RealCast-bench Leaderboard

Time series foundation models evaluated zero-shot on proprietary Carbon Arc commercial panels: 10 source panels, 81 datasets, 290 tasks across three boards (152 weekly, 102 monthly, 36 daily). Test year: October 2025 to September 2026. This table mirrors Table 2 of the paper RealCast-bench: Evaluating Time Series Foundation Models on Proprietary Commercial Panels. The panels themselves are proprietary and are not released.

Columns

• *_rel_mase: geometric mean over tasks of each task's MASE divided by Seasonal Naive's MASE. Lower is better; Seasonal Naive = 1.000. overall pools all 290 tasks, and the board columns score each board on its own. Ratios are not comparable across boards: daily Seasonal Naive repeats last week's value, a much stronger reference than weekly and monthly Seasonal Naive, which repeat the value from a year earlier.
• *_rank: rank among the 21 configurations.
• win_rate: share of (task, other configuration) pairings in which this configuration has the lower MASE, with ties counted as half.
• license, license_url: the license of the released weights, with a link to its text.
• commercial_use: yes (Apache-2.0), no (non-commercial license) or restricted (commercial use permitted under conditions; see the license).
• covariates: none means the model saw only the target series. past + calendar means it also saw the panel's other measured series up to the forecast origin, plus calendar features.

Per-task results (tasks config)

One row per (board, dataset, horizon, configuration): 290 tasks × 21 configurations. A dataset is one panel at one entity level (original, company, product, service, retailer). rel_mase and rel_wql are the task's MASE and WQL divided by Seasonal Naive's; their geometric mean over tasks gives the leaderboard scores. coverage_80pct is the share of actuals inside the 10–90% interval. Receipt Spend is scored separately in the paper and is not included.

AutoETS's WQL explodes on 11 tasks (up to about 1e35, worst on daily app_downloads at 91 days): on long horizons a few windows produce runaway prediction intervals. These values are reported as computed. The leaderboard ranks by MASE, which they do not affect.

Pairwise comparison (pairwise config)

Pairwise win rate and skill score between configurations, computed from tasks.csv with fev-bench's method (fev.analysis.pairwise_comparison, commit f9e2c7d, reimplemented; it reproduces fev's output exactly). One row per (board, metric, model_1, model_2); board is all, weekly, monthly or daily, and metric is mase or wql. win_rate is the share of tasks on which model_1 has the lower error than model_2, ties counted as half. skill_score is 1 minus the geometric mean over tasks of model_1's error divided by model_2's, with ratios clipped to [0.01, 100]. _lower and _upper bound the 95% interval from 1,000 bootstrap resamples of tasks (seed 123). As on fev-bench, each (board, metric) holds the 15 configurations with the highest average win rate, plus Seasonal Naive. + cov. means past + calendar covariates.

Setup

• No model is fine-tuned on Carbon Arc data.
• Symmetric averaging (averaging the forecast of a series with the sign-restored forecast of its negation) is enabled for every model that offers it: TimesFM-2.5, TimesFM-3.0, Timestral-209M and the three t0 models. Other settings are each library's defaults.

Submitting a model

Point us to the model's public weights and forecasting package. We run it through the same wrapper, entities and windows as every other configuration and add it to all three boards.