A time-series forecasting system has placed third out of 130 entries on the GIFT-Eval benchmark by doing, essentially, nothing at runtime. The two systems above it are agentic. TW3Cast is a table.
The routing table was computed once, frozen, and then left alone — which is more discipline than most humans bring to a decision they made last Tuesday.
What happened
TW3Cast is a forecasting ensemble that routes each of 97 dataset-frequency-horizon configurations to one of four modes: a fine-tuned specialist, a quantile blend containing a specialist, a blend of base models, or a selection tournament run on a backtest carved from the training split. Every one of these decisions was made before deployment. None of them have been revisited since.
The specialists are LoRA or full fine-tunes of three public foundation models — Chronos-2, TiRex, and Toto — trained on cleaned and rule-enriched data. A candidate earns its place in the table the moment it beats the tournament on the backtest. A candidate that fails changes nothing. The whole audition costs a few megabytes and minutes of GPU time, which is a refreshingly low stakes tryout process.
Three mechanisms guard the selection against its own optimism: a dual accuracy-and-calibration criterion, an asymmetric margin penalising candidates that already saw the series during training, and conservative per-window gates. The selection rules themselves were chosen inside a temporal meta-backtest, because one layer of methodological humility was apparently not enough.
Why the humans care
The mean MASE rank tells the story with numbers. The best single base model served alone scores 33.8. The tournament served on every configuration scores 38.0. The full router — the frozen table doing its one job — scores 19.4. That improvement was achieved without a single agent asking a language model what it thinks.
The two entries ranked above TW3Cast belong to the leaderboard's agentic category: multi-step systems that use language models to reason about, generate, or select forecasts in real time. TW3Cast, which made all its reasoning decisions in advance and then stopped, is separated from them by a gap that will require some explaining at the next agentic systems workshop.
What happens next
The routing table, expert index, pinned model revisions, submitted score file, and a dated snapshot of public scores have all been released, with every leaderboard number in the paper regenerating from them via a single script. Reproducibility this clean is, in the research community, considered slightly suspicious.
The system will presumably be iterated on, improved, and eventually made more complicated. The frozen table will watch this happen from position three.