interesting details that could be made more accessible. This is particularly prevalent in the appendix.

Questions for authors
  1. Are there any comparisons between PathFM with stratified noise ("ensemble sampling") against baselines with "fair" sampling? This seems like an unfair comparison as stratified noise could also improve baselines. I would prefer all evaluations to use fair sampling.

  2. Can you please provide more details on the training data? As you say this includes real-world time series, is there any risk of leakage/contaimination of the evaluation benchmarks?

  3. The baselines with the Gaussian copulas are interesting. How do these compare to the simple method from [1]? Also, for the baselines, you used "constant extension beyond the outer quantiles". Does the IQF approach from [1] improve the baselines?

[1] Baron et al. Efficiently Generating Correlated Sample Paths from Multi-step Time Series Foundation Models. https://arxiv.org/pdf/2510.02224

2 comments
AuthorsOpenReview ↗

Response to Reviewer NtmG

Thank you for your careful and encouraging review. We address your concerns as below.

We have thoroughly revised the draft, with significant changes highlighted in blue. We extended the pretraining from 1M to 2M steps with the same data and architecture, denoted as "updated PathFM".

Reply to W1 and Q2 (training data, code and weights)

Training data. The general models are trained on an 83-million-row corpus of real and synthetic time series. The real part is drawn from public univariate datasets: energy (Electricity, Solar, Wind Farms, London Smart Meters, Buildings 900K), transport (Mexico City Bikes, Pedestrian Counts, Taxi, Uber TLC, LargeST, Q-Traffic), nature (KDD Cup 2018, USHCN, WeatherBench, Temperature-Rain), cloud operations (Alibaba Cluster Trace 2018, Azure VM Traces 2017, Borg Cluster Data 2011), web (Wiki), and the M4 daily, hourly, weekly and monthly sets. The synthetic part has two kinds of data. Univariate series come from a generator that combines random trend, seasonality and irregular components, and from one that samples random temporal causal graphs and generates series autoregressively. Multivariate series are built from univariate base generators (AR, ETS, the trend–seasonality generator, and KernelSynth) by adding dependence either at the same time step or across time, such as lead–lag relationships and cointegration.

Leakage. The corpus was built to exclude all fev-bench datasets, so the fps-bench tasks are also excluded, and the test portions of every GIFT-Eval task. It does overlap with the training portions of some GIFT-Eval datasets.

Data versus model design. We agree that comparisons with other pretrained models cannot separate these, but Chronos-2 is a close control in both training data (same corpus) and model size. PathFM-126M significantly improves FPS over Chronos-2 with its strongest completion (0.5386 vs. 0.5274; 95% CI [0.0028, 0.0200]). Since the training still differs, this is suggestive.

Code and weights. Thank you for raising this. We now provide the code and weights at https://github.com/pathfm/pathfm-supplementary-submission. The code is also uploaded to the supplementary material of the revised submission. We welcome you to try it directly!

Reply to W2 (dense writing)

Thank you for your thorough review and for recognizing the merits of the Appendix. We agree that parts of the paper, especially the Appendix, are dense. The paper reports a large number of results, and presenting them all compactly is not easy. We have nonetheless done our best to improve the presentation. Specifically, we have:

  1. added an overview at the start of the Appendix that says the motivation and claims of each section;
  2. made the motivation for our design choices and experiments more explicit throughout the paper, including but not limited to PDF lines 223, 261, 902, 958, and 1111;
  3. consolidated the copula, tail, and sampling comparisons into one grouped table (Appendix E.3, Table 14), and placed marginal-score uncertainty beside the benchmark scores (Appendix B.6, Tables 6–7);
  4. incorporated new experiments into the paper and revised it throughout.

Reply to Q1 (fair sampling for every model)

Thank you for pointing out an issue in the original comparison. We now compare both PathFM and TimesFM-3 (the strongest baseline) under the same sampling scheme (iid or stratified). Specifically, TimesFM-3 uses a history-fitted AR(1) copula and IQF tails, which is also the strongest completion method we have tested.

Sampling (both models)PathsUpdated PathFM-300M FPSTimesFM-3 + AR(1)/IQF FPSDifference95% CI
iid320.53410.5373−0.0031[−0.0176,0.0107]
iid1280.54970.5458+0.0038[−0.0079,0.0161]
iid5120.55230.5480+0.0043[−0.0071,0.0162]
iid2,0000.55430.5486+0.0057[−0.0053,0.0174]
Stratified320.54580.5446+0.0011[−0.0100,0.0129]
Stratified5120.55430.5486+0.0057[−0.0054,0.0173]

We can see that PathFM leads the average FPS in every case except iid sampling + 32 paths. None of these differences from TimesFM-3 is significant. We have refined our claim in the manuscript to "PathFM ranks first in FPS, although its advantage over the second-place method is not significant." The revised manuscript now reports iid and stratified results for all models.

Reply to Q3 (AR completion and IQF tails)

Thank you for the reference. We have cited it and implemented the history-based AR(1) construction of Baron et al. (2025) for quantile baselines. It is the strongest completion we have tested, ahead of the other Gaussian copulas we fitted (residual AR(1), Toeplitz, and nonstationary covariance). We have therefore updated the FPS result table in the paper and in our reply to Q1 above. We specifically compared PathFM with TimesFM-3 + AR(1)/IQF to confirm that PathFM keeps the highest average FPS, though not statistically significant. We have updated our draft to be accurate on this claim.

