ChronosX: Adapting Pretrained Time Series Models with Exogenous Variables

TL;DR

This work addresses the challenge of incorporating exogenous covariates into pretrained time-series forecasting models. It introduces ChronosX, a modular adapter framework that injects past covariates into token embeddings and future covariates into logits, allowing covariate-aware forecasting without heavy modification to the pretrained backbone. The authors demonstrate strong improvements on a 32-dataset synthetic benchmark and 18 real covariate datasets, and show that the same framework extends to other pretrained models (TimesFMX, MOMENTX). Overall, ChronosX enables efficient covariate integration with pretrained time-series models, offering improved probabilistic forecasts and faster downstream adaptation, while highlighting trade-offs between adapter training and zero-shot capabilities.

Abstract

Covariates provide valuable information on external factors that influence time series and are critical in many real-world time series forecasting tasks. For example, in retail, covariates may indicate promotions or peak dates such as holiday seasons that heavily influence demand forecasts. Recent advances in pretraining large language model architectures for time series forecasting have led to highly accurate forecasters. However, the majority of these models do not readily use covariates as they are often specific to a certain task or domain. This paper introduces a new method to incorporate covariates into pretrained time series forecasting models. Our proposed approach incorporates covariate information into pretrained forecasting models through modular blocks that inject past and future covariate information, without necessarily modifying the pretrained model in consideration. In order to evaluate our approach, we introduce a benchmark composed of 32 different synthetic datasets with varying dynamics to evaluate the effectivity of forecasting models with covariates. Extensive evaluations on both synthetic and real datasets show that our approach effectively incorporates covariate information into pretrained models, outperforming existing baselines.

Figures (33)

• Figure 1: Architecture of ChronosX. It consists of two blocks: the first block adds covariates of the past and updates token embeddings. The second block adds future covariates updating the estimated distribution of the pretrained model.
• Figure 2: Input and Output injection blocks used in ChronosX. Observe that IIB and OIB take past and future covariates, respectively. \ref{['fig:IIB']} shows that the input injection block takes a pretrained tokenized embedding together with covariates of the past, whereas \ref{['fig:OIB']} takes the pretrained logits (expressed as the final hidden state multiplied by a pretrained matrix) together with covariates of the future. The blue color indicates that the module is taken from the pretrained model.
• Figure 3: Illustration of the generation of synthetic time series. Different combinations of the $4$ possible main signals modified by one of the $4$ possible external covariates through a chosen operator ($+$ or $\times$), result in realistic synthetic time series with covariates.
• Figure 4: Evaluations on Simple and Complex datasets as introduced in \ref{['sec:synthetic-data-explanation']}. Scores are normalized by Chronos Small. We can see that our proposed models ChronosX and ChronosX(FF) effectively incorporate covariates.
• Figure 5: Evaluations on real datasets with covariates. We report the Aggregated Relative WQL and MASE.