Pith. sign in

REVIEW 4 major objections 6 minor 21 references

Time Series Foundation Models are Flow Predictors

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that generic zero-shot time-series foundation models, given only the temporal evolution of origin–destination flows, outperform specialized spatiotemporal deep learning models on crowd-flow prediction across three…

desk verdict Plausible zero-shot TSFM result on flow prediction, but the evaluation protocol is so under-specified that the headline margins are not yet reproducible. read the letter →

arxiv 2507.00945 v1 pith:NSEE22FT submitted 2025-07-01 cs.LG cs.CY

Massimiliano Luca , Ciro Beneduce , Bruno Lepri This is my paper · ORCID
classification cs.LGcs.CY
keywords crowdflowpredictiontimeseriesfoundationmodelszero-shotforecastingorigin-destinationflowsMoiraiTimesFMhumanmobilityspatiotemporal
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that general-purpose time-series foundation models can serve as crowd-flow predictors without any training on the target city. The claim is that Moirai and TimesFM, used zero-shot and fed only the temporal history of each origin–destination flow, beat classical statistical models and specialized spatial deep networks on Bike NYC, Taxi Beijing, and Spanish nationwide origin–destination data. If the claim holds, accurate flow forecasting no longer requires annotated target-city data, a spatial graph, or a grid-based architecture; a pretrained forecaster can be pointed at the raw flow series directly. On the reported numbers the advantage is up to 33% lower RMSE, up to 39% lower MAE, and up to 49% higher common part of commuters (CPC) against the baselines.

What carries the argument

The central object is the per-origin time series reformulation (Definition 2.2): instead of modeling the OD tensor jointly, each origin tile $i$ yields a sequence of vectors $s^{(i)}_\tau = [T_{\tau,i,1}, \dots, T_{\tau,i,n}]$, and the task is to forecast $s^{(i)}_{t+1}$ from the history $s^{(i)}_1, \dots, s^{(i)}_t$. This reduces one spatiotemporal tensor problem to $n$ independent multivariate forecasting tasks, which is exactly the input format that pretrained transformer forecasters such as Moirai and TimesFM consume. The mechanism doing the work is thus the decomposition plus zero-shot transfer: no fine-tuning, no spatial adjacency, and no convolution over a grid.

What would settle it

Retrain CrowdNet and MSAGGN on the same Bike NYC and Taxi BJ splits with a documented hyperparameter search and repeated runs, then compare their best RMSE, MAE, and CPC to Moirai-L's values of 6.09/3.04/0.72 and 9.32/7.34/0.62; if the tuned baselines match or beat those numbers, the paper's central claim of zero-shot TSFM superiority on these benchmarks would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that per-location decomposition turns crowd-flow prediction into a set of independent multivariate time-series tasks and that pretrained time-series foundation models solve those tasks better than the spatial baselines. For each origin tile the outgoing flows to all destinations form one series; Moirai and TimesFM forecast the next vector from that series alone. In Table 1, Moirai-L achieves the best score on every dataset and metric—RMSE 6.09, MAE 3.04, CPC 0.72 on Bike NYC; RMSE 9.32, MAE 7.34, CPC 0.62 on Taxi BJ; RMSE 21.74, MAE 9.94, CPC 0.67 on Spain OD—while TimesFM lands close behind. Because the deep-learning baselines have explicit spatial structure and the foundation models do not, the authors interpret this as evidence that the temporal regularities in origin–destination flows carry most of the predictive signal, and that zero-shot transfer to new cities and scales is feasible.

Load-bearing premise

The load-bearing premise is that the baseline deep-learning models were trained well enough to represent their intended performance; the paper reports no hyperparameters, validation procedure, or number of runs for ST-ResNet, DMVSTNet, ACMF, CrowdNet, or MSAGGN.