Thank you again for your review!

Dear Reviewer NtmG,

Thank you again for your review. In the revision, we released the code and model weights, clarified the training data, added iid-sampling comparisons for all models, evaluated IQF tails and the Baron et al. completion, and revised the writing (including a new appendix overview). We would be glad to clarify anything further before the discussion period ends.

Best regards, Authors

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

This paper proposes PathFM, a time series foundation model that generates complete joint future paths over both the forecast horizon and channels via conditional flow matching, rather than the per-step point estimates or marginal quantiles most current TSFMs predict. The motivation is that marginal forecasts leave the temporal and cross-variable dependence of the future unspecified, so path-dependent quantities like cumulative demand, peak load, or event timing cannot be recovered from them. To make generative forecasting practical, PathFM distills the flow matching teacher model down to a 1–2 evaluation sampler via parallel decoding distillation, share and cache history computation across paths and flow steps, backpropagate a fair CRPS calibration loss directly through the one-step sampler, and stratify the source noise for small ensembles. A second contribution is fps-bench, which augments 94 fev-bench tasks with six path-level scores to evaluate trajectory distributions. Empirically, PathFM ranks first on fps-bench while staying competitive on marginal benchmarks (3rd on fev-bench, 4th on TIME), and on 135 chaotic DYSTS systems it beats fine-tuned Chronos and the specialized PANDA model on both point accuracy and long-run dynamical fidelity — all from a 126M-parameter model that generates 32 paths in about 52 ms.

Critical strengths
  • The evaluation framework addresses a real gap, and it's built the right way. Marginal quantiles genuinely can't pin down the joint law of the future, so path-dependent quantities (totals, peaks, event timing) are invisible to existing benchmarks. fps-bench fixes this using established tools (energy/variogram scores, fair estimators, CRPS on path functionals) — not metrics invented to favor the model — and the independent vs. comonotone completions are a clean control: same marginals, different dependence, so score differences really do isolate what's being tested. The finding that the marginal winner (TimesFM-3) is not the path winner is exactly the blind spot this benchmark exposes. This contribution stands on its own regardless of where PathFM ranks.

  • The pipeline hangs together, the ablations are unusually honest, and there's independent evidence. The pieces are borrowed (PDD from image/video, scoring-rule post-training from the RMMD lineage) but they genuinely interlock: the 1–2 NFE sampler is what makes backpropagating fair CRPS through the deployed sampler possible, and the paper carefully maintains that loss's statistical preconditions. The ablation chain self-reports its own warts — the accuracy tax of history-only attention, the teacher cap, the ~9-point coverage gap left after calibration, distillation degrading on chaotic rollouts — which makes us trust the rest of the reporting. And the DYSTS results, beating the specialized PANDA at matched parameter count, are independent evidence pointing the same way as the main claims.

Critical weaknesses
  • The headline fps-bench margin is smaller and shakier than presented. The whole value proposition rests on the path advantage, because in the marginal view PathFM is beaten by Chronos-2 (higher fev skill, ~2.4× faster). But the ~10% headline margin over TimesFM-3 shrinks to ~2.5% once TimesFM-3 gets even a cheap per-task AR(1) Gaussian copula on top of its frozen quantiles (App. E.3: 0.494 → 0.531 vs. PathFM's 0.545) — and that result sits in an appendix while the main text shows only the weakest baselines. On top of that, the quantile completions use constant tails beyond the 0.1–0.9 levels, which mechanically hurts exactly the Sum/Max/first-passage scores where PathFM gains most — so part of the margin may be interface truncation, not dependence modeling. And no cross-model table has confidence intervals, so even the remaining ~2.5% might not be significant. If the real advantage is ~2.5% and possibly noise, this paper's case is much weaker.

  • The "fast, calibrated" model isn't consistently the one being validated. On chaotic systems, all flagship results use the 10-evaluation teacher; the released 1–2 NFE student is strictly worse there (Table 14), so Section 6 validates the architecture, not the product carrying the "fast" claim. The chaotic models are retrained from scratch but the training data is never stated — DYSTS systems are known ODEs, so on-the-fly trajectory generation would be an unassessable data advantage — and the Chronos fine-tuning protocols are also unspecified. Worse, the long-run diagnostics structurally favor any stochastic sampler over deterministic continuation (PANDA collapses to a small orbit), so part of the "56–74% reduction vs. PANDA" is model class, not model quality — the margin vs. the stochastic Chronos-1 is ~28%, not ~70%. Finally, the "calibrated" claim itself retains a ~9-point coverage gap, and the calibration loss is purely marginal — the path quality that motivates the whole paper is never directly optimized, even though the one-step sampler makes that feasible. Both title words end up demonstrated only in settings where they aren't challenged.

Questions for authors

Q1: App. E.3 fits only a stationary Gaussian AR(1) temporal copula — how much of the remaining ~2.5% gap to PathFM would stronger dependence models (empirical/vine copulas, non-stationary) close, and why are these stronger baselines absent from the main comparison?

Q2: What exactly were the chaotic-domain PathFM models trained on (fixed DYSTS split vs. on-the-fly ODE generation; volume), and how is this decontaminated w.r.t. test trajectories? Please also specify the Chronos-1/2 fine-tuning protocol (data, budget, provenance), and state whether PANDA/Chronos numbers were re-run under your unified protocol or copied from prior work.

