We present a Bayesian approach to analyze the impact of order flow imbalances from multiple levels of the order book on stock price changes. By examining up to ten levels of the order book, we show that additional levels beyond the best bid and ask enhance model performance, with the average out-of-sample \(R^2\) increasing from approximately 55% to around 80% when all ten levels of order flow imbalances are incorporated. Bayesian model selection shows that, among the candidate models with one through ten levels, posterior model probabilities are concentrated on the ten-level specification. Furthermore, we demonstrate that incorporating cross-sectional order flow imbalances leads to a small but statistically significant performance advantage. Conducting feature attribution analysis using the Shapley value, we find that order flow imbalances from deeper levels and the cross-section yield an average improvement in out-of-sample \(R^2\) of about 19% and 3.4%, respectively.
