The foreign exchange (FX) market is the world’s largest financial market, far surpassing equities and bonds in daily turnover. In the spot FX segment, market makers manage risk by hedging externally, often at the cost of crossing the bid-offer spread and exposing trading intent. To navigate this trade-off, machine learning models are increasingly used to forecast adverse price movements and reduce market impact, while mitigating the risk of information leakage.
This challenge of extracting signal from noise parallels high-energy physics at CERN’s Large Hadron Collider, where real-time algorithms trigger on rare events—such as Higgs boson decays, produced in about 1 of every 10 billion collisions, and phenomena beyond the Standard Model—amid 600 million particle collisions per second. These are modelled as Poisson point processes, assuming event independence. In contrast, financial time series are inherently sequential, requiring models that capture temporal dependencies. In this talk, I’ll introduce the structure of the spot FX market and present a case study on using machine learning to predict alpha-generating opportunities. I’ll also discuss the limitations of applying AI to time-dependent data and explore the conceptual bridges and key differences between experimental physics and electronic trading.