Q3: Since no cross-model comparison reports uncertainty, could you add task-bootstrap CIs to all main tables — in particular, is the residual PathFM-vs-fitted-TimesFM-3 margin significant?

Q4: The calibration loss is purely marginal — did you try path-level losses (energy score or Sum/Max CRPS through the one-step sampler), and if they failed (e.g., variance at m=8, interference with the PDD anchor), what happened?

Q5: What is the inference protocol for horizons beyond the 480-step training limit on general benchmarks (chaining? how is sampling error handled?), and does it match how baselines are served?

Q6: One-step generative multivariate forecasters (e.g., TARFVAE) already exist outside foundation scale — did you consider them, and is there a fundamental reason they can't be scaled into foundation models, making distillation unnecessary?

4 comments
AuthorsOpenReview ↗

Response to Reviewer hBUW (1/2)

Thank you for the detailed review and for recognizing fps-bench and the honesty of our ablations. We respond to each specific point below.

We have thoroughly revised the draft, with significant changes highlighted in blue. We extended the pretraining from 1M to 2M steps with the same data and architecture, denoted as "updated PathFM".

Reply to W1, Q1 and Q3 (margin over fitted copulas, stronger dependence models, tails, and uncertainty)

We agree that the fitted completions belong in the main comparison, and that the margin over them is smaller. The main table now gives each marginal baseline its best completion among independent, comonotone, and history-fitted AR(1) with IQF tails.

Stronger dependence models (Q1). We added an empirical block copula, a D-vine copula for TimesFM-3 and Chronos-2. D-vine copulas are built from pair copulas, which can be Gaussian or non-Gaussian (Student-t, Clayton, Gumbel, Frank, Joe). We also tested the residual AR(1), a Toeplitz covariance and a nonstationary covariance, all of which did not beat the history AR(1). The table shows FPS at 32 stratified paths, with each completion at its better tail setting. Updated PathFM-300M scores 0.5458.

Completion (tails)Chronos-2 FPSTimesFM-3 FPSPathFM-300M minus TimesFM-3 (95% CI)
Independent (constant)0.47470.4918+0.0539 [0.0402, 0.0687]
Empirical block (constant)0.50660.5257+0.0201 [0.0077, 0.0329]
D-vine (IQF)0.52050.5358+0.0100 [−0.0025, 0.0234]
History AR(1) (IQF)0.52740.5446+0.0011 [−0.0100, 0.0129]

Both new classes improve on independent completion, but neither beats the history-fitted AR(1) copula, which remains the strongest baseline for both models. So they do not close the gap. PathFM is significantly ahead of the empirical block copula, and it leads the D-vine and the history AR(1), although not significantly.

Tails. We added the IQF tail extrapolation of Baron et al. (2025). For TimesFM-3, it raises the AR(1) completion from 0.5288 to 0.5373 FPS but lowers the independent completion from 0.4867 to 0.4623, so better tails help only together with a dependence model. The strongest baseline below uses IQF tails.

Uncertainty (Q3). The cross-model tables in the revision now carry paired bootstrap intervals. Here is the comparison with the strongest baseline, using the same sampling scheme and number of paths for both models on 94 tasks:

Sampling (both models)PathsUpdated PathFM-300M FPSTimesFM-3 + AR(1)/IQF FPSDifference95% CI
iid320.53410.5373−0.0031[−0.0176, 0.0107]
iid1280.54970.5458+0.0038[−0.0079, 0.0161]
iid5120.55230.5480+0.0043[−0.0071, 0.0162]
iid2,0000.55430.5486+0.0057[−0.0053, 0.0174]
Stratified320.54580.5446+0.0011[−0.0100, 0.0129]
Stratified5120.55430.5486+0.0057[−0.0054, 0.0173]

PathFM has the higher average score in every case except iid sampling at 32 paths. None of these differences is significant. We have therefore refined our claim: PathFM has the highest average FPS, but its lead over this baseline is not statistically significant.

What a path model adds. We want to credit the AR(1) completion, as it is the strongest completion we tested. However, its limitation is structural. Any method that turns marginal quantiles into paths must impose a dependence model from outside the forecaster. A path model, like PathFM, assumes no such form.

As evidence, we evaluated the path score under different context lengths. This matters because the AR(1) copula is fitted on the context. The advantage over the strongest baseline is largest at the shortest context. This is expected, because a history-fitted completion needs a long, representative history and a lag-one structure that actually holds, while PathFM relies on neither.

Context limitUpdated PathFM-300M FPSTimesFM-3 + AR(1)/IQF FPSΔFPSTask 95% CI
640.43430.4028+0.0315[0.0127, 0.0537]
2560.50140.4895+0.0120[−0.0015, 0.0262]
10240.54170.5288+0.0129[0.0017, 0.0248]
81920.55240.5451+0.0073[−0.0032, 0.0186]

Reply to W2 and Q2 (chaotic systems)

Training data. The chaotic models are trained on a fixed, stored collection, GilpinLab/dataset-panda72M-repro, with 190,006 trajectories; no trajectories are generated on the fly. It is split into 156,784 training and 33,222 validation trajectories. PathFM receives 2m flow-matching updates and 100k lr-decay updates, and the distilled student gets another 100k steps. The point-forecast test set has 9,347 separate trajectories, and none of them are used in training or validation.