Editorial extensions

If this is right

  • A city or region with no labeled flow history and no spatial map can still get flow forecasts from a pretrained model, which lowers the data barrier for mobility analysis.
  • Decomposing OD tensors per origin makes prediction scalable: adding a region adds one more independent time series instead of requiring a retrained joint spatiotemporal model.
  • Because the winners ignore spatial adjacency, spatial structure appears less essential for these benchmarks than the surveyed deep-learning literature assumed.
  • The same zero-shot recipe becomes a natural baseline for future crowd-flow studies, and a candidate for other per-pair mobility prediction tasks such as demand and next-location forecasting.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A consequence left implicit is that if the temporal signal dominates these flow series, adding spatial context to Moirai or TimesFM as extra input channels should give only small gains, and the size of that gain is directly measurable.
  • The comparison may partly reflect baseline tuning rather than TSFM superiority, so retraining the deep baselines with a documented hyperparameter search is the decisive check of the claimed advantage.
  • At finer temporal or spatial granularities the OD series become sparser and noisier, so the zero-shot advantage should be retested before extending the conclusion to real-time or street-level prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes a reformulation of crowd flow prediction as per-origin multivariate time series forecasting and applies two pretrained time series foundation models, Moirai and TimesFM, in a strict zero-shot setting. On three public datasets (Bike NYC, Taxi Beijing, Spanish national OD flows), the authors report that the TSFMs outperform statistical and deep learning baselines across RMSE, MAE, and CPC, despite not using explicit spatial information. The paper argues that TSFMs offer a scalable, generalizable alternative to specialized spatio-temporal deep learning models.

Significance. The finding is potentially significant: if confirmed, it would show that general-purpose TSFMs can outperform task-specific deep spatio-temporal models in flow prediction, which has practical implications for data-scarce scenarios. The paper's strengths are its clear problem formulation and the use of real-world datasets. However, the current evidence is undermined by an underspecified evaluation protocol and internal numerical inconsistencies, so the significance is not yet firmly established.

major comments (4)
  1. [Section 2.4 and Abstract/Introduction] The input representation given to the TSFMs is ambiguous. Definition 2.2 defines a multivariate per-origin series s(i)_tau = [T_{tau,i,1}, ..., T_{tau,i,n}], but the Abstract and Section 1 state that the models receive 'only the temporal evolution of each OD flow' and 'individual time series.' The paper never states whether the TSFMs receive one univariate series per OD pair or an n-dimensional vector per origin. This distinction is critical because Moirai and TimesFM have different multivariate capabilities and normalization schemes, and because the per-origin vector contains destination information, which is arguably spatial structure. The evaluation protocol (context length, prediction horizon, instance normalization) is also not reported. Without these details, the 'strict zero-shot' scenario and the comparison to baselines are not reproducible as described.
  2. [Section 4 and Table 1] The percentage improvements reported in Section 4 do not match Table 1. For example, comparing Moirai-L to the best baseline MSAGGN on Bike NYC, the RMSE reduction is (8.02 - 6.09)/8.02 = 24.1%, not the reported -23.44%, and the MAE reduction is (3.59 - 3.04)/3.59 = 15.3%, not -23.01%. On Taxi BJ, the RMSE reduction is (14.12 - 9.32)/14.12 = 34.0%, not -32.86%, and the MAE reduction is (11.95 - 7.34)/11.95 = 38.6%, not -39.86%. On Spain OD, the RMSE reduction is (28.05 - 21.74)/28.05 = 22.5%, not -20.83%, and the MAE reduction is (13.26 - 9.94)/13.26 = 25.0%, not -21.33%. These discrepancies affect the abstract's headline numbers (e.g., 'up to 39% lower MAE') and must be corrected with a clear statement of the comparison basis.
  3. [Section 3.3] The baseline training setup is severely underspecified. For ST-ResNet, DMVSTNet, ACMF, CrowdNet, and MSAGGN, the paper reports no hyperparameters, validation procedure, or number of runs; it merely says the data are split 'as proposed in [7].' Without this information, under-training or mis-tuning of the baselines cannot be ruled out, which would invalidate the claim that TSFMs outperform 'state-of-the-art competitors.' The authors should provide exact configurations or cite the original implementations, and they should report error bars or significance tests, as Table 1 appears to contain single-run results.
  4. [Section 3.3 (Taxi Beijing)] The Taxi Beijing dataset is one week of data (168 hourly points), which is shorter than the typical context length of TSFMs (e.g., TimesFM uses a default context of 512). The paper does not state how the models are applied to this short series—whether the data are aggregated to a daily resolution, whether a shorter context is used, or whether the evaluation uses an expanding window that includes test observations. Without this information, the validity of the zero-shot evaluation on this dataset is unclear, and the reported gains could be an artifact of the chosen protocol.
minor comments (6)
  1. [Title page] The CCS Concepts and Keywords sections still contain template placeholders ('Do Not Use This Code' and 'Do, Not, Us, This, Code') and must be replaced with correct terms.
  2. [Throughout] The model name is written inconsistently as 'TimeFM' in Section 4 and 'TimesFM' elsewhere; please use one spelling consistently.
  3. [Section 1] There is a typo: 'Common Parts of Communters' should be 'Common Part of Commuters'.
  4. [Figure 1] The caption and the figure itself are not fully described in the text; the reader is not told what the axes or colors represent, and the figure appears to be a schematic with insufficient explanation.
  5. [Section 3.2] The manuscript does not specify which exact versions/checkpoints of Moirai and TimesFM are used (e.g., Moirai-Large, TimesFM-2.0), nor the patch size and forecast horizon settings. These details are necessary for reproducibility.
  6. [End matter] There is no data availability or code availability statement. Given the reproducibility claims, at least the configuration scripts and evaluation code should be made available.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular step found: the central claim is an externally benchmarked zero-shot comparison, not a fit or self-citation chain.

