Truace
Subscribe

Covariate-informed prompt design for LLM time series forecasting

Source article: Informed forecasting: Leveraging auxiliary knowledge to boost large language model performance on time series forecasting

The rapid adoption of large language models (LLMs) has sparked growing interest in extending their capabilities beyond traditional natural language tasks, including applications in time series forecasting. This work explores the enhancement of time series forecasting with LLMs by incorporating time-dependent covariates. We propose a set of representative prompting strategies that span a wide range of formats while incorporating covariates, and then implement the proposed framework to evaluate performance across…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1460· cite & verify
Informed forecasting: Leveraging auxiliary knowledge to boost large language model performance on time series forecasting
Lake Champlain Water Level at Rousse-Pointe(June-02-2011) by US National Weather Service. Public domain
Trace impact readingContested
P 66The P score combines the specificity and measured human impact of the grounded problem claim with the strength of this Trace’s cited sources.

Both sides are scored from claims and sources, not community votes.

G 69The G score combines the specificity and measured human impact of the grounded gain claim with the strength of this Trace’s cited sources.

In brief

Researchers tested several prompting formats that feed time-dependent covariates, meaning related background variables, to GPT-4o-mini when forecasting time series. They evaluated these strategies on three real-world datasets drawn from healthcare, service operations, and transportation.

The results suggest general-purpose language models can be adapted to forecasting without specialized training, if the prompt supplies suitable auxiliary information. The benefit depends on matching the covariate and prompt structure, and the abstract reports only one model, so how widely the gains hold is not established.

Main points

  1. The study proposes a set of prompting strategies spanning many formats that incorporate time-dependent covariates.
  2. Experiments used GPT-4o-mini on three real-world datasets: healthcare, service operations, and transportation.
  3. Errors fell by up to 70% compared with a baseline that used no covariates.
  4. Both the choice of covariate and the prompt structure were found to be critical to results.
  5. Sensitivity analyses covered censored settings and uncertainty quantification, with statistically significant gains over existing approaches.

The gain

Including relevant time-dependent covariates in GPT-4o-mini prompts reduced prediction errors by up to 70% versus a no-covariate baseline across three real-world datasets.

The problem

Poorly aligned choices of covariate and prompt structure can degrade forecasting performance rather than improve it.

The rundown

The work starts from growing interest in applying large language models beyond conventional language tasks, with time series forecasting as the target. Its approach is to add covariates to the prompt rather than relying on the series alone.

The authors report that suitable covariates combined with an appropriate prompt design substantially improve accuracy. The 70% figure is the upper end of the reported error reduction, not a uniform result.

The analysis also shows that mismatched covariate and prompt configurations can make forecasts worse. Additional sensitivity analyses address censored data settings and uncertainty, and the authors conclude that covariate integration helps bridge general-purpose LLMs and forecasting tasks.

What this doesn’t fix

The improvements depend on selecting well-aligned covariates and prompt structures, since poorly aligned configurations can reduce accuracy.

Sources

  1. Peer-reviewedPLOS One2026-10-01

The debate