Covariates Without Retraining: Efficient Conditioning of Time Series Foundation Models

Many forecasting problems provide information about the future through known covariates, such as calendar events, promotions, prices, or weather forecasts. Many pretrained univariate foundation models cannot directly consume this information, while native covariate mechanisms can incur inference costs that grow with the number of covariates. We ask what can be recovered when the deployed forecaster is frozen, univariate, and cannot afford that growth. We introduce a covariate adapter that decomposes each target into a component explained by the covariates and a residual forecast by a frozen, univariate time series foundation model (TSFM). The covariates never enter the TSFM. A ridge regression fitted to the observed context compresses them into a single target-aligned signal. No backbone parameters are updated, the TSFM input length is unchanged, and the procedure requires two univariate forecast calls regardless of the number of covariates, plus regression overhead. The corrected and uncorrected forecasts are then combined by per-step inverse-variance weighting derived from the predictive quantiles, with an additional regression-uncertainty term for the corrected path. Across 30 covariate-bearing tasks from fev-bench and seven TSFM configurations, the adapter reduces geometric-mean MASE by 7.6–9.2%, with median task-level improvements of 2.2–3.2%. It closes 79% of the gap to TiRex-2's native covariate path and 62% of Chronos-2's, with respective CPU speedups of $4.1\times$ and $1.7\times$. Gains vary across tasks but are consistent across backbones, suggesting that a substantial component of useful covariate information can be extracted efficiently outside the foundation model itself.