ModelsarXiv
DualCast: Dual-Path LM Introduced for Bimodal Financial Time-Series Forecasting
Researchers have released DualCast, a dual-path framework combining fast numerical forecasting and optional news-conditioned revisions using a frozen Qwen3-8B language model extended with a custom financial vocabulary. The model achieves state-of-the-art performance across multiple asset classes and forecasting horizons.
The DualCast model introduces a dual-path approach for financial time-series forecasting by extending a frozen general-purpose language model (Qwen3-8B) with a discrete financial vocabulary. It pairs advanced representation techniques with adaptive quantization methods to enhance codebook utilization, and offers both a fast and a slow prediction path.
Key Features
- Extends a frozen Qwen3-8B model with a discrete financial token vocabulary.
- Represents log-return patches using a learned summary token and three residual shape tokens, allowing pattern sharing and drift/volatility preservation.
- Implements adaptive frequency-equalizing residual vector quantization to improve utilization of discrete codes without sacrificing reconstruction accuracy.
- Fast path: trains only financial-token embeddings and output heads on the frozen backbone for quick numerical forecasts.
- Slow path: employs a toggleable LoRA adapter to revise predictions by conditioning on fast outputs and contemporaneous news, initialized by supervised fine-tuning and further optimized via return-space group relative policy optimization.
Performance and Findings
- Zero-shot evaluations across equities and energy prices at multiple time resolutions show the slow path achieves the lowest mean absolute percentage error in 8 out of 12 dataset-horizon settings, including all longest-horizon scenarios.
