Machine Learning and Speed in High-Frequency Trading

Abstract

The creative destruction wrought by high-frequency algorithmic tradinghas raised increasing concerns about the eect of machine learning behaviorsand ultra high-frequency trading in finnancial markets. By employing a geneticalgorithm with a classifer system as an adaptive learning tool, we address someof these concerns by studying a dynamic limit order market model with asymmetricinformation and varying speeds of high-frequency trading (HFT). We showthat HFT benefits uninformed traders, improves information effciency but reducesmarket liquidity. We find that there is a trade-off where a competition effect erodesthe information and speed advantages of high-frequency traders, increasing tradingspeeds of HF traders, reduces market liquidity but generates a hump-shaped relationshipto the protability of high-frequency traders and information eciency.This research finds there may be potential benefits to throttling the trading speedarms race to improve market efficiency. We also find that strategic algorithmictrading compensates for diminishments in speed advantages, providing an insighton machine behavior in the FinTech age.