TW3Cast: Lightly Fine-Tuned Foundation Models Power Time-Series Forecasting to Rank 3 on GIFT-Eval

Summary:TW3Cast is a novel time-series forecasting system that ranks third on the GIFT-Eval benchmark by mean MASE. Unlike top-performing entries that rely on agents or language models, TW3Cast utilizes a routing table computed on the training split and frozen thereafter. The experts are public foundation models lightly fine-tuned on the training data. TW3Cast incorporates three mechanisms to guard against selection biases and optimizes selection rules through meta-backtesting. Its performance significa

---

TW3Cast: Lightly Fine-Tuned Foundation Models for Time-Series Forecasting

Key Breakthroughs

TW3Cast has demonstrated its prowess in time-series forecasting by ranking third on the GIFT-Eval benchmark with a mean MASE score. The system boasts the following key features:

• No Agents, No Language Models: TW3Cast relies solely on a routing table computed on the training split and frozen thereafter, without the need for agents or language models for reasoning or forecast selection.
• Lightly Fine-Tuned Experts: The experts selected by the routing table are public foundation models (e.g., Chronos-2, TiRex, Toto) that have been lightly fine-tuned using LoRA or full fine-tuning, with training data cleaned and enriched by explicit rules.
• Multi-Modal Selection Mechanisms: The routing table offers four modes for 97 dataset, frequency, and horizon configurations: specialist, quantile blend containing a specialist, base model blend, and tournament selection on a backtest.
• Bias Mitigation: TW3Cast employs dual accuracy and calibration criteria, an asymmetric margin against candidates that saw the series during training, and conservative per-window gates to guard against selection biases.

Technical Highlights

1. Routing Table Design: The routing table is designed through meta-backtesting to ensure the effectiveness of selection rules over time.
2. Lightweight Fine-Tuning: The fine-tuning process for expert models is efficient and cost-effective, with each candidate requiring only a few megabytes and minutes of GPU time.
3. Performance Advantage: TW3Cast's overall performance significantly surpasses using the best base model alone (mean MASE rank of 33.8) and tournament selection alone (38.0), with the full router achieving a mean MASE rank of 19.4.

Industry Impact

TW3Cast's release brings new perspectives to the field of time-series forecasting, particularly in scenarios where resource constraints and real-time requirements are critical. Its lightweight design and efficient performance make it a strong contender in AI-driven forecasting systems.

Developer Recommendations

• Model Integration: Developers can integrate TW3Cast's routing table mechanism into existing forecasting systems to improve prediction accuracy and efficiency.
• Fine-Tuning Strategy: TW3Cast's lightweight fine-tuning strategy can serve as a reference for optimizing the performance of other foundation models.
• Bias Mitigation: TW3Cast's bias mitigation mechanisms provide new insights for AI model selection and evaluation, helping to enhance model fairness and reliability.

Conclusion

TW3Cast demonstrates the powerful potential of combining lightly fine-tuned foundation models with intelligent routing mechanisms in time-series forecasting, paving the way for new directions in AI-driven forecasting systems.

---

Source: ArXiv AI (cs.AI) (2026-09-25)

— END —

Tags: #Time-Series Forecasting #Foundation Models #Lightweight Fine-Tuning #GIFT-Eval #Routing Mechanism