Built on the study “Re(Visiting) Time Series Foundation Models in Finance” (Rahimikia, Ni, and Wang, 2025), FinText advances the frontiers of AI-driven financial forecasting through large-scale, domain-specific Time Series Foundation Models (TSFMs). These large scale mdoels models are pre-trained from scratch on massive datasets—spanning over two billion daily excess return observations across 89 markets—to capture the complex temporal dynamics of global finance.
This study shows that generic foundation models fail to generalize effectively to financial data, whereas finance-native pre-training leads to significant improvements in predictive accuracy and portfolio performance. Each model is chronologically pre-trained to prevent look-ahead bias, ensuring realistic, time-consistent forecasting. The resulting FinText-TSFM suite includes hundreds of Chronos and TimesFM variants aligned with different years, markets, and datasets empowering researchers and practitioners with transparent, bias-free, and reproducible forecasting tools.
Developed through collaboration between the University of Manchester, University College London, Shanghai University, and partners including, UKRI, Isambard-AI, and N8 Bede, FinText embodies the integration of financial economics, machine learning, and high-performance computing. Together, these efforts establish a new foundation for reliable, interpretable, and scalable AI in finance.