Predicting Stock Price Movements in High-Frequency Trading

This research focuses on utilizing techniques to predict mid-price movements and bid-ask spread crossings in high-frequency trading. The study explores the use of support vector machines, random forests, and gradient boosting algorithms on limit order book (LOB) data, capturing stock price movements at the millisecond scale. The dataset is split into two time periods: 9:30–11:00 and 11:00–12:00, with the former used for model fitting and assessment. Various performance measures, including recall, precision, and F1-measure, are employed to evaluate the machine learning models. Additionally, summary statistics are derived to fit a logistic regression model and compare its results with the machine learning approaches. The findings indicate that while the models perform well during the 9:30–11:00 period, achieving precision levels of 70%–80%, they yield unsatisfactory results during the 11:00–12:00 strategy timeframe. The discussion highlights the challenges in achieving high backtesting accuracy and suggests potential avenues for improvement, such as incorporating additional time-sensitive features and considering time-series techniques. Furthermore, the study addresses the importance of factoring in transaction fees when evaluating real-life profitability, proposing an approach to reclassify spread cross labels to account for transaction costs.