arxiv: 2604.08400 · v1 · submitted 2026-04-09 · 💻 cs.LG · cs.AI

Recognition: unknown

Zero-shot Multivariate Time Series Forecasting Using Tabular Prior Fitted Networks

Mayuka Jayawardhana , Nihal Sharma , Kazem Meidani , Bayan Bruss , Tom Goldstein , Doron Bergman

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:01 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords multivariate time series forecastingzero-shot forecastingtabular foundation modelsTabPFNscalar regressioninter-channel interactionstime series prediction

0 comments

The pith

Tabular foundation models can perform zero-shot multivariate time series forecasting by recasting the task as scalar regression problems.

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

The paper introduces a framework that converts the multivariate time series forecasting problem into a collection of scalar regression tasks. These tasks are then solved directly and without task-specific training by tabular foundation models such as TabPFN. The approach is presented as an advance over prior methods that handled each variable as an independent univariate series. A sympathetic reader would care because it offers a uniform way to apply large tabular models to forecasting while addressing interactions across channels.

Core claim

By recasting the multivariate time series forecasting problem as a series of scalar regression problems which can then be solved zero-shot by any tabular foundation model with regression capabilities, the method incorporates inter-channel interactions. Results are shown using the TabPFN-TS backbone and compared against current state-of-the-art tabular methods.

What carries the argument

The recasting of multivariate time series forecasting into scalar regression problems on tabular inputs, which allows prior-fitted networks to produce predictions that reflect cross-channel structure.

Load-bearing premise

That converting the forecasting task into scalar regressions on tabular data will let the model capture and exploit dependencies between different channels without any extra mechanisms.

What would settle it

A head-to-head test on a multivariate dataset with strong cross-channel correlations where the method shows no accuracy gain over independent univariate applications of the same tabular model.

Figures

Figures reproduced from arXiv: 2604.08400 by Bayan Bruss, Doron Bergman, Kazem Meidani, Mayuka Jayawardhana, Nihal Sharma, Tom Goldstein.

**Figure 1.** Figure 1: Reformulating multivariate time series forecasting as a scalar regression problem: A pictorial and tabular representation of a Lorenz system (a time series with three covariates x, y, z introduced in Lorenz (1963)) is presented on the left parts of the figure. The time-indexed table is “rolled out” by introducing a channel indicator column (as seen in the table on the right). We note that this transformati… view at source ↗

**Figure 2.** Figure 2: Performance of TabPFN-TS compared to ours on multivariate datasets in the gifteval benchmark our approach lowers MASE on 60% of the datasets. We refer the reader to [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Performance of TabPFN-TS-MV compared to other SOTA time series approaches: The y-axis is log scale, Tabpfn-TS-MV (ours) and Chronos 2 supports channel dependent (CD) multivariate predictions while TabPFN-TS and TempoPFN only support univariate (channel independent) predictions [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Performance of TabPFN-TS-MV compared to TimePFN: While TimePFN supports joint multivariate time series forecasting it only supports a maximum context window of 96 time steps. This poses a significant limitation as a majority of the datasets in Gift-Eval consist of much longer time series. All approaches are limited to a context window of 96 time steps (per variate) for fair comparison 9 [PITH_FULL_IMAGE:f… view at source ↗

**Figure 5.** Figure 5: Performance comparison between different strategies for mitigating cross-channel distributional shifts. The y-axis is in log scale. We observe that while the standardizing approach greatly improves performance, first order differencing degrades performance. We incorporate the standardization approach into our framework. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

read the original abstract

Tabular foundation models, particularly Prior-data Fitted Networks like TabPFN have emerged as the leading contender in a myriad of tasks ranging from data imputation to label prediction on the tabular data format surpassing the historical successes of tree-based models. This has led to investigations on their applicability to forecasting time series data which can be formulated as a tabular problem. While recent work to this end has displayed positive results, most works have limited their treatment of multivariate time series problems to several independent univariate time series forecasting subproblems, thus ignoring any inter-channel interactions. Overcoming this limitation, we introduce a generally applicable framework for multivariate time series forecasting using tabular foundation models. We achieve this by recasting the multivariate time series forecasting problem as a series of scalar regression problems which can then be solved zero-shot by any tabular foundation model with regression capabilities. We present results of our method using the TabPFN-TS backbone and compare performance with the current state of the art tabular methods.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

The paper recasts multivariate forecasting as scalar regressions for TabPFN zero-shot use, but the writeup gives no concrete construction of the tabular inputs needed to capture cross-channel effects.

read the letter

The core move is turning the multivariate problem into many separate scalar regressions so a tabular foundation model can tackle each one zero-shot. That is the main new framing they offer, and it directly addresses the independent-univariate limitation they note in earlier tabular time-series papers. They back it with a TabPFN-TS backbone and some comparison to other tabular baselines, which at least shows they ran the idea on real data rather than leaving it at the abstract stage. Credit for keeping the approach general enough that any regression-capable tabular model could plug in. The execution looks clean on the surface and the motivation is straightforward. The soft spot is exactly the one the stress-test flags. Nothing in the abstract or the summary description spells out how the feature vectors are assembled for each scalar target. If the rows only contain lagged values from the target channel itself, the method collapses to the independent case they say they are improving on. If they do pack in lags from all channels, there is no reported check that the zero-shot TabPFN actually learns or generalizes those cross terms, and no ablation that isolates the effect. Without that detail the central claim rests on an unverified assumption about what the tabular structure encodes. The results section is mentioned but not broken down here, so it is impossible to tell whether the gains are real or just from a different data packing. This is the kind of paper that would interest people already working on tabular foundation models or zero-shot forecasting pipelines. A practitioner looking for a quick way to try TabPFN on multivariate series might pick up the framework and test it themselves. A reader who wants evidence that inter-channel dependencies are actually being used will find the current version thin. I would send it to peer review because the idea is simple enough to evaluate quickly and the gap it targets is real; the referees can ask for the missing input construction and ablations without much trouble.

Referee Report

2 major / 2 minor

Summary. The paper proposes a generally applicable framework for zero-shot multivariate time series forecasting by recasting the problem as a collection of scalar regression tasks that can be solved directly by tabular foundation models such as TabPFN. It implements this using a TabPFN-TS backbone and reports comparisons against existing state-of-the-art tabular forecasting methods.

Significance. If the tabular reformulation demonstrably encodes cross-channel dependencies and the zero-shot results hold under proper controls, the approach would provide a simple, architecture-light route to multivariate forecasting that reuses strong tabular priors without task-specific training or custom interaction modules. The zero-shot property and use of existing PFN regression capabilities are clear strengths if supported by evidence.

major comments (2)

[Framework description] The manuscript provides no explicit description of how the tabular feature matrix is constructed for each scalar regression target. In particular, it is not stated whether the input row for predicting y_t^{(i)} contains lagged values from channels j ≠ i. Without this construction (e.g., in the framework section), the claimed ability to capture inter-channel interactions cannot be verified and the method risks reducing to independent univariate forecasting.
[Experiments / Results] No ablation or controlled experiment isolates the contribution of cross-channel features. The results section compares TabPFN-TS only against other tabular baselines; an ablation that removes all non-target-channel lags and re-runs the same zero-shot evaluation is required to substantiate the central claim that the recasting overcomes the independent-univariate limitation.

minor comments (2)

[Abstract] The abstract states that results are presented and compared with SOTA tabular methods but supplies neither dataset names, forecast horizons, nor quantitative metrics; adding these would improve readability.
[Introduction / Framework] Notation for the multivariate series (e.g., channel index, lag structure) is introduced only informally; a short formal definition would clarify the scalar-regression mapping.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the framework and strengthen the evidence for cross-channel modeling. We address each point below and will revise the manuscript to incorporate the requested details and experiments.

read point-by-point responses

Referee: [Framework description] The manuscript provides no explicit description of how the tabular feature matrix is constructed for each scalar regression target. In particular, it is not stated whether the input row for predicting y_t^{(i)} contains lagged values from channels j ≠ i. Without this construction (e.g., in the framework section), the claimed ability to capture inter-channel interactions cannot be verified and the method risks reducing to independent univariate forecasting.

Authors: We agree that an explicit description of the feature matrix construction is essential. In the proposed framework, for each scalar target y_t^{(i)}, the input row is formed by concatenating lagged values from the target channel i with lagged values from all other channels j ≠ i (using the same lag window), plus any static covariates. This construction directly encodes inter-channel dependencies within the tabular input, allowing the TabPFN regression head to learn and exploit them zero-shot. We will add a dedicated subsection (with pseudocode and a small illustrative example) to the framework section in the revision to make the matrix construction fully transparent. revision: yes
Referee: [Experiments / Results] No ablation or controlled experiment isolates the contribution of cross-channel features. The results section compares TabPFN-TS only against other tabular baselines; an ablation that removes all non-target-channel lags and re-runs the same zero-shot evaluation is required to substantiate the central claim that the recasting overcomes the independent-univariate limitation.

