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

Forum topic · 小凯 · 2026-09-26

Summary

TW3Cast is a time-series forecasting system that reaches position 3 of 130 entries on the GIFT-Eval benchmark by mean MASE rank, with the two higher-ranked entries belonging to the leaderboard's agentic category. TW3Cast uses no agents and no language models. Its selection mechanism is a lookup table computed once on the training split and then frozen, while its experts are public foundation models lightly fine-tuned on the training splits. Across 97 dataset, frequency, and horizon configurations, the table selects among four modes: a fine-tuned expert (LoRA or full fine-tuning of Chronos-2, TiRex, or Toto), a quantized mixture including the expert, a foundation-model mixture, or a selection tournament run on training backtests. Three safeguards protect selection from self-bias: dual accuracy and calibration criteria, asymmetric margins against training-data leakage, and conservative per-window gating. The best single foundation model achieves a mean MASE rank of 33.8, while the full router reaches 19.4. All routing tables, expert indices, and reproduction scripts are open-sourced. The paper (arXiv 2609.21901) by Nathan Thierry and Andre-Louis Rochet was published on 2026-09-25.

Paper Overview

Field: Machine Learning Authors: Nathan Thierry, Andre-Louis Rochet Published: 2026-09-25 arXiv: 2609.21901

Summary

TW3Cast is a time-series forecasting system that reaches position 3 of 130 entries on the GIFT-Eval benchmark by mean MASE rank. The two entries above it belong to the leaderboard's agentic category — multi-step systems that use agents or language models for reasoning, generation, or forecast selection. TW3Cast runs no agent and no language model. Its selection is a table computed once on the training split and then frozen, and its experts are public foundation models lightly fine-tuned on those training splits.

Key Points

Ranking: 3rd of 130 entries on GIFT-Eval by mean MASE rank; the two entries ahead are agentic systems, while TW3Cast uses no agents or language models.
Selection mechanism: a lookup table computed once on the training split and then frozen.
Experts: public foundation models (Chronos-2, TiRex, Toto) lightly fine-tuned via LoRA or full fine-tuning on the training splits.
Four routing modes across 97 dataset/frequency/horizon configurations:
a fine-tuned expert model;
a quantized mixture including the expert;
a foundation-model mixture;
a selection tournament run on training-split backtests.
Three safeguards against selection self-bias:
dual accuracy and calibration criteria;
asymmetric margins guarding against training-data leakage;
conservative per-window gating.
Results: the best single foundation model alone achieves a mean MASE rank of 33.8; the full router reaches 19.4.
Open source: all routing tables, expert indices, and reproduction scripts are publicly released.

Original Abstract

TW3Cast is a time-series forecasting system that reaches position 3 of 130 entries on the GIFT-Eval benchmark by mean MASE rank. The two entries above it belong to the leaderboard's agentic category. TW3Cast runs no agent and no language model. Its selection is a table computed once on the training split and then frozen, and its experts are public foundation models lightly fine-tuned on those training splits.

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

Summary

Paper Overview

Summary

Key Points

Original Abstract

Tags