Editorial illustration for TW3Cast Freezes Router Selection for Time-Series Forecasting on 97 Datasets
TW3Cast Freezes Router Selection for Time-Series...
A team behind a system called TW3Cast landed third out of 130 entries on the GIFT-Eval time-series forecasting leaderboard, ranked by mean MASE, as of September 14, 2026. The two systems ranked above it both belong to the benchmark's agentic category, setups that lean on language models or agents to reason through, generate, or select forecasts at inference time. TW3Cast does none of that. No agent, no language model, just a frozen router sitting on top of a set of lightly fine-tuned foundation models with fixed base-model revisions.
That framing matters because GIFT-Eval has increasingly rewarded systems that add reasoning layers on top of forecasting models, often at real computational cost. TW3Cast's pitch is that a fixed routing table and a static expert index, selected entirely on the training split rather than tuned against the leaderboard itself, can hold its own against systems built around active model selection or generation. The team released the routing table, the expert index, the pinned model revisions, the submitted score file, and a dated snapshot of the public scores, so every number in the result can be regenerated from a single script.
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.
Why this matters
TW3Cast's third-place finish is a useful data point for anyone building forecasting systems on a budget: a frozen lookup table beats most agentic pipelines while running none of the inference-time reasoning those systems depend on. The two entries ahead of it use agents or LLMs to generate or select forecasts at runtime, which costs compute and adds latency. TW3Cast does its work once, on the training split, then locks the routing table and reuses LoRA-tuned public models per dataset-frequency-horizon combination.
For teams shipping forecasting products, that's a real engineering tradeoff worth studying: precomputed routing over 97 configurations versus paying an inference tax for every prediction. It also raises a fair question about how GIFT-Eval's rankings should be read, since a static router selected entirely offline is competing directly against systems doing live reasoning, and coming out ahead. Anyone citing leaderboard position as a proxy for "smarter" models should look at what's actually happening at inference time before drawing conclusions.
The method matters as much as the rank here.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Latent-Space Distillation Compresses Streaming Neural Audio Encoders
From Archives PrismML Brings Its Tiny, 4x-Smaller LLMs to Qualcomm Smart Glasses
Further Reading
- SalesforceAIResearch/gift-eval - GitHub
- A Benchmark For General Time Series Forecasting Model Evaluation - arXiv
- Time Series Forecasting Benchmark: Introducing GIFT-Eval - Salesforce Blog
- GIFT-Eval Leaderboard — All Models, Filters & Rankings - TSFM.ai
- TSRouter: Dynamic Modality-Model Selection for Time Series Forecasting - arXiv
