Benchmarks

TiRex is evaluated on two public leaderboards that cover a broad mix of time-series domains and scoring rules.

GIFT Eval

What it is: The GIFT Eval provides a unified zero-shot evaluation over 23 datasets, ~144k time series and 177 M data points, spanning seven domains (from from retail, energy, traffic, up to macroeconomic sources), 10 frequencies, multivariate inputs, and short- to long-term horizons (see Sec. 4 of the TiRex paper).
Metrics: Models are reported and ranked using MASE (point forecasts) and CRPS (probabilistic forecasts); the public leaderboard exposes MASE_Rank and CRPS_Rank for comparison.
Reproducibility: The evaluation uses fixed splits (final 10% of each dataset as test) with non-overlapping rolling windows; the code, data, and leaderboard are publicly available to reproduce results. Have a look at the TiRex GIFT Eval Notebook to reprocude the results.

What it is: The fev-bench leaderboard reports results on 100 forecasting tasks across seven domains, including 46 tasks with covariates and both univariate and multivariate settings, with varied sampling cadences and horizons (see fev-bench paper). It includes telemetry, finance, climate, and competition datasets with varying sampling cadences.
Metrics: Tasks use MASE for point forecasts and SQL (Scaled Quantile Loss) for probabilistic forecasts; leaderboard summaries emphasize win rates and skill scores with bootstrap confidence intervals for statistically rigorous aggregation.
Reproducibility: Submissions follow the lightweight fev library and fully specified task definitions and splits; extended tabular and interactive results are hosted on the HF Space. Have a look at the fev-bench TiRex Example to reprocude the results.

TiRex consistently ranks among the top zero-shot forecasters across common leaderboards without task-specific fine-tuning.
The public evaluation scripts, datasets, and checkpoints referenced above allow anyone to reproduce the official scores end-to-end.