TW3Cast Ranks Third on GIFT-Eval with a Frozen Router of Fine-Tuned Forecasting Models

Summary

TW3Cast is a time-series forecasting system that ranks third among 130 entries on GIFT-Eval by mean MASE rank as of September 14, 2026. The two entries ahead of it are in the benchmark's agentic category, while TW3Cast uses neither an agent nor a language model. Its routing table is computed once from training data and then frozen, with lightly fine-tuned public foundation models serving as experts. Across 97 dataset, frequency, and horizon configurations, the table selects among four modes: a specialist, a quantile blend containing a specialist, a blend of base models, or a tournament based on a training-split backtest. Specialists are LoRA or full fine-tunes of Chronos-2, TiRex, or Toto, with training data cleaned and enriched by explicit rules. A candidate is admitted as soon as it beats the tournament on the backtest, while failed candidates have no effect; each candidate requires only a few megabytes and minutes of GPU time. The system uses dual accuracy and calibration criteria, an asymmetric margin for candidates that saw a series during training, and conservative per-window gates. These selection rules were themselves chosen in a temporal meta-backtest. The best base model alone achieved a mean MASE rank of 33.8, the tournament applied everywhere reached 38.0, and the full router reached 19.4. The authors released the routing table, expert index, pinned model revisions, submitted score file, and dated public-score snapshot, with a single script reproducing every leaderboard result.