Machine Learning-Based Latency-Aware Smart Order Routing for U.S. Treasury Securities

In its traditional form, the smart order router only cares about price when deciding where to send a trade; send the order to the exchange currently showing the tightest spread. Implicit in this rule is the assumption that the order comes quickly, rarely the case. Quoted prices go stale between the time you make a decision and the time you get there, and anticipated savings become unanticipated slippage. In this paper, we build a latency aware smart order router for US Treasury securities and compare its performance to a traditional price only router. We take a hybrid approach: real daily constant-maturity as the dealer-to-customer platforms of the Treasury do not provide a free live orderbook feed as listed exchanges do. Specifically, volatility for Treasury yields for the 2-, 5-, 10- and 30-year tenors estimated from the Federal Reserve Economic Data (FRED) service is used to calibrate a simulated intraday order book over six representative dealer venues with their own latency profile. Orders are routed according to a random forest classifier trained on spread, depth, order book imbalance and latency variables; execution outcomes are predicted by anticipating the price that each venue would have displayed when its latency had elapsed. The classifier achieves a test accuracy of 80.1% on the simulated data set of 57,600 rows over 40 trading days, but the generated router only beats the price-only baseline on 2.2% of held-out rounds. The difference in effective execution cost is small, but statistically significant in the opposite direction (paired t-test, t = −3.86, p = 0.0001). The latency-aware router has an average effective execution cost of 6.561 basis points, while the price-only baseline has an average of 6.548, despite routing to venues with lower average latency (47.6 ms versus 49.9 ms). We don't call this a win, we just state it as it is, and point to a classifier that has high recall but low precision on the minority class "best venue", meaning it often sends us confidently to a place that only seemed to be the best. We discuss the implications of this for anyone using a similarly naive latency aware classifier in a bond market setting, and what would need to be changed to fix it.