Ant International Launches Falcon TST 2.0, Tops Global Benchmark for Time-Series Forecasting

Falcon TST 2.0 builds on Ant International's earlier work. In 2025, the model family was first validated in forex exposure and treasury management. In 2026, the company introduced Falcon-2.0 to improve univariate forecasting efficiency and Falcon-X for heterogeneous multivariate relationship modeling. Falcon TST 2.0 represents a specialized attempt that combines general-purpose time-series foundation capabilities with validated financial workflows, the company said.

Falcon-2.0 is an encoder-only univariate TSFM based on the ORBIT training framework, according to Ant International. It is designed to reduce latency and error accumulation compared with autoregressive generation, and its public API outputs 21 quantile levels from 0.01 to 0.99, with an input_mask parameter to explicitly handle missing values such as holidays and trading-day misalignment. Falcon-X, introduced in a paper released on May 26, 2026, converts variables with different physical meanings into a unified latent-space representation and uses differential attention to identify both positive and negative relationships. Its largest publicly tested version has 591 million parameters, and the paper reports a MASE of 0.687 on GIFT-Eval.

Ant International has documented finance-sector deployments. Barclays integrated TST into its BARX NetFX forex hedging platform in May 2025, enabling hour-, day- and week-ahead forecasts of cash flows and forex exposure. Citi piloted Falcon with its fixed forex rate solution for airline clients; the bank said the first airline customer lowered hedging costs, with Ant International reporting savings of about 30 percent. Standard Chartered connected Falcon to its SCALE liquidity engine in August 2025, and the bank said forecast accuracy for Ant International's forex exposure exceeded 90 percent, with Falcon covering more than 60 percent of Ant International's forex conversion transactions. Related forex costs fell by up to 60 percent and liquidity management costs by up to 50 percent, according to the bank.

The release comes as global tech firms push time-series foundation models in new directions. Amazon's Chronos-2, Google's TimesFM 2.5, Salesforce's Moirai 2.0 and IBM's FlowState have all added multivariate, long-context, probabilistic or cross-sampling-rate capabilities. However, academic research has cautioned that financial time series pose special challenges. A working paper by researchers at the University of Manchester and University College London, published in November 2025, found that off-the-shelf pretrained TSFMs performed weakly in zero-shot financial forecasting, and performance improved notably after re-pretraining on financial data. A separate study in June 2026 of five highly liquid U.S. stocks found that TSFMs won eight of ten tasks but their gains over random-walk baselines were generally small and rarely statistically significant.

Quantile outputs from models like Falcon TST are useful for constructing pessimistic, neutral and optimistic scenarios, but they should not be directly equated with regulatory value-at-risk or expected shortfall measures. In a real risk-control system, coverage tests, stress scenarios and model risk management are still required to determine whether the quantiles are stably calibrated. Multivariate inputs also need careful variable selection: a study by Santa Clara University using Chronos-2 found that jointly inputting related variables generally outperformed univariate inputs, but directly mixing stock and interest-rate panels degraded predictive accuracy.