Machine Learning from Randomized Experiments : The Case of the Tick Size Pilot Program

In this paper, I investigate the Tick Size Pilot Program with the goal of policy evaluation beyond average treatment effect. Using a machine learning approach, I study policy effects stock-by-stock on three major market quality measures, percentage quoted spread, consolidated displayed depth, and high-low volatility. For each pilot stock, I test whether it receives significant treatment effects. I find less than half of the pilot stocks in the treatment groups show positive significance for percentage quoted spread; more than 80% shows positive significance for consolidated displayed depth; only less than 5% shows significance for high-low volatility in either direction; the control group stocks rarely show significance for all the outcomes, revealing no spillover effect at the individual level. Tick constrainedness turns out to be useful in explaining differing significance only for percentage quoted spread, but not for consolidated displayed depth. Percentage realized spread, though, appears to explain for the both outcomes: the lower percentage realized spread, the more likely is the null hypothesis rejected, indicating less-profitable stocks for liquidity providers in the pre-intervention periods tend to receive significant effects in the post-intervention periods