AuthorsOpenReview ↗

Response to Reviewer hBUW (2/2)

Reply to W2 and Q2 (chaotic systems, cont'd)

Baselines. All baseline numbers come from forecasts we generated ourselves under one evaluation protocol; none are copied from prior work. PANDA is the released 72M checkpoint without further training. Chronos-1 Base is finetuned with its token cross-entropy loss (context 512, horizon 64, learning rate 1e-4) for 7,000 updates (chosen on the validation set). Chronos-2 starts from the released checkpoint and is fully fine-tuned with its quantile loss on the 156,784 training trajectories (context 512, horizon 128, learning rate 1e-5, 40,000 updates). The baselines therefore saw at least as much of the collection as PathFM.

The fast model. The main chaotic table uses the ten-step sampler, but the two-step student is already in the paper (Table 14 and Figure 12). As you note, it is less accurate than the ten-step model. Yet it is still ahead of PANDA, Chronos-1 and Chronos-2 on every metric.

ModelsMAPE H=64H=128H=256H=512W1PSD H²CD errorSegment analog
Chronos-1 FT24.8031.8639.6947.240.4110.1480.75016.40
Chronos-2 FT21.0726.7833.2840.079.6860.2690.56911.12
PANDA19.0924.4231.9441.220.5880.1950.53710.03
PathFM-126M, ten steps16.4521.5227.4334.500.2510.0610.2734.80
PathFM-126M, two steps16.6921.8327.8634.970.3160.0660.3266.56

We have included both 10-step and 2-step model entries in Table 2 in the main text to be explicit about the tradeoff.

Stochastic and deterministic baselines. We agree that part of the gap to PANDA comes from the deficit of deterministic models. That's why we include Chronos-1 FT, a stochastic baseline, in our comparison. Over all 135 systems, the two-step student lowers the four long-run errors by 23%, 55%, 57% and 60% relative to Chronos-1 FT, and all four differences are significant. We have added this point to the revised manuscript.

Reply to W2 (remaining coverage gap)

We agree that the fair-CRPS refinement improves calibration but does not reach nominal coverage. We have measured empirical coverage for PathFM-300M on fev-bench:

Paths(iid)80%90%
3272.29%82.00%
51276.47%86.30%

Part of the gap comes from the finite ensemble: going from 32 to 512 paths cuts the gap from ~8 points to ~3.5. The rest remains even at 512 paths, so the model is still somewhat under-dispersed. We have therefore revised the wording to say that the CRPS refinement improves calibration.

Reply to Q4 (path-level losses)

Yes, we tried them. Starting from the same distilled 126M model, we backpropagated an energy score, and separately a CRPS on the horizon mean and maximum, through the one-step sampler, alongside marginal CRPS and the distillation loss. Loss weights and gradient clipping were tuned over twenty validation configurations.

Refinement objectiveFPS
Marginal CRPS0.5252
Marginal CRPS + energy score0.5253
Marginal CRPS + mean/max CRPS0.5221

The energy score gives no measurable gain, and the mean/max loss significantly lowers FPS, although it does improve the Max-CRPS component on its own. We therefore keep marginal CRPS.

Reply to Q5 (horizons beyond 480 steps)

PathFM. For horizons beyond 480 steps, PathFM chains blocks autoregressively. The quantiles are calculated post-hoc.

Baselines. Each baseline is run through its released interface, with its own native block length and its own procedure for longer horizons. We follow the protocols used in existing benchmarks.

Scope. Chaining is used for 23 GIFT-Eval configurations and six TIME tasks, with maximum horizons of 900 and 864. Every fev-bench horizon is at most 288 steps, so no fps-bench result involves chaining.

Reply to Q6 (one-step baselines such as TARFVAE)

Thank you for the pointer; we have cited TARFVAE (Wei et al., 2025) in the related work. Its experiments train one model per dataset, while PathFM is a single pretrained model applied without task-specific training, so the two address different settings. We know of no fundamental reason that TARFVAE could not be scaled up. We chose flow matching because its training objective is a simple regression that is known to scale, with no variational bound to balance during training. We hope future work will explore more architectures for generative time-series foundation models.

We also ran a controlled comparison on ETTh1, training both TARFVAE and a flow model from scratch with the same data exposure. We provide the mean of the six path-loss ratios of the flow model divided by TARFVAE's; values below 1 favor the distilled flow model. From the table, the two are level at horizon 96, and the distilled flow model is significantly better at other horizons, especially long horizons. We have included this comparison in the revised manuscript.

HorizonPath loss ratio (PathFM / TARFVAE; <1 favors PathFM)
961.020
1920.911
3360.810
7200.855

Thank you again for your review!

Thank you for the detailed response and the substantial revision. I appreciate that many concerns have been addressed: the data provenance and baseline protocols are now clear; the strong completions, IQF tails, path-level loss experiments, and TARFVAE comparison are all useful additions. I also appreciate the honest acknowledgment that PathFM's remaining path advantage is not statistically significant.

A few issues still need to be fixed in the main text:

  • Calibration wording. The contribution bullet and Figure 5 still describe the model as “calibrated,” while coverage at 512 paths is 86.3% versus the nominal 90%. Please state the residual gap and soften the claim consistently.
  • Presentation. Confidence intervals should appear in the main result or figure with the correct table reference. The chaotic section should clearly distinguish the one-step product used in the main benchmarks from the two-step student used there. The context-length result is setting-dependent and should not be presented as rescuing the headline result.

