In Network ML project: LOBIN

Work by Xinpeng Hong, Changgang Zheng, Stefan Zohren and Noa Zilberman 

In the algorithmic trading industry, there are two types of market participants. One specializes in high-frequency trading (HFT) approaches while using quick and simple decision logic. The other utilizes powerful machine learning (ML) algorithms, while paying with increased decision's latency. In-network ML provides algorithmic trading a midway solution: using machine learning, while at the same time providing low latency. 

A common problem in algorithmic trading is predicting future price movements from market microstructure signals. While directly using order-based market messages (MBO) doesn't achieve satisfactory forecasting performance, electronic limit order books (LOBs) were shown to provide good prediction performance based on ML models. For a particular stock, a real-time LOB is constructed from unmatched limit orders that are predetermined with specific prices.  

LOBIN studies the application of in-network ML to market prediction using LOBs. It is a proof of concept for future price movement prediction by maintaining and updating LOBs in the programmable data plane based on high-frequency market data feeds. 

 LOBIN was implemented in P4 on BMv2 and on Intel Tofino/Tofino2 switch-ASIC, with its workflow shown in the figure below. Historical market data feeds are used to train an ML model on a server, using LOB-based features. Using Planter, the trained model is mapped to the data plane, and table entries are generated. For a given stock, when a new order is received, LOBIN updates the LOB and extracts features from it for ML inference.

LOBIN provides high prediction accuracy (compared with a server-based benchmark), while achieving microsecond-scale latency.