On the evaluation of intraday market quality in the limit-order book markets: a collaborative filtering approach

This study proposes a methodology for evaluating market quality of individual stocks in the high-frequency domain (short term) by applying a recommender system that has become ubiquitous in our daily lives, especially when running internet apps. In the first place, it is not easy to evaluate market quality such as the "true" liquidity of individual stocks. In particular, in situations where liquidity for a short-time period is to be evaluated using high-frequency data, the lack of observations can become severe for numerous stocks. Since stocks that have exhibited similar behavior in the past are expected to perform so in the future as well, one can expect that collaborative filtering, which is nowadays the main approach of recommender systems, can work effectively for the market quality measure "estimation" problem for stocks. However, in some occasions, "standard-type" collaborative filtering methods may not work well, especially when data sparsity (or scarcity) is severe. Specifically, in this paper we adopt a regression-based latent factor model (RLFM), a "hybrid-type" collaborative filtering proposed by Agarwal and Chen (Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD'09), 2009). It has a hierarchical linear structure designed to address the so-called "cold-start problem" in the recommender systems literature. In this study, we take on liquidity and volatility as market quality measures. The liquidity measure used in this paper is the inverse limit order book slope proposed by Deuskar and Johnson (Rev Financial Stud 24(3):721-753, 2011). For volatility, we adopt the realized volatility based on tick-by-tick mid-quote prices. To investigate the effectiveness of the method in consideration, empirical analysis was performed using high-frequency limit-order book data from the Tokyo Stock Exchange. The data period is the first 3 months of the year 2019, with which regularly-spaced 5-min aggregate datasets were formed. The explanatory variables in the regression term are six variables related to observed market activities of individual stocks such as logarithmic return and share volume, three variables related to the static attributes of individual stocks such as whether it is an ingredient of the Nikkei 225 index, and the industry category it belongs to. There are also time polynomial terms up to the order of 6 to capture the average movements of the whole market along the time axis. As a result of the empirical analysis, various characteristics that characterize market quality were identified from the estimated regression coefficients obtained by fitting the RLFM model to the training dataset. The results suggest that our approach is robust to the degree of sparsity (or scarcity) and flexible enough to deal with various data environments. There was room for improvement of the methodology in terms of prediction accuracy.