Ant International's FalconTST 2.0: Forecasting AI Built for Finance, and Why Finance Needs Different AI

Ant International has just announced the introduction of Falcon Time-Series Transformer (FalconTST) 2.0, an AI model focused on finance.

In particular, the model is specialised in forecasting, foreign exchange (FX) movements and risk management. Ant International announced more industry applications to come, but in the meantime, the model seems already at the top of relevant benchmarks.

What MASE and SOTA are, and how benchmark evaluation works

As per the official release, Ant International defined FalconTST 2.0 as SOTA (State-of-the-Art) for its performance with time series data, which is the type of data used in finance and forecasting for its particular characteristics. This kind of data is crucial in these fields because it reveals patterns and trends over time, which is what forecasts are built on.

An AI model is SOTA when it scores higher in performing a specific task. It is not a universal performance measure, but it is based on specific metrics and test conditions.

In this specific case, FalconTST 2.0 is considered SOTA for its performance against the Mean Absolute Scaled Error (MASE) metric. MASE measures how well a model forecasts future data compared to a simple baseline. This simple baseline is really the simplest kind of forecast: it predicts that the next value will be the same as the most recent observed one.

When it comes to evaluating AI models, it is important to understand that there is no official and global kind of testing. Creators initially test their models against publicly available sets of data, in order to allow for peer review. Usually this is the initial step, along with tests against privately owned data, and data from partners who agree to join the launching phase.

Even though teams usually don't reveal which competitors they want to benchmark against, to protect their competitive advantage, we do know the scores that other AI bigs have shared, exactly as Ant International did.
Here's how FalconTST 2.0 ranks against other bigs, owners of models optimised for time-series data and forecasting:

The banks that already run it

As for the external partners using the model, four global banks built FalconTST 2.0 into their own FX systems.

Barclays has built FalconTST into BARX NetFX, its FX hedging platform. Citi runs it together with its own Fixed FX Rates product. At Standard Chartered, the model works with the bank's SCALE FX system inside PathFin.ai, a programme run by the Monetary Authority of Singapore. Deutsche Bank has adopted it too.

For these banks, the work is cross-border currency risk: a client earning revenue in one currency and paying costs in another must hedge, and hedging well depends on forecasting how much of each currency will arrive and when. Hedge for more currency than actually turns up and the company has paid to protect money it never held. Hedge for less and the shortfall sits open to the market. Ant reports that version 2.0 lifted forecasting accuracy above 93 percent for these partners.

Why finance needs its own forecasting AI

The institutions and companies that need a model focused on finance and financial forecasting can't rely on text-based models. A payment institution deals in figures that never sit still: balances, incoming and outgoing settlements, the size of each transaction, the position it holds in every currency.

To hold the right amount of each currency, it has to read where those figures are heading well before they arrive. Ant International is a global payments and financial technology provider, and FalconTST reads that record to forecast what a business will need next.

Ant International reached a milestone in the AI race, and in particular, in financial AI.

Editor's note

Every piece published on The Bright Minded goes through careful verification, but mistakes can happen. If you spot an error, have additional information, or want to flag anything, write to rosalia@thebrightminded.com.