Intel / article

TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting on GIFT-Eval, Selected Entirely on the Training Split

Source: arxiv.org
Author: Nathan Thierry, Andre-Louis Rochet
Date: Published Sep 25, 2026

Context

TW3Cast tests whether a time-series forecasting system can compete by choosing suitable models in advance rather than using an at runtime. It builds a fixed selection table using backtests drawn from training data. For each forecasting configuration, that table selects a specialized model, a blend, or a model-selection procedure. It runs no agent or to make those choices.

In the authors' September 14, 2026 snapshot of the GIFT-Eval leaderboard, the system placed third among 130 entries by average rank on scaled absolute forecasting error. The two entries above it belong to the leaderboard's agentic category. It placed eighth on the separate measure of probabilistic forecasts, so its standing depends on which kind of forecast quality is measured.

Key takeaways · AI-distilled

TW3Cast ranked 3rd of 130 entries on GIFT-Eval by mean MASE rank as of 2026-09-14. The two entries above it use agents or language models; TW3Cast uses neither.
For each of 97 dataset, frequency and horizon configurations, a frozen table serves one of four modes: a fine-tuned Chronos-2, TiRex or Toto specialist, a quantile blend containing a specialist, a blend of base models, or a backtest tournament.
Every choice is made on a backtest carved from the training split. include a joint accuracy and criterion, an asymmetric margin against candidates that saw the series in training, and conservative per-window gates.
The best single base model reaches a mean MASE rank of 33.8 and the tournament alone 38.0, against 19.4 for the full router. The authors say every leaderboard number regenerates from released artifacts with one script.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.

Read the source arxiv.org