Researchers from the Bank for International Settlements have utilized AI to generate daily forecasts of market dysfunction up to 60 business days in advance.
Predicting financial market stress has historically been challenging. However, AI’s ability to process large datasets and uncover hidden nonlinear patterns reveals promise for change.
In a working paper, BIS researchers developed a two-step tool for forecasting market stress and elucidating the rationale behind their predictions. Initially, a recurrent neural network (RNN) learns from over one hundred daily market indicators to predict the average size of gaps between euro-yen traded directly and euro-dollar-yen traded via the US dollar.
These “triangular arbitrage parity” gaps typically close within seconds under normal conditions, while significant or persistent gaps indicate rising market frictions.
Next, the model identifies which market indicators are most influential in generating its signals. This information can guide a large language model to search for recent news about these crucial indicators to provide timely context.
The system can flag periods of potential dysfunction up to 60 business days in advance. In tests conducted with data from 2021-2024, it accurately identified episodes associated with real events, including the banking issues of March 2023.
When alerts are triggered, the model’s top-weight indicators inform targeted news searches. Case studies demonstrated that these searches uncovered discussions on relevant drivers prior to the emergence of turbulence.
The team concludes: “In short, the tool detects risk early and explains it in accessible terms, helping authorities focus their surveillance and prepare responses.”
Read the paper:
Download the document now 2.1 mb (Chrome HTML Document)