Abstract

Abstract: Machine learning (ML) is driving the evolution of algorithmic trading, but conflicts with the demand for fast execution speed. Although both aim to drive higher profitability, embedding more powerful ML approaches and lowering trading latencies are hard to achieve simultaneously. Offloading ML inference to programmable network devices, also called in-network ML, provides a delicate balance between the two ends of this trade-off. In this talk, I present LOBIN , providing ML-based market movement prediction using high-frequency market data feeds. LOBIN builds limit order books and conducts ML-based inference within programmable switches. This talk describes our solution to the challenging task of mapping LOB constructs and suitable ML models to a commodity switch. Compared with existing solutions, LOBIN predicts stock price movements with lower latency, higher throughput, and a minor impact on ML performance.

Bio: Xinpeng Hong is a second-year DPhil student in the Computing Infrastructure Group at the University of Oxford. His research interest lies in time-sensitive applications of in-network ML.