Time-Aware Prior Fitted Networks for Zero-Shot Forecasting with Exogenous Variables

ICLR 2026 Conference Submission18625 Authors

Keywords: Neural Forecasting, Time Series, PFN, Zero-shot, Foundational Model

TL;DR: Through our synthetic data generation procedure and architectural modifications we incorporate time series inductive biases to our PFN model for accurate forecasting

Abstract: In many forecasting settings, the target series comes with exogenous covariates: promotions and prices for retail demand, temperature for energy load, calendar/holiday flags for traffic or sales, and grid load or fuel costs for electricity prices. Ignoring such exogenous covariates can seriously degrade forecasting accuracy, especially when they signal phase changes or spikes in the target series. Most current time-series foundation models (e.g., \texttt{Chronos}, \texttt{Sundial}, \texttt{TimesFM}, \texttt{TimeMoE}, \texttt{TimeLLM}, and \texttt{LagLlama}) ignore exogenous covariates and make forecasts solely from the time-series history, limiting their performance. In this paper we focus on bridging this gap by developing \texttt{ApolloPFN}, a prior-data fitted network (PFN) that is time-aware (unlike prior PFNs) and that natively incorporates exogenous covariates (unlike prior univariate forecasters). Our design introduces two major advances: (i) a synthetic data generation procedure tailored to resolve the failure modes that arise when tabular (non-temporal) PFNs are applied to time-series, and (ii) time-aware architectural modifications that embed the inductive biases needed to fully exploit the time-series context. We demonstrate that \texttt{ApolloPFN} achieves state-of-the-art results across benchmarks containing \emph{exogenous} information such as M5 and electric price forecasting.

Primary Area: learning on time series and dynamical systems

Submission Number: 18625