Analyzing the Importance of Broker Identities in the Limit Order Book Through Deep Learning

Limit order books (LOBs) have been widely adopted as a trading mechanism in global securities markets, and the degree of LOB transparency is one of the most studied topics in market design. In the past, this issue was mainly researched through the comparison of LOB transparency in a market before and after a policy change, although such instances were rare and occurred decades ago. This article analyzes the importance of broker identities (IDs) in the LOB with respect to price movement predictability by proposing a different approach. By analyzing raw LOB data, an enormous dataset of selected Hong Kong stocks is divided into two parts, namely the prices and order volumes (anonymous LOBs), and a list of broker IDs in the bid and ask queues. A deep learning model is then employed to predict the mid-price movement after 20 ticks. Our result indicates that the best F1 scores of the anonymous LOB and broker ID models are fairly high, ranging from 57.63% to 68.70% and from 53.70% to 59.39%, respectively. When comparing the performance of both datasets, surprisingly, the overall F1 prediction performance based solely on the broker ID dataset can reach, on average, 85.13% that of the anonymous LOB dataset. The contributions of this study are twofold. First, a machine learning-based tool for finance researchers is proposed to quantitatively measure the price predictability of LOB features, and the results of the impact of LOB transparency on traders' profitability are novel as this study is empirical. Second, the empirical result strongly suggests that the broker ID queues in the LOB consist of significant information content for price prediction, and thus, the study provides insights for regulators to determine the appropriate degree of LOB transparency to guarantee a fair market for all investors.