Execution Algorithms: A Real-Time ML Approach

This blog is for testing.

---

A milestone write-up on real-time execution modeling using limit order book features and dynamic thresholding.
🧠 Written by Binxi Li, Sirui Zhang, and Sheng Zhang @ Cornell Financial Engineering (Spring 2025)

---

📄 Execution Algorithms: A Real-Time ML Approach

📘 Project Overview

In this project, we built a real-time execution algorithm to decide when to buy (or sell) one share per minute using limit order book (LOB) and message flow data.

We combined market microstructure intuition with machine learning classifiers, and designed fallback mechanisms to guarantee exactly one trade per minute. Our primary goal: achieve better prices than a TWAP benchmark by exploiting order-level features.

---

🧮 Features from LOB

We engineered several per-tick features to drive our model decisions:

• Spread: ask − bid (proxy for transaction cost)
• Order Imbalance: (bid_volume - ask_volume)/(bid_volume + ask_volume)
• Momentum: 5-tick change in mid-price
• VWAP: Size-weighted average trade price (last 10 orders)
• Execution Intensity: Rolling count of recent execution messages
• Order Direction Mean