Ant International has launched FalconTST 2.0, a time-series AI model for cashflow and FX forecasting.
Ant International has introduced the Falcon Time-Series Transformer (TST) Model 2.0, an updated version of its forecasting model aimed at improving foreign exchange risk management in cross-border payments. The company said the model is also intended for wider use in sectors such as logistics, aviation, and ecommerce, beyond its original financial services applications.
Benchmark performance
FalconTST 2.0 has recorded a score of 0.666 on the Mean Absolute Scaled Error (MASE) metric, a measure used to assess the accuracy of time-series forecasting models. Ant International said this places the model at the top of a public global benchmark for time-series foundational models, ahead of comparable models from other technology companies. In addition, the company noted that, unlike large language models, which are built to process text, time-series models such as FalconTST are designed to interpret continuously changing numerical data, including transaction volumes, account balances, and currency positions.
Adoption by banking partners
FalconTST was first deployed internally at Ant International to manage its own cashflow and FX exposure on an hourly, daily, and weekly basis. It has since been integrated by four banks into their respective FX hedging systems. Barclays has incorporated the model into its BARX NetFX platform, while Citi has combined it with its Fixed FX Rates solution. Both are applied primarily to FX risk management for ecommerce platforms and airlines. Moreover, Standard Chartered uses FalconTST alongside its SCALE FX system, as part of a joint participation with Ant International in the PathFin.ai programme run by the Monetary Authority of Singapore.
According to Ant International, all four banking partners have now adopted the 2.0 version of the model, which the company said delivers a forecast accuracy rate of more than 93% on a consistent basis.
Technical updates in version 2.0
Ant International outlined three areas of technical development in the new version. The model distinguishes missing data from genuine zero values, addressing distortions that can arise, for example, when no bank transactions occur over a weekend. Through a framework the company refers to as ORBIT, FalconTST is designed to identify shared time-series patterns across finance, retail, energy, and tourism data, which Ant International said supports its application in new business contexts without requiring separate models per sector. The model also processes data across multiple time frequencies within a single architecture, ranging from second-level payment data to monthly economic indicators.
Broader industry rollout
Ant International said the aviation sector, where revenue and costs are often spread across multiple currencies, is one of the areas where FalconTST is already applied for FX and liquidity forecasting, alongside emerging use in ecommerce and logistics. The company has made an API trial of FalconTST 2.0 available to developers through GitHub, alongside further case studies published on its dedicated FalconTST site.
Jiang-Ming Yang, Chief Innovation Officer at Ant International, said the model reflects a broader effort to apply predictive AI to forecasting-related decisions, including liquidity planning and capital allocation, across payments and financial services more generally.
