TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting on GIFT-Eval, Selected Entirely on the Training Split
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
A frozen lookup table that routes each of 97 forecasting configs to one of four modes (specialist fine-tune, quantile blend, base-model blend, or tournament pick) hits rank 3/130 on GIFT-Eval — beating everything except two agentic LLM-based systems — while running zero agents and zero LLMs at inference. Routing is the entire win: the best single base model ranks 33.8 and per-config tournament selection ranks 38.0, but the guarded router pushes mean MASE rank to 19.4, meaning you get near-top forecasting accuracy with cheap, deterministic, fully reproducible inference and no runtime reasoning cost. For anyone shipping time-series forecasting, this says invest offline compute in a backtest-driven selection table over lightly fine-tuned public models rather than paying for agentic inference loops.
TW3Cast achieves a mean MASE rank of 19.4 on the GIFT-Eval benchmark using a frozen routing table that selects between lightly fine-tuned foundation models, avoiding costly agentic or LLM-based approaches. This enables production-grade time-series forecasting with minimal GPU overhead—each candidate model costs only megabytes and minutes to test—while maintaining competitive performance against more complex systems.