full rationale

The paper's central claim is an empirical benchmark claim, not a derivation. Section 2.4 defines per-origin forecasting with input series s_tau^(i) = [T_tau,i,1, ..., T_tau,i,n]; Section 3.2 specifies Moirai and TimesFM as externally pretrained checkpoints; Section 3.4 defines RMSE, MAE, and CPC; Section 4 reports Table 1. None of these equations defines the TSFM predictions in terms of the reported metrics or the test labels, and no fitted parameter is renamed as a prediction. The dataset split is borrowed from [7], whose authors overlap with the current paper, and CrowdNet [7] is a published baseline, but adopting an external benchmark protocol is not a circular derivation: the reported improvement over CrowdNet is not implied by adopting CrowdNet's split. The other self-citations ([2]-[4], [12]) concern tooling, LLM next-location prediction, and a survey; they do not justify the central result. The internal inconsistencies in the Section 4 percentage summary and the underspecified TSFM context/horizon/normalization are validity and reporting concerns, not circularity. Hence no circular step can be exhibited with a specific reduction, and the score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper fits no parameters and introduces no entities. Its load-bearing assumptions are transferability of pretrained TSFMs, fair baseline training, and sufficiency of the per-origin decomposition.

assumptions (3)
  • domain assumption Pretrained TSFM checkpoints transfer to mobility OD flow distributions without fine-tuning.
    Invoked in Sections 3.2 and 4; no analysis of distribution shift between LOTSA and Google pretraining corpora and bike, taxi, or national mobility data.
  • domain assumption Baseline deep learning models are trained and tuned as in prior work [7], on a comparable split.
    Section 3.3 delegates the train/test protocol to [7] but no hyperparameters, runs, or validation are reported; if mis-tuned, the comparison is unfair.
  • ad hoc to paper Decomposing the OD tensor into per-origin multivariate time series loses no information critical for prediction.
    Definition 2.2 removes inter-origin and explicit spatial dependencies; the paper then treats the result as a strength because TSFMs still win, but this is the central design assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Time Series Foundation Models are Flow Predictors." pith.science (2026). https://pith.science/paper/NSEE22FT

@misc{pith2026250700945,
  author       = {Pith},
  title        = {Pith review of: Time Series Foundation Models are Flow Predictors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NSEE22FT}},
  note         = {Machine review of arXiv:2507.00945}
}
read the original abstract

We investigate the effectiveness of time series foundation models (TSFMs) for crowd flow prediction, focusing on Moirai and TimesFM. Evaluated on three real-world mobility datasets-Bike NYC, Taxi Beijing, and Spanish national OD flows-these models are deployed in a strict zero-shot setting, using only the temporal evolution of each OD flow and no explicit spatial information. Moirai and TimesFM outperform both statistical and deep learning baselines, achieving up to 33% lower RMSE, 39% lower MAE and up to 49% higher CPC compared to state-of-the-art competitors. Our results highlight the practical value of TSFMs for accurate, scalable flow prediction, even in scenarios with limited annotated data or missing spatial context.

Figures

Figures reproduced from arXiv: 2507.00945 by the authors.

