Time series and forecasting · insight-1251
Check that a zero-shot foundation forecaster can use your future-known drivers before expecting promotion peaks
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, ML engineer, Applied AI
- Industries
- retail, energy, e-commerce
Summary: situation, action, reason, boundary
Situation
A team evaluates an off-the-shelf pretrained forecaster on weekly sales with a known promotion calendar and finds that it misses every promotion peak.
Action
Read the model documentation for support of past and future covariates. Use a covariate-capable variant, or combine the model with a covariate-aware model such as a boosted tree. Report a no-covariate baseline so the value of the drivers is visible.
Reason
The original Chronos paper focused on univariate forecasting and proposed stacking with covariate-capable models such as LightGBM, while later releases of Chronos-2 and TimesFM list covariate support.
Boundary
Retail, demand and energy forecasting with known promotions, prices, holidays or weather.
Symptom or risk
A forecaster that only extrapolates the history ignores the drivers that carry the signal, and it does so without any error message.
Exceptions
Series whose variation is almost entirely seasonal and trend need no covariates.
Trade-offs
Covariate-capable variants add inference cost and a new place for leakage.
Pitfalls
Feeding actual past covariates where future-known ones are required, and calling the result zero-shot accuracy.
Lessons and concepts
Claims and their sources
The Chronos paper focuses on univariate forecasting of uniformly spaced series and suggests stacking with covariate-capable models such as LightGBM as a way to handle exogenous information.
As read on 2026-10-12, the Chronos repository README describes Chronos-2 (released 20 Oct 2025) as offering zero-shot univariate, multivariate and covariate-informed forecasting.
As read on 2026-10-12, the TimesFM README lists covariate support through XReg for TimesFM 2.5 (added back 29 Oct 2025) and covariate support for TimesFM 3.0; it does not say whether earlier releases support covariates.