FinCast introduces the first foundation model specifically designed for financial time-series forecasting, addressing key challenges of non-stationarity, multi-domain diversity, and varying temporal resolutions. The model is a 1B-parameter decoder-only Transformer with three novel components: (1) a Point-Quantile Loss that jointly optimizes point forecasts and quantile-based probabilistic estimates to improve robustness under distribution shifts; (2) learnable frequency embeddings that encode temporal resolution information, enabling the model to adapt to patterns at different frequencies (e.g., minute, hourly, daily); and (3) a token-level sparse Mixture-of-Experts mechanism that increases capacity efficiently and allows experts to specialize across financial domains such as stocks, cryptocurrencies, forex, and futures. The model is pretrained on over 20 billion time points from diverse financial and non-financial datasets.

Extensive experiments demonstrate that FinCast achieves state-of-the-art performance in both zero-shot and supervised forecasting benchmarks. In zero-shot settings, it reduces MSE by an average of 20% and MAE by 10% compared to existing general-purpose time-series foundation models like TimesFM, Chronos-T5, and TimesMOE. In supervised settings, fine-tuned FinCast outperforms specialized models such as PCIE, PatchTST, and D-Va, with a 23% average error reduction. Ablation studies confirm the importance of each proposed component, and qualitative analyses show that FinCast adapts well to shifting patterns across domains and temporal resolutions. The model also maintains practical inference efficiency, running on consumer-grade GPUs despite its large scale.