New Time-Series Forecasting System Achieves Top Ranks on GIFT-Eval Benchmark

VERA Brief

AI-generated. Grounded in the article and its cited sources.

TW3Cast, a new time-series forecasting system, achieved third place on the GIFT-Eval benchmark. It uses a frozen router with fine-tuned foundation models and does not employ agents or language models.

Key facts

• TW3Cast ranked third out of 130 entries on the GIFT-Eval benchmark for time-series forecasting.
• The system uses a static selection table that directs forecasts from foundation models.
• TW3Cast does not use agents or language models, unlike the top two entries.
• Specialist models are used if they perform better than a tournament on a backtest.
• The selection process includes accuracy and calibration criteria.