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

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

Introduction

The paper "Zero-shot Multivariate Time Series Forecasting Using Tabular Prior Fitted Networks" (2604.08400) presents a unified framework for addressing multivariate time series (MTS) forecasting by leveraging advances in tabular foundation models, specifically Prior-data Fitted Networks (PFNs). While recent research has demonstrated that tabular PFNs such as TabPFN-TS exhibit competitive performance on univariate time series by treating forecasting as a scalar regression task, existing multivariate approaches predominantly adopt a channel independence (CI) assumption, decomposing MTS into multiple univariate regressions and thereby neglecting cross-channel (intra-sample) dependencies. This work proposes and empirically validates a methodology to serialize multivariate sequences into a format amenable to zero-shot prediction with tabular foundation models, offering new insights and competitive benchmarks against state-of-the-art techniques.

Methodology

The core contribution is a problem reformulation that enables MTS forecasting using generic tabular PFNs—without domain-specific retraining or architectural modification. The methodology transforms the $d$-variate time series at each time step into $d$ scalar regression samples, each corresponding to a (timestamp, channel index, value) tuple. A channel indicator is introduced as a categorical variable, permitting each row to encode both temporal and channel information. The regression target is then the subsequent value in the flattened sequence, thus recasting the forecasting problem as scalar regression over a tabular context while implicitly encoding inter-channel structure.

A central element is the normalization of each channel prior to the rolling/unrolling process via z-score normalization, with inverse transformation applied post-prediction. This normalization is critical for MTS with disparate scaling across variates. The approach (denoted TabPFN-TS-MV) allows joint prediction of all future samples for all channels, facilitating simultaneous spatial and temporal correlation modeling using only off-the-shelf PFNs.

Distinct from existing work that either modifies architectural priors or adapts training data to introduce temporal modeling into PFN frameworks, this approach is explicitly channel-dependent (CD), thereby modeling inter-variable dependencies at inference time, which is infeasible in vanilla CI univariate decompositions.

Experimental Results

Experiments utilize the Gift-Eval benchmark (Aksu et al., 2024), spanning 23 datasets, over 144,000 time series, and 177 million data points across diverse application domains and frequencies. The authors focus on multivariate settings and compare TabPFN-TS-MV against TabPFN-TS (CI), and specialized SOTA models: TimePFN [see (2604.08400) refs], TempoPFN, and Chronos-2.

• Point Forecast Accuracy (MASE): TabPFN-TS-MV yields lower MASE compared to TabPFN-TS on 60% of the Gift-Eval multivariate datasets, indicating that explicit modeling of inter-channel dependencies via tabular serialization confers consistent improvements.
• Probabilistic Forecast Accuracy (WQL): For weighted quantile loss, TabPFN-TS-MV lags TabPFN-TS and advanced models slightly, suggesting a calibration deficit in the quantile regression context. This is flagged as an avenue for further model and inference pipeline enhancement.
• Comparison to SOTA Models: When averaged across datasets excluding the problematic high-variance Jena-Weather set, TabPFN-TS-MV attains competitive MASE metrics, outperforming or matching specialized time series models except for TempoPFN and Chronos-2 (the latter only in some configurations). Its rank approaches that of models with much heavier adaptation and temporal context engineering.

Notably, the increased context length due to serialization is highlighted as a primary bottleneck—this scales with the number of channels and is bounded by the PFN context window. With modern PFN architectures (e.g., TabPFN-2.5), expanded context lengths are becoming practical, further mitigating this limitation.

Practical and Theoretical Implications

This framework confirms that general-purpose tabular foundation models have untapped potential for generic structured data forecasting. The approach is model-agnostic, requiring neither new parameterization nor retraining to enable multivariate, channel-dependent forecasting in a strict zero-shot regime. It provides a principled pathway for integrating heterogeneous time series sources (with arbitrary covariate structure) into tabular learning paradigms, simplifying MTS forecasting pipelines in contexts with limited data and/or resources for retraining foundational models.

Theoretically, the work exposes a nuanced tradeoff between explicit context modeling (using CD serialization) and memory/compute efficiency (as in CI decompositions). It reveals that context length, rather than model capacity or supervision resource, is the chief scalability constraint, which will diminish with continued scale-up in tabular foundation models.

Furthermore, the framework is extensible: categorical, continuous, and arbitrarily indexed variables can be handled under the same general serialization protocol, and incorporation of new calibration strategies (e.g., improved probabilistic regression loss functions) may close the remaining gap with highly optimized SOTA models.

Limitations and Future Directions

Limitations stem mainly from computational overhead induced by context sequence expansion and potential sensitivity to serialization ordering—though ensemble properties within the PFN family mitigate some of the latter. Quantile regression calibration and uncertainty estimation within the proposed serialization remain incomplete. In addition, the empirical observation that CI methods may outperform CD methods under specific dataset regimes suggests dataset-adaptive strategies may be necessary for best-in-class generalization.

Future research should investigate:

• Extensions to more diverse variable types and higher-dimensional dependencies.
• Efficient serialization and batching for memory-aware foundation model inference.
• Automated adaptive selection among CI and CD formulations depending on data structure.
• Further assessment of the method with cutting-edge PFN variants and larger-scale tabular LLMs.

Conclusion

This paper demonstrates a general, efficient, and model-agnostic methodology for multivariate time series forecasting using tabular prior-fitted networks, establishing that zero-shot foundation models can effectively handle arbitrarily structured MTS data when recast as scalar regression over serialized (flattened) tabular contexts. The empirical results confirm that explicit context modeling of inter-channel dependencies confers measurable point forecast accuracy gains with only minor calibration deficits in probabilistic settings. This work suggests that as tabular foundation contexts continue to expand, such paradigms will become increasingly competitive with bespoke time series models for MTS in real-world applications.