The Conduits of Price Discovery: A Machine Learning Approach

Abstract

When examining information flow into prices, empirical literature usually focusses on direct conduits such as order submissions. Meanwhile, theory suggests that market conditions should have substantial additional effects. Empirical analyses of such effects are methodologically challenging and therefore uncommon. We bypass these challenges using a machine learning technique that allows for multiple conditioning variables. Consistent with theory, price discovery is notably affected by such conditions as the state of the limit order book, price history, bid-ask spread, and order arrival frequency. The state of the book and price history stand out as conduits, whose magnitudes rival that of order submissions.