PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2604.08400v1cs.LGcs.AI

TL;DR

Multivariate forecasting often ignores inter-channel interactions by decomposing series into independent univariate tasks. This paper serializes channels into scalar regression examples for zero-shot PFN forecasting, and outperforms standard TabPFN-TS on 60% of multivariate Gift-Eval datasets, with context-length costs that grow with the number of variates.

  • Problem

    Many multivariate forecasting methods decompose channels into independent univariate problems, ignoring intra-sample dependencies among covariates.

    Source: §1 · p. 1

  • Method

    The framework flattens multivariate observations into scalar regression examples with timestamp, channel indicator, and value, enabling zero-shot PFN forecasting without retraining.

    Source: §1 · p. 2, §2.2 · p. 3

  • Results

    The method lowers MASE versus standard TabPFN-TS on 60% of multivariate datasets in Gift-Eval.

    Source: §3.1 · p. 4

  • Takeaways & Limitations

    The approach adds inter-channel interaction to off-the-shelf PFN forecasting, trading zero-shot applicability for increased context length.

    Source: §1 · p. 2, §4 · p. 4

    Channel-independent forecasting performs better on some datasets, and the conditions favoring either approach remain unresolved.

    Source: §4 · p. 5

Abstract

from arXiv · show

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.

1 INTRODUCTION

The paper addresses multivariate forecasting methods that often ignore inter-channel dependencies by introducing a zero-shot tabular reformulation for PFN regressors.

1 INTRODUCTION

  • Multivariate time series contain both temporal and intra-sample dependencies, but many prior methods decompose them into independent channel forecasts.This channel-independent strategy is attractive when training data are scarce because modeling cross-channel interactions generally requires higher-capacity models and more data.Source: §1 · p. 1
  • Prior-data Fitted Networks can solve scalar regression in context, motivating their use for time series after tabular reformulation.Earlier work applied PFNs to univariate forecasting by treating future values as scalar target labels.Source: §1 · pp. 1–2
  • The proposed framework flattens each multivariate time step into rows containing timestamp, covariate index, and value, with the next flattened value as target.This converts spatial correlations among variables into sequential dependencies interpretable by a PFN, while increasing context length and computational cost.Source: §1 · p. 2, Figure 1 · p. 2
  • The study benchmarks the method against TabPFN-TS and specialized time-series-prior architectures including TimePFN, TempoPFN, and Chronos-2.The comparisons are designed to evaluate the zero-shot multivariate framework against both a univariate decomposition baseline and trained alternatives.Source: §1 · p. 2

2 METHODOLOGY

The methodology converts multivariate forecasting into scalar regression for tabular PFNs, preserving cross-channel context without retraining or architectural changes.

2 METHODOLOGY

  • The method adapts TabPFN-TS, which treats time as an input feature and the time-series value as a scalar regression target.TabPFN-TS uses automatically generated temporal features, but lacks a mechanism for unconstrained cross-channel forecasting.Source: §2 · pp. 2–3

2.1 UNIVARIATE TIME SERIES FORECASTING AS A REGRESSION PROBLEM

  • Univariate forecasting is expressed as next-step conditional density estimation, allowing a tabular PFN to predict each future value from time-indexed context.The regression formulation uses time explicitly as a sample feature and the observed series value as the target.Source: §2.1 · p. 3

2.2 MULTIVARIATE TIME SERIES FORECASTING AS A UNIVARIATE REGRESSION PROBLEM

  • For multivariate forecasting, an indicator η identifies the component being predicted while features contain time and η and the target remains a single scalar value.This avoids retraining a foundation model to natively predict multiple targets.Source: §2.2 · p. 3
  • The rolled-out sequence lets autoregressive predictions draw from the joint multivariate distribution, and the formulation extends to more variates through a categorical channel feature.Different channels and the full horizon can be predicted simultaneously with a tabular foundation model.Source: §2.2 · p. 3
  • Each channel is independently z-score normalized before TabPFN processing and transformed back to its original scale before evaluation.The authors report that this helps when variates have different scales.Source: §2.2 · p. 3
  • The resulting TabPFN-TS-MV approach enables off-the-shelf tabular foundation models to perform multivariate forecasting.Its main design trade-off is zero-shot applicability versus longer contexts caused by unrolling.Source: §2.2 · p. 3

3 EXPERIMENTS

Experiments use the Gift-Eval benchmark and show that the proposed method improves MASE over standard TabPFN-TS on most evaluated multivariate datasets.