Authors: We acknowledge that a controlled ablation isolating cross-channel lags would provide direct evidence for the multivariate benefit. While the current comparisons are against other tabular methods (some of which also operate on multivariate inputs), we will add the requested ablation: re-running the zero-shot evaluation on the same datasets and horizons after removing all non-target-channel lags from the feature matrix. The performance drop (if any) will be reported alongside the main results to quantify the contribution of inter-channel information. revision: yes

Circularity Check

0 steps flagged

No circularity: reformulation applies existing tabular models without self-referential derivations

full rationale

The paper frames its contribution as recasting multivariate time series forecasting into scalar regression problems solvable zero-shot by tabular foundation models such as TabPFN. No equations, parameter fits, or derivations are presented that reduce the claimed performance or inter-channel capture to inputs by construction. The approach is positioned as an application of prior tabular PFN capabilities rather than a closed derivation chain. No load-bearing self-citations, ansatzes, or uniqueness theorems from the authors' prior work are invoked in a way that forces the result. This is a standard non-circular application paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that tabular foundation models trained for regression can implicitly capture time-series cross-channel structure when the problem is cast as scalar regressions; no free parameters or invented entities are described in the abstract.

axioms (1)

domain assumption Tabular prior-fitted networks such as TabPFN can perform zero-shot regression on appropriately formatted tabular inputs derived from time series.
Invoked when stating that any regression-capable tabular foundation model can solve the recast problems.

pith-pipeline@v0.9.0 · 5478 in / 1141 out tokens · 54221 ms · 2026-05-10T17:01:53.529991+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages · 1 internal anchor

[1]

Gift-eval: A benchmark for general time series forecasting model evaluation.arXiv preprint arXiv:2410.10393,

Taha Aksu, Gerald Woo, Juncheng Liu, Xu Liu, Chenghao Liu, Silvio Savarese, Caiming Xiong, and Doyen Sahoo. Gift-eval: A benchmark for general time series forecasting model evaluation. arxiv preprint arxiv:2410.10393,

work page arXiv
[3]

Chronos-2: From Univariate to Universal Forecasting

URLhttps://arxiv.org/abs/2510.15821. Samuel Dooley, Gurnoor Singh Khurana, Chirag Mohapatra, Siddartha V Naidu, and Colin White. Forecastpfn: Synthetically-trained zero-shot forecasting. InAdvances in Neural Information Pro- cessing Systems,

work page internal anchor Pith review arXiv
[4]

B., M ¨uller, S., Salinas, D., and Hutter, F

URLhttps://arxiv.org/ abs/2501.02945. Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. Deep learning for time series classification: a review.Data Mining and Knowledge Discovery, 33(4):917–963, March

work page arXiv
[5]

doi: 10.1007/s10618-019-00619-1

ISSN 1573-756X. doi: 10.1007/s10618-019-00619-1. URLhttp://dx.doi.org/10.1007/s10618-019-00619-1. Bryan Lim and Stefan Zohren. Time-series forecasting with deep learning: a survey.Philo- sophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sci- ences, 379(2194):20200209, 02

work page doi:10.1007/s10618-019-00619-1
[6]

Defu Cao, Yujing Wang, Juanyong Duan, Ce Zhang, Xia Zhu, Congrui Huang, Yunhai Tong, Bix- iong Xu, Jing Bai, Jie Tong, et al

ISSN 1364-503X. doi: 10.1098/rsta.2020.0209. URL https://doi.org/10.1098/rsta.2020.0209. Edward N. Lorenz. Deterministic nonperiodic flow.Journal of Atmospheric Sci- ences, 20(2):130 – 141,

work page doi:10.1098/rsta.2020.0209 2020
[7]

doi: 10.1175/1520-0469(1963)020⟨0130:DNF⟩2.0. CO;2. URLhttps://journals.ametsoc.org/view/journals/atsc/20/2/ 1520-0469_1963_020_0130_dnf_2_0_co_2.xml. 5 Max Planck Institute for Biogeochemistry. Jena climate dataset,

work page doi:10.1175/1520-0469(1963)020 1963
[8]

While our standardization strategies prove beneficial (Appendix E), our approach demonstrates a perfor- mance deficit compared to TabPFN-TS (Table 1)

6 A PROBABILISTICFORECASTACCURACY We further evaluate probabilistic forecast accuracy via the Weighted Quantile Loss (WQL). While our standardization strategies prove beneficial (Appendix E), our approach demonstrates a perfor- mance deficit compared to TabPFN-TS (Table 1). Future work is required to isolate the drivers of this behavior and investigate po...

work page arXiv 2025