Figure 1
Figure 1. A high-level depiction of the task and our approach. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages

  1. [7]

    Marco Cardia, Massimiliano Luca, and Luca Pappalardo. 2022. Enhancing crowd flow prediction in various spatial and temporal granularities. In Companion proceedings of the web conference 2022 . 1251–1259

    work page 2022

  2. [1]

    Laura Alessandretti. 2022. What human mobility data tell us about COVID-19 spread. Nature Reviews Physics 4, 1 (2022), 12–13

    work page 2022

  3. [2]

    Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. 2025. Large language models are zero-shot next location predictors. IEEE Access (2025)

    work page 2025

  4. [3]

    Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. 2025. Urban Safety Percep- tion Through the Lens of Large Multimodal Models: A Persona-based Approach. arXiv preprint arXiv:2503.00610 (2025)

    arXiv 2025

  5. [4]

    Ciro Beneduce, Tania Gullón Muñoz-Repiso, Bruno Lepri, and Massimiliano Luca

    work page

  6. [5]

    Luís MA Bettencourt, José Lobo, Dirk Helbing, Christian Kühnert, and Geoffrey B West. 2007. Growth, innovation, scaling, and the pace of life in cities. Proceedings of the national academy of sciences 104, 17 (2007), 7301–7306

    work page 2007

  7. [6]

    Juan Camilo Bohorquez, Sean Gourley, Alexander R Dixon, Michael Spagat, and Neil F Johnson. 2009. Common ecology quantifies human insurgency. Nature 462, 7275 (2009), 911–914

    work page 2009

  8. [8]

    Vittoria Colizza, Alain Barrat, Marc Barthelemy, Alain-Jacques Valleron, and Alessandro Vespignani. 2007. Modeling the worldwide spread of pandemic influenza: baseline case and containment interventions. PLoS medicine 4, 1 (2007), e13

    work page 2007

Show all 21 references
  1. [9]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2024. A decoder- only foundation model for time-series forecasting. In Forty-first International Conference on Machine Learning

    2024

  2. [10]

    Nicolò Alessandro Girardini, Arkadiusz Stopczynski, Olga Baranov, Cornelia Betsch, Dirk Brockmann, Sune Lehmann, and Robert Böhm. 2024. Using smart- phones to study vaccination decisions in the wild. PLOS digital health 3, 8 (2024), e0000550

    2024

  3. [11]

    Lingbo Liu, Ruimao Zhang, Jiefeng Peng, Guanbin Li, Bowen Du, and Liang Lin

  4. [12]

    Massimiliano Luca, Gianni Barlacchi, Bruno Lepri, and Luca Pappalardo. 2021. A survey on deep learning for human mobility. ACM Computing Surveys (CSUR) 55, 1 (2021), 1–44

    2021

  5. [13]

    Mattia Mazzoli, Alex Molas, Aleix Bassolas, Maxime Lenormand, Pere Colet, and José J Ramasco. 2019. Field theory for recurrent mobility. Nature communications 10, 1 (2019), 3895

    2019

  6. [14]

    Ministerio de Transportes y Movilidad Sostenible. 2023. Informe metodológico del Estudio de Movilidad con Big Data . Informe técnico. Ministerio de Transportes y Movilidad Sostenible, Madrid, España. https://www.transportes.gob.es/recursos_mfom/paginabasica/recursos/ a3_inform...

    2023

  7. [15]

    Zain Ul Abideen, Xiaodong Sun, and Chao Sun. 2025. Crowd flow prediction: an integrated approach using dynamic spatial–temporal adaptive modeling for pattern flow relationships. Journal of Forecasting 44, 2 (2025), 556–574

    2025

  8. [16]

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. 2024. Unified training of universal time series forecasting transformers. (2024)

    2024

  9. [17]

    Huaxiu Yao, Fei Wu, Jintao Ke, Xianfeng Tang, Yitian Jia, Siyu Lu, Pinghua Gong, Jieping Ye, and Zhenhui Li. 2018. Deep multi-view spatial-temporal network for taxi demand prediction. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  10. [18]

    Jing Yuan, Yu Zheng, Xing Xie, and Guangzhong Sun. 2011. T-drive: Enhancing driving directions with taxi drivers’ intelligence. IEEE Transactions on Knowledge and Data Engineering 25, 1 (2011), 220–232

    2011

  11. [19]

    Junbo Zhang, Yu Zheng, and Dekang Qi. 2017. Deep spatio-temporal residual net- works for citywide crowd flows prediction. In Proceedings of the AAAI conference on artificial intelligence, Vol. 31

    2017

  12. [2018]

    InProceedings of the 26th ACM international conference on Multimedia

    Attentive crowd flow machines. InProceedings of the 26th ACM international conference on Multimedia. 1553–1561

  13. [2025]

    arXiv:2506.13385 [cs.CY] https://arxiv.org/abs/2506.13385

    pySpainMobility: a Python Package to Access and Manage Spanish Open Mobility Data. arXiv:2506.13385 [cs.CY] https://arxiv.org/abs/2506.13385

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.