The Bank for International Settlements published a working paper proposing a two-stage artificial intelligence framework for financial market monitoring that combines an interpretable recurrent neural network (RNN) with a large language model (LLM) to forecast and explain emerging episodes of market dysfunction. The approach is demonstrated on foreign exchange markets using deviations from triangular arbitrage parity (TAP) in the Euro-Yen currency pair with the US dollar as the vehicle currency. An RNN produces daily forecasts of a 20-day average TAP deviation measure 60 business days ahead using more than 100 daily financial variables, while a dynamic, time-varying variable-weighting mechanism is used to make the model’s inputs and shifting drivers transparent and to guide LLM-based searches of contemporaneous text sources. In pseudo out-of-sample testing over 2021–2024, the model’s signals were smoother than an autoregressive benchmark and pointed to elevated dysfunction risk ahead of episodes including the UK liability-driven investment stress in September 2022 and the March 2023 banking turmoil, while formal tests found the RNN’s forecast losses slightly higher than the AR benchmark but its forecasts remained statistically relevant predictors of future TAP deviations (R-squared around 0.53). To illustrate operational use, the paper shows how an LLM (Gemini 2.5 Pro) can be prompted with a large corpus of July 2023 financial news and the RNN’s high-weight variables to surface key developments for supervisors to monitor ahead of the October 2023 “Treasury tantrum”, and notes plans to release code via the BIS gingado open-source library.
Bank for International Settlements 2025-09-24
Bank for International Settlements outlines RNN and large language model approach to forecast Euro-Yen FX market dysfunction 60 business days ahead
The Bank for International Settlements released a working paper proposing a two-stage AI framework for financial market monitoring, combining an interpretable recurrent neural network (RNN) with a large language model (LLM). Demonstrated on foreign exchange markets, it forecasts and explains market dysfunction using deviations from triangular arbitrage parity in the Euro-Yen currency pair. The paper highlights the model's ability to predict elevated dysfunction risk and plans to release code via the BIS gingado open-source library.