With these corrections in the final version, I am willing to raise my score to 3. I hope the authors will improve the main text accordingly.

Thank you very much for your constructive follow-up and for your willingness to raise your score. We appreciate your careful reading, and we have revised the manuscript as you suggested:

  1. Calibration. We have removed the “calibrated” wording from the contribution bullet (Line 108), Figure 2, Figure 5 and other places (Line 232, 261, 868, 888, 958). Section 5.2 now refers to Appendix C.2 about the coverage gap (Line 355).
  2. Uncertainty. The main FPS claim (Line 351), Figure 3, Table 12 now include a significance footnote, pointing to Table 9 for the paired 95% confidence interval.
  3. Misc. Section 6 now clearly identifies the ten-step teacher and the two-step student (Line 437, 455, 458, 466, 470, 502). The main text now presents the context-length experiment as a separate comparison (Line 361).

Thank you again for your help in improving the clarity and rigor of the paper.

Reviewer_FaG6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PathFM models complete multivariate future trajectories using conditional flow matching. Its transformer shares history computation across draws; parallel decoding distillation reduces sampling to one or two evaluations, and fair-CRPS refinement improves marginal predictions. The paper introduces fps-bench, combining six temporal-path scores on 94 forecasting tasks, and reports marginal forecasting, cross-variable, latency, and chaotic-dynamics evaluations.

Critical strengths
  1. The paper identifies a genuine mismatch between marginal forecasting evaluation and path-dependent applications. Its benchmark evaluates temporal changes, totals, maxima, and crossing times rather than relying on marginal accuracy to imply joint-distribution quality. The reported gains are not confined to a single arbitrary aggregate: Table 10 shows advantages on all six components against the listed fixed-completion/native-sampling baselines, while retaining competitive marginal performance.

  2. The engineering and evaluation are unusually complete. Distillation, history packing, calibration, and ensemble construction are supported by ablations. The 126M model returns 32 paths in 51.6 ms in the stated L4 single-history setting. The authors also include native generative baselines, fitted temporal copulas, and smaller-model chaotic-system experiments. These controls strengthen the case that the contribution is a practical trajectory forecaster rather than merely a comparison against unsuitable marginal outputs.

Critical weaknesses
  1. The training-data description is insufficient to assess generalization and reproduce the benchmark claims. Appendix B.2 describes an 83M-row mixture of real and synthetic series but does not identify its constituent datasets, cutoffs, synthetic-generation procedures, or overlap checks against the evaluation suites. Retaining benchmark leakage-imputation rules is not a substitute for identifying which PathFM training sources overlap each benchmark. The chaotic-system section likewise needs an explicit training/evaluation trajectory split. This is a material uncertainty, not an allegation of contamination.

  2. Some headline interpretations exceed what the evaluation most directly establishes. The six aggregate FPS components score each channel's temporal path; cross-variable quality is evaluated separately. In Table 11, PathFM's cross-variable variogram score is 0.1443 versus 0.1446 for independent TimesFM-3 and 0.1420 for task-calibrated TimesFM-3. The fitted AR(1) copula also narrows temporal FPS from 0.4942 to 0.5311, compared with PathFM's 0.5446. Finally, Appendix D.1.1 acknowledges that the fair-score unbiasedness argument does not apply to the stratified main comparison, and that the finite-sample variogram estimator has model-dependent bias. A common iid sensitivity comparison and paired uncertainty against the strongest copula baseline would make the remaining advantage more convincing. These qualifications do not erase the useful temporal-path results.

Questions for authors
  1. Please provide the training-source manifest, date cutoffs, train/validation separation, and benchmark-specific overlap exclusions, including how data used for post-training were handled. For chaotic systems, which trajectories, initial conditions, and systems were available during training?

  2. Does the FPS ranking persist with iid paths for all samplers, several ensemble sizes, and paired task-bootstrap uncertainty against TimesFM-3 with the fitted AR(1) copula? Please distinguish finite-ensemble product quality from estimates of the population predictive law.

  3. What evidence supports a practically meaningful improvement in cross-variable dependence beyond the temporal-path gains? Please report uncertainty for Table 11 and clarify whether any further joint cross-channel metric was evaluated.

3 comments
AuthorsOpenReview ↗

Response to Reviewer FaG6 (1/2)

Thank you for the thoughtful review and for recognizing the completeness of our engineering and evaluation. We respond to each question below.

We have thoroughly revised the draft, with significant changes highlighted in blue. We extended the pretraining from 1M to 2M steps with the same data and architecture, denoted as "updated PathFM".

Reply to Q1 (training sources)

General models. The general models are trained on an 83-million-row corpus of real and synthetic time series. The real part is drawn from public univariate datasets: energy (Electricity, Solar, Wind Farms, London Smart Meters, Buildings 900K), transport (Mexico City Bikes, Pedestrian Counts, Taxi, Uber TLC, LargeST, Q-Traffic), nature (KDD Cup 2018, USHCN, WeatherBench, Temperature-Rain), cloud operations (Alibaba Cluster Trace 2018, Azure VM Traces 2017, Borg Cluster Data 2011), web (Wiki), and the M4 daily, hourly, weekly and monthly sets. The synthetic part has two kinds of data. Univariate series come from a generator that combines random trend, seasonality, and irregular components, and from one that samples random temporal causal graphs and generates series autoregressively. Multivariate series are built from univariate base generators (AR, ETS, the trend–seasonality generator, and KernelSynth) by adding dependence either at the same time step or across time, such as lead–lag relationships and cointegration.