3 EXPERIMENTS

  • Gift-Eval contains 23 datasets spanning over 144,000 time series, 177 million data points, seven domains, and ten sampling frequencies.Across valid datasets, frequencies, and horizons, it provides 97 benchmarking tasks.Source: §3 · pp. 3–4

3.1 POINT FORECAST ACCURACY

  • The evaluation focuses on multivariate datasets because the proposed method differs from TabPFN-TS specifically in multivariate forecasting.Point forecast accuracy is measured with Mean Absolute Scaled Error (MASE).Source: §3 · p. 4, §3.1 · p. 4
  • 60% of multivariate Gift-Eval datasets have lower MASE with the proposed method than with standard TabPFN-TS.Both approaches use the same number of past time steps in this comparison.Source: Figure 2 · p. 4, §3.1 · p. 4

4 DISCUSSION AND FUTURE WORK

The framework provides zero-shot PFN-based multivariate forecasting with inter-channel interaction, while increasing context requirements and leaving conditions favoring channel dependence unresolved.

4 DISCUSSION AND FUTURE WORK

  • The framework reformulates multivariate forecasting into scalar regressions, enabling zero-shot use with PFN-based regressors and inter-channel interaction.Its context length scales linearly with the number of variates and is constrained by the base model’s maximum context.Source: §4 · p. 4
  • Unlike related sequence-flattening work such as MORPHEUS Patil et al. (2025), the framework is model-agnostic and requires neither retraining nor architectural modification.This preserves compatibility with off-the-shelf tabular foundation models.Source: §4 · p. 5
  • Channel-independent forecasting performs better on some datasets, so when channel-dependent or channel-independent methods are superior remains unresolved.The authors also identify calibration of quantile-regression probabilistic forecasts as a future research direction.Source: §4 · p. 5

A PROBABILISTIC FORECAST ACCURACY

The approach shows a performance deficit against TabPFN-TS on probabilistic forecast accuracy measured by WQL, despite beneficial standardization strategies.

B SENSITIVITY OF TABPFN-TS-MV TO COLUMN PERMUTATIONS.

Although the rolled-out representation is sensitive to column permutations, TabPFN v2's inference ensemble provides robustness, so no ordering ablation is performed.

C DETAILED RESULTS

Table 1 compares TabPFN-TS-MV with TabPFN-TS on multivariate Gift-Eval datasets using point and probabilistic forecast metrics, where lower values are better.

  • Table 1 compares the proposed approach with TabPFN-TS on multivariate datasets from Gift-Eval.Source: Table 1 · p. 7
  • Point forecast accuracy is evaluated using Mean Absolute Scaled Error (MASE).Source: Table 1 · p. 7
  • Probabilistic forecast accuracy is evaluated using Weighted Quantile Loss (WQL), with lower values preferred for both metrics.Source: Table 1 · p. 7

D COMPARISON WITH SOTA TIME SERIES MODELS

The SOTA comparison reports MASE and WQL across multivariate Gift-Eval datasets, includes averages excluding Jena Weather because of Chronos-2's poor performance, and applies matched context limits for fair comparisons.

E INVESTIGATING STRATEGIES FOR MITIGATING DISTRIBUTIONAL SHIFTS

The framework evaluates first-order differencing and channel-wise standardization to mitigate cross-channel distributional shifts. Standardization improves accuracy, whereas differencing degrades performance, so standardization is adopted.

  • The evaluated strategies target distributional discrepancies between channels in joint multivariate forecasting.The alternatives predict either first-order increments or independently standardized channel values.Source: E INVESTIGATING STRATEGIES FOR… · p. 10
  • Standardization improves accuracy, while first-order differencing degrades performance; the framework therefore adopts standardization.The comparison uses a log-scaled y-axis, and the transformations are applied independently per channel before inverse transformation for evaluation.Source: E INVESTIGATING STRATEGIES FOR… · p. 10, Figure 5 · p. 10
  • The selected standardization strategy is incorporated into the final framework after the comparative evaluation.Source: E INVESTIGATING STRATEGIES FOR… · p. 10
Loading 2604.08400v1…