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

Summary

TW3Cast is a forecasting system covering all 97 dataset/frequency/horizon configurations of GIFT-Eval. A frozen routing table assigns each configuration to one of four serving modes: a designated specialist (a lightly LoRA- or fully fine-tuned Chronos-2, TiRex or Toto trained on cleaned, rule-enriched data), a quantile blend containing a specialist, a blend of base models, or a tournament run on a backtest carved from the training split. Every decision was taken on that backtest; the selection rules and all thresholds were themselves chosen in a temporal meta-backtest, with guards including a 12% anti-memorization margin, a dual MASE+CRPS criterion and per-window gates. Inserted into the 2026-09-14 leaderboard snapshot, it reaches a mean per-configuration MASE rank of 19.4, position 3 of 130, behind two agentic systems; by weighted quantile loss it is position 8. Ablations attribute the gain to routing: best single member 33.8, tournament everywhere 38.0, full router 19.4, and against a test-side per-configuration oracle it wins 37 of 97 configurations. The full release (router table, checkpoints, score file, dated snapshot, figure script) makes every number regenerable. Main limitations: a single benchmark, no uncertainty quantification behind the 0.3-rank gap to position 4, calibration lagging point accuracy, and one expert shipped only as verified quantiles.