The corpus was built to exclude all fev-bench datasets and the test portions of every GIFT-Eval task; it does overlap with the training portions of some GIFT-Eval datasets. Exclusions are by dataset and by test period. All training stages use the same corpus: 2,000,000 flow-matching updates, then 100,000 terminal-decay, 100,000 distillation and 100,000 CRPS-refinement updates.

Chaotic models. These are trained separately on a fixed, stored collection, GilpinLab/dataset-panda72M-repro, with 190,006 trajectories. Reserving the trajectories that involve ten parent systems gives 156,784 training and 33,222 validation trajectories. The point-forecast test set has 9,347 separate trajectories (56,082 windows), and none of its source files appear in the training collection.

The long-run evaluation covers 135 systems. Among these, 105 also occur in the training data and 30 do not occur in training. This should not be confused with trajectories: no training trajectory appears in the test set. The split is also fair to the baselines (PANDA, Chronos-1 and Chronos-2 fine-tune).

The table below reports the long-run errors for the two groups separately. Each cell shows seen / unseen systems; lower is better. PathFM has the lowest error on all four diagnostics in both groups, so its advantage holds whether seeing a system during training or not.

ModelSliced W1PSD H²CD errorSegment analog
Chronos-1 FT0.434 / 0.3320.150 / 0.1400.723 / 0.84816.68 / 15.44
Chronos-2 FT12.135 / 1.1120.269 / 0.2690.596 / 0.47711.22 / 10.76
PANDA0.585 / 0.6020.191 / 0.2080.560 / 0.4559.85 / 10.67
PathFM-126M0.237 / 0.2990.058 / 0.0710.280 / 0.2504.44 / 6.31
AuthorsOpenReview ↗

Response to Reviewer FaG6 (2/2)

Reply to Q2 (common iid sampling, ensemble sizes, and paired uncertainty)

We ran the requested comparison against TimesFM-3 with a history-fitted AR(1) copula and IQF tails on its frozen quantiles. Both models use the same sampling scheme and the same number of paths on FPS.

SamplingPathsPathFM-300M FPSTimesFM-3 + AR(1)/IQF FPSDifference95% CI
iid320.53410.5373−0.0031[−0.0176, 0.0107]
iid1280.54970.5458+0.0038[−0.0079, 0.0161]
iid5120.55230.5480+0.0043[−0.0071, 0.0162]
iid2,0000.55430.5486+0.0057[−0.0053, 0.0174]
Stratified320.54580.5446+0.0011[−0.0100, 0.0129]
Stratified5120.55430.5486+0.0057[−0.0054, 0.0173]

PathFM has the higher average score in every case except iid sampling at 32 paths: it leads with iid sampling from 128 paths upward, and with stratified sampling. None of these differences is significant. We therefore refine our claim: PathFM has the highest average FPS, but its lead over this baseline is not statistically significant. We have updated the paper to make this scope explicit.

This also answers your question on distinguishing finite-ensemble product quality from estimates of the population predictive law. The population law cannot be evaluated directly, because that would require infinitely many samples. Instead, we can observe how the score converges as the number of iid paths increases. From 32 to 2,000 paths, PathFM rises from 0.5341 to 0.5543, and the baseline from 0.5373 to 0.5486. We then extrapolate by fitting a − b/m, where m is the number of paths. The fitted limit is about 0.554 for PathFM and about 0.549 for TimesFM-3 with AR(1)/IQF. These limits approximate the score under the population predictive law.

Reply to Q3 (cross-variable dependence and Table 11 uncertainty)

fps-bench focuses on temporal dependence by design. Its six scores evaluate each channel's path over time, which lets us compare PathFM with univariate forecasters on the same benchmark. We also report cross-variable results for completeness, but they are not part of our main contribution. We are happy to test additional metrics if the reviewer specifies them.

We reran Table 11 with the updated models under the same protocol (35 multivariate tasks, 32 iid paths). CrossVS (Cross-variable Variogram Score) compares pairs of channels at the same time step; it does not measure temporal dependence. Lower is better.

MethodsCrossVS
Chronos-2, independent/constant0.1496
Chronos-2, joint history-AR(1)/IQF0.1461
TimesFM-3, independent/constant0.1452
TimesFM-3, joint history-AR(1)/IQF0.1415
Sundial, native0.1632
Toto-1.0, native0.2008
Updated PathFM 126M0.1411
Updated PathFM 300M0.1407

We also provide the uncertainty of PathFM's advantage against strong baselines:

PathFM 300M compared withReduction95% CI
Chronos-2, independent/constant5.91%[3.62%, 11.03%]
Chronos-2, joint history-AR(1)/IQF3.67%[0.34%, 9.90%]
TimesFM-3, independent/constant3.06%[1.00%, 7.13%]
TimesFM-3, joint history-AR(1)/IQF0.53%[−2.21%, 5.50%]

PathFM has the lowest average CrossVS and is significantly better than the independent completions of both models. Against the fitted joint copula, it is ahead of Chronos-2 but not significantly different from TimesFM-3, and the same holds for TimesFM-3 when both models use stratified sampling. PathFM is also significantly better than the native generative baselines, Sundial and Toto-1.0. We have updated the paper to make this scope explicit.

