Falcon TST 2.0获世界权威测评第一名，推动时间序列基础模型从通用预测走向金融应用

TL;DR - Ant International released Falcon TST 2.0, a time-series foundation-model family that reports a 0.666 MASE on GIFT-Eval and targets production financial forecasting. Its significance lies in combining benchmark performance with deployments for cash-flow and foreign-exchange risk management at major banks and enterprises.

• Falcon-2.0 uses an encoder-only architecture for efficient univariate forecasting, while Falcon-X models heterogeneous multivariate relationships through a shared latent space.
• The models provide probabilistic forecasts across 21 quantiles and explicit missing-value masks, supporting scenario analysis and irregular financial data.
• Reported forecasting accuracy exceeds 93%, with deployments involving Barclays, Citi, Deutsche Bank, and Standard Chartered; the technology is also being extended to aviation, logistics, and e-commerce.
• The article cautions that benchmark leadership does not guarantee financial returns: domain adaptation, rolling backtests, tail-risk calibration, and leakage controls remain necessary.