Timestral

The history stays in the context. Forecast tokens read that history. A missing observation stays missing, rather than being treated as a value the model must also predict.

Each series has a role. A known-future covariate, such as the month or a bank holiday, can see the horizon. The target cannot. Covariates do not read the target’s own future, so a calendar does not leak the print.

Interactive runs and the files on these tabs use Timestral-28M until the nightly 209M job overwrites them. Zero-shot. There is no fine-tune in this build. One forecast runs at a time.

Payrolls are scored on the first print, not the revision. The context is the latest published vintage. UK monthly GDP is scored on the official month-on-month series already used by the high-speed nowcast: walk-forward, calm regime, latest vintage, Covid window excluded from that nowcast’s own backtest.

The inference code is Apache-2.0. The weights are Modified MIT and stay on the forecast host.Model card · Mistral’s note.

This page does not add an Official Macro tool, or a change to inflation, the payrolls lab, high-speed, oil, or housing.