Thank you again for your review!

Dear Reviewer FaG6,

Thank you again for your review. In the revision, we clarified the training data for general and chaotic systems, added common-iid comparisons across ensemble sizes with confidence intervals against TimesFM-3 with AR(1)/IQF, and reported uncertainty for the cross-variable results. We would be glad to clarify anything further before the discussion period ends.

Best regards, Authors

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

PathFM generates joint future trajectories using conditional flow matching, with shared history computation and parallel decoding distillation for efficient sampling. Fair-CRPS refinement improves marginal forecasting, and fps-bench evaluates trajectory quality using six path scores. Experiments cover marginal benchmarks, path-dependent quantities, inference cost, and chaotic dynamics.

Critical strengths

1.The paper clearly motivates joint trajectory forecasting through the dependence information that marginal forecasts leave unspecified. 2.Joint generation, shared computation, and distillation enable efficient trajectory sampling while retaining competitive marginal accuracy. 3.The evaluation combines path scores, fitted-copula comparisons, and long-run dynamical diagnostics to assess trajectory quality. 4.The visualizations clearly explain the motivation and illustrate the behavior of generated trajectories and their differences from baseline forecasts.

Critical weaknesses

1.PathFM's lead on the authors' fps-bench requires sensitivity analysis of metric aggregation and dependence completions to assess whether the benchmark design favors the method, particularly given its lower marginal accuracy and the smaller gap after fitting copulas. The method combines established techniques, and the PathFM's lead on the authors' fps-bench requires sensitivity analysis of metric aggregation and dependence completions to assess whether the benchmark design favors the method, particularly given its lower marginal accuracy and the smaller gap after fitting copulas. The method combines established techniques, and the cumulative ablations leave the contribution of their integration insufficiently isolated. 2.The submission lacks an anonymous code repository for independent verification of the sampling, scoring, and baseline-completion protocols. cumulative ablations leave the contribution of their integration insufficiently isolated. 3.The submission lacks an anonymous code repository for independent verification of the sampling, scoring, and baseline-completion protocols.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer SXUS (1/2)

Thank you for the careful review, and for raising the questions of benchmark robustness, the contribution of the integrated design, and reproducibility. We respond to each point below.

We have thoroughly revised the draft, with significant changes highlighted in blue. We extended the pretraining from 1M to 2M steps with the same data and architecture, denoted as "updated PathFM".

Reply to W1 (does the benchmark design favor PathFM?)

Thank you for the careful reading. We did not select scores or tasks to favor PathFM. fps-bench uses every fev-bench task on which all six scores can be computed (94 of 100; there are 6 tasks excluded because of NaN). The path scores are established and widely used: the energy score [1, 2], and the variogram score [3]. The three functionals were chosen to represent common path-dependent tasks: the horizon sum (total demand) [4], the maximum (peak load) [5], and the first-passage time of a threshold (time to failure) [6]. These are the only three we considered, and we are happy to evaluate any further functional you suggest.

Thank you also for pointing out that the gap narrows once quantile forecasters are given a fitted copula. We agree, and we now make this explicit. We now evaluate marginal baselines with the best of independent, comonotone, and AR(1)/IQF. We followed the geometric mean aggregation from fev-bench. We report 95% task-bootstrap intervals for each model in Table 12 (Appendix E.1), and paired intervals for matched-sampling comparisons in Table 9 (Appendix C.3).

ForecasterCompletionESVSDESSumMaxFPTFPS
Chronos-2AR(1)/IQF0.49280.60770.44570.53820.47240.58560.5274
TimesFM-3AR(1)/IQF0.50610.61710.45740.55860.48600.61710.5446
Toto-2.0 (2.5B)AR(1)/IQF0.45690.56390.43030.50220.42420.58810.4984
Updated PathFM 126MNative0.49170.62440.45570.53830.51030.58960.5386
Updated PathFM 300MNative0.49950.63050.45950.54760.52340.59240.5458

Our method leads the aggregate score against the strongest completion (TimesFM-3 with history AR(1) and IQF tails) at 32 stratified paths, with a positive but insignificant difference of +0.0011 (95% CI [−0.0100, 0.0129]). The revised paper makes this scope explicit in Section 5.2 and Appendix E.1.

We also want to credit the AR(1) completion, as it is the strongest completion we tested, as used in [7]. However, its limitation is structural. Any method that turns marginal quantiles into paths must impose a dependence model from outside the forecaster; here, a stationary Gaussian copula governed by a lag-one coefficient estimated from the history. A path model, like PathFM, assumes no such form.

To test PathFM's structural advantage directly, we took PathFM's own samples, converted them into marginals, and fitted them into paths with AR(1). Because both versions share identical marginals, any difference comes from the dependence alone. PathFM's native paths score higher than the AR(1) re-coupled ones, and the improvement is significant on all six components.

Path construction (same marginal samples)FPS95% CI
Native PathFM pairing0.5524/
Re-coupled with history AR(1)0.5321[0.0145, 0.0274]
Re-coupled independently0.4668[0.0690, 0.1047]

As further evidence, we evaluated the path score under different context lengths. This matters because the AR(1) copula is fitted on the context. The advantage over the strongest baseline is largest at the shortest context. This is expected, because a history-fitted completion needs a long, representative history and a lag-one structure that actually holds, while PathFM relies on neither.

