TW3Cast Beats 128 Systems on GIFT-Eval Time-Series Benchmark Without Agents or LLMs

TW3Cast is a time-series forecasting system that ranks third of 130 entries on the GIFT-Eval benchmark with a mean MASE rank of 19.4. It relies on a frozen routing table computed once on the training split and uses experts drawn from public foundation models (Chronos-2, TiRex, Toto) lightly fine-tuned via LoRA or full fine-tuning.

The two entries ranked above it belong to the agentic category, which uses agents or language models to reason about, generate, or select forecasts. TW3Cast runs no agent and no language model, only a frozen selection table, making it lighter and cheaper to operate.

The paper releases the routing table, expert index, pinned base-model revisions, and the submitted score file, and every benchmark number regenerates by one script. That eases reproduction and verification, but production adoption still requires evaluating performance on the organization's own data.

GIFT-Eval is a public benchmark for evaluating time-series forecasting systems, covering 97 configurations that combine datasets, frequencies, and horizons. The paper's ranking uses mean MASE rank, and as of 2026-09-14 TW3Cast stood third of 130 entries.