Contextual Time Series Forecasting with Exogenous Covariates

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 299 papers110 conceptsfrozen 2026-10-04

Traditional time series forecasting operates on a fundamental assumption: the future of a variable is encoded within its own past. Whether using classical statistical methods like ARIMA or deep learning architectures like LSTMs, these "univariate" models treat the target signal as a closed system. However, real-world systems—from stock markets to power grids—are rarely closed. They are buffeted by external shocks, policy shifts, and environmental changes that historical price or demand data alone cannot anticipate.

A growing body of research is moving beyond this univariate trap by integrating exogenous covariates: external variables that provide the necessary context to explain why a trend might deviate from its historical pattern.

The Limits of Endogenous History

The primary failure mode of univariate models is "information lag." As Xudong Yang notes in the context of agricultural commodities, models relying solely on historical prices fail to react to sudden supply-side shocks. In financial markets, Rudy Morel, Stéphane Mallat, and Jean‐Philippe Bouchaud argue that purely endogenous models cannot anticipate market reactions to geopolitical shifts or policy changes because these events are not yet encoded in the price series.

To bridge this gap, researchers are evolving models to "see" the environment. Jing Fan and colleagues demonstrated this by transitioning from ARIMA to ARIMAX, incorporating real estate data as an explanatory variable to predict service industry trends with significantly higher precision.

Hybridizing Linear Rigor and Non-Linear Complexity

One effective strategy for integrating external context is the hybrid model. These systems use a "divide and conquer" approach: a statistical model handles the stable, linear trends, while a machine learning model captures the volatile, non-linear residuals influenced by external factors.

Fangyi Li illustrates this with an ARIMA-LSTM hybrid for fund net value prediction. ARIMA filters out the predictable linear patterns, leaving a "residual" signal that contains the complex impact of market indices and trading volumes. The LSTM then focuses exclusively on these residuals. Similarly, Zheng Yin and Lei Dai combined ARIMA with Support Vector Machines (SVM) to predict coal prices, using the SVM to account for non-linear influences like freight indices and trade policy changes.

Extracting Causal Signals from Unstructured Text

Perhaps the most significant leap in contextual forecasting is the use of Large Language Models (LLMs) to digest unstructured data—news, earnings calls, and social media—into structured signals.

Models like STRUX, developed by Yiming Lu and Fei Liu 0004, transform verbose corporate transcripts into "fact tables" with assigned impact strengths. This allows the model to "de-bias" executive optimism by grounding it in external market realities. Taking this further, the TiMi framework by Jiafeng Lin and Zhongyi Pei uses a Multimodal Mixture-of-Experts (MMoE) to extract causal insights from text to guide numerical predictions, effectively allowing the model to understand the "narrative" driving the numbers.

Adapting Foundation Models for Covariates

The current frontier involves bringing covariate awareness to "Zero-Shot" Time Series Foundation Models (TSFMs). While models like Chronos or TimesFM are powerful, they are often pre-trained exclusively on univariate data.

Qin Guo and Yong Liu introduced CoRA (Covariate-awaRe Adaptation) to solve this. CoRA uses a causal weighting mechanism to inject external information—like weather or images—into pre-trained univariate models without destroying their learned weights. This allows a general-purpose model to become context-aware for specific tasks, such as Andreas Auer and Raghul Parthipan demonstrate by enabling in-context learning with external economic shifts.

Why It Works: Reducing Aleatoric Uncertainty

The intuition behind this research direction is the reduction of aleatoric uncertainty—the "randomness" in a signal that is actually caused by variables the model simply hasn't seen. By treating factors like interest rates, weather, or news sentiment as covariates rather than noise, these models transform "random" fluctuations into predictable responses. As shown in SIFusion, integrating wind speed and sea temperature doesn't just add data; it provides the physical "why" that allows the model to predict Arctic ice shifts that historical ice data alone would miss.

Papers in this version

Showing 60 of 99 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 131 papers
  2. Version 1 · 2026-10-04 · 99 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_exogenous_variable_exclusion_8fd338a4,
  title = {exogenous variable exclusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/exogenous-variable-exclusion/rdr_8fd338a4969245a1b0848844cc2a979b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_contextual_time_series_forecasting_with__8fd338a4_v1,
  title = {Contextual Time Series Forecasting with Exogenous Covariates},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/contextual-time-series-forecasting-with-exogenous-covariates/rdr_8fd338a4969245a1b0848844cc2a979b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}