Context limitUpdated PathFM-300M FPSTimesFM-3 + AR(1)/IQF FPSΔFPS95% CI
640.43430.4028+0.0315[0.0127, 0.0537]
2560.50140.4895+0.0120[−0.0015, 0.0262]
10240.54170.5288+0.0129[0.0017, 0.0248]
81920.55240.5451+0.0073[−0.0032, 0.0186]

References

  1. Gneiting & Raftery (2007). Journal of the American Statistical Association, 102(477), 359–378.
  2. Gneiting et al. (2008). TEST, 17(2), 211–235.
  3. Scheuerer & Hamill (2015). Monthly Weather Review, 143(4), 1321–1334.
  4. Boylan & Babai (2022). International Journal of Production Economics, 252, 108586.
  5. Hyndman & Fan (2010). IEEE Transactions on Power Systems, 25(2), 1142–1153.
  6. Si et al. (2011). European Journal of Operational Research, 213(1), 1–14.
  7. Baron et al. (2025). arXiv:2510.02224.
  8. Pic et al. (2025). Advances in Statistical Climatology, Meteorology and Oceanography, 11(1), 23.
AuthorsOpenReview ↗

Response to Reviewer SXUS (2/2)

Reply to W1, continued (contribution of the integration)

We agree that many building blocks come from prior work, but each needed changes to work for forecasting, and they depend on each other. History-only attention with late entry lets the history be computed once for all sampled paths, which makes drawing many paths cheap. Parallel decoding distillation, applied to joint paths, reduces sampling to one step, and that one-step sampler enables backpropagating fair CRPS through the generated forecasts. The resulting four-stage pipeline performs well: updated PathFM-300M ranks second on fev-bench and third on TIME for marginal accuracy, has the highest average path score on fps-bench, and improves on previous generative foundation models by a wide margin. We release the code and the model weights (see the next answer), and we hope the pipeline serves as a starting point for later work on generative forecasting models.

Beyond the model, we ran a range of studies to help understand generative time series foundation models. In the paper: the teacher–student comparison at different step budgets, the ensemble-size and sampling study, the inference speed-up breakdown, and the many chaotic-systems experiments. In the rebuttal: a fixed-marginal experiment that isolates the learned dependence, a context-length study, matched iid and stratified comparisons from 32 to 2,000 paths, stronger copula baselines, path-level training losses, and a comparison with TARFVAE.

Our further contributions are: The paper draws attention to a problem that existing benchmarks miss: the joint distribution of forecast paths, beyond their marginals. It provides fps-bench to measure the gap. And we show on chaotic systems that a general forecasting architecture can outperform a specialized model while preserving long-run variability.

Reply to W2/3 (anonymous code repository)

Thank you for raising this. We now provide the code and weights at https://github.com/pathfm/pathfm-supplementary-submission. The code is also uploaded to the supplementary material of the revised submission. You are welcome to try it out.

Thank you again for your review!

Dear Reviewer SXUS,

Thank you again for your review. In the revision, we added aggregation and dependence-completion sensitivity analyses (including IQF tails and stronger copulas), ablations on native dependence vs. refitting and restricted context, and an anonymous release of code and weights. We would be glad to clarify anything further before the discussion period ends.

Best regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

Dear Area Chair and Reviewers,

Thank you for the thoughtful feedback and discussion. We appreciate the reviewers’ recognition that fps-bench addresses an important gap in evaluating forecast paths beyond marginal accuracy (SXUS, FaG6, hBUW, NtmG). Reviewers also highlighted fast trajectory sampling with competitive marginal accuracy (SXUS, FaG6, NtmG), and the thorough evaluation, ablations, and supporting evidence from chaotic dynamics (FaG6, hBUW).

We have revised the manuscript to incorporate the additional experiments, clarify the evaluation protocols, and specify the scope of our claims. Revisions are highlighted in blue. The main updates are:

  1. Stronger baselines and matched comparisons. Following requests from FaG6, hBUW, and NtmG, we tested richer copulas and IQF tails, matched sampling schemes and ensemble sizes, and added uncertainty estimates for temporal and cross-variable scores. In the main comparison, PathFM-300M has the highest mean FPS; its gap to TimesFM-3 with AR(1)/IQF is not statistically significant. Under restricted history length, the advantage is significant in the 64-step context, so we treat it as setting-dependent.
  2. Evidence for PathFM’s design. PathFM-126M significantly improves FPS over Chronos-2's strongest completion, which uses the shared corpus and similar model size (NtmG). PathFM’s native dependence significantly improves all six path scores over AR(1) reconstruction from PathFM’s marginals (SXUS, hBUW). The path-level loss experiments support retaining marginal CRPS (hBUW). In a matched-data comparison with TARFVAE, our model is comparable at horizon 96 and significantly better at the three longer horizons (hBUW).
  3. Chaotic-system evaluation. In response to FaG6 and hBUW, we clarified data separation and baseline training, and added results for 105 systems represented in training and 30 unrepresented systems. The ten-step model leads on all four diagnostics in both groups. The two-step student reduces long-run errors by 23%–60% over stochastic Chronos-1 FT, with all four improvements significant.
  4. Reproducibility and clarity. Addressing the reviewers’ requests, we released code and weights with documentation. Following NtmG’s feedback, we added an appendix overview, clarified experimental and design motivations, and revised the writing throughout.

Thank you again for your time!

Yours, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.