Abstract:
Traders need timely financial information, yet relevant knowledge is distributed across institutions, markets, and devices and may be absent from centrally trained models. We propose a federated learning application that collaboratively updates a shared language model while retaining raw data locally. Intermittent participation and heterogeneous data create a central challenge: incorporating fresh information before it becomes stale. We ask how availability constrains access to emerging financial knowledge, whether availability-adaptive scheduling improves exposure compared with random selection, and whether this improves adaptation without degrading general financial-language quality. Using heterogeneous client streams and controlled semantic drift, our model achieves lower final private-drift completion perplexity than random scheduling and a stale CrossLM-inspired teacher–student baseline restricted to public financial text, supporting its potential for timely trading decision support.
Bio:
Stefan is doing his 2nd PhD in the Systems Research Group at the University of Cambridge; He has had a prior PhD in management science at the University of Strasbourg. He previously did a Master of Engineering at ETH Zurich and the University of Alberta.
Online meeting link:
