TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting

Overview

TW3Cast is a time-series forecasting system that reached 3rd place out of 130 entries on the GIFT-Eval benchmark by mean MASE rank, as of 2026-09-14. Notably, the two entries ranked above it belong to the leaderboard's agentic category—multi-step systems that use agents or language models to reason about, generate, or select forecasts.

TW3Cast, by contrast, runs no agent and no language model.

Design Philosophy

The system's core innovation is its selection mechanism: a routing table computed once on the training split and then frozen. Its experts are public foundation models, lightly fine-tuned on those training splits.

How the Router Works

For each of the 97 dataset, frequency, and horizon configurations, the table serves one of four modes:

1. Specialist — a LoRA or full fine-tune of Chronos-2, TiRex, or Toto, whose training data was cleaned and enriched by explicit rules.
2. Quantile blend — a blend that contains a specialist.
3. Base-model blend — a blend of base models.
4. Selection tournament — a tournament played on a backtest carved from the training split.



Every decision in the table was taken on that backtest. A specialist is admitted the moment it beats the tournament there, so a candidate costs only a few megabytes and minutes of GPU time—and a failed candidate changes nothing.




Guarded Mechanisms




Three guarded mechanisms protect the selection from its own biases:



• A dual accuracy and calibration criterion
• An asymmetric margin against candidates that saw the series during training
• Conservative per-window gates



The selection rules themselves were chosen inside a temporal meta-backtest.




Results




| Configuration | Mean MASE Rank |

|---|---|

| Best base model served alone | 33.8 |

| Tournament served on every configuration | 38.0 |

| Full router | 19.4 |




Reproducibility




The routing table, expert index, pinned base-model revisions, submitted score file, and a dated snapshot of the public scores are all released. Every leaderboard number in the paper regenerates from these artifacts via a single script.




Metadata



• Subjects: Artificial Intelligence (cs.AI)
• Cite as: arXiv:2609.28506 [cs.AI]
• Submitted: 16 Sep 2026
• DOI: https://doi.org/10.48550/arXiv.2609.28506

via ArXiv AI