TW3Cast ranks third on GIFT-Eval time-series forecasting benchmark without using agents or language models

A new arXiv preprint introduces TW3Cast, a time-series forecasting system that ranks third on the GIFT-Eval benchmark by mean MASE rank as of 2026-09-14. Unlike the two entries above it, which belong to the leaderboard’s agentic category, TW3Cast does not use agents or language models. Instead, it employs a frozen routing table computed once on the training split and selects among experts that are public foundation models lightly fine-tuned on those training splits.

The system serves one of four modes for each of the 97 dataset, frequency, and horizon configurations: a specialist (a LoRA or full fine-tune of Chronos-2, TiRex, or Toto with cleaned and enriched training data), a quantile blend containing a specialist, a blend of base models, or a selection tournament played on a backtest carved from the training split. Every decision in the routing table is made using only the training split backtest.

TW3Cast admits a specialist only if it beats the tournament on the training backtest, with candidate evaluation costing a few megabytes and minutes of GPU time. Failed candidates do not alter the system. Three safeguards mitigate selection bias: a dual accuracy and calibration criterion, an asymmetric margin against candidates that saw the series during training, and conservative per-window gates. The selection rules themselves were chosen via a temporal meta-backtest.

The authors report mean MASE ranks for ablations: the best base model alone reaches 33.8, the tournament alone reaches 38.0, and the full router reaches 19.4. The routing table, expert index, pinned base-model revisions, submitted score file, and dated snapshot of public scores are released, and all leaderboard numbers can be regenerated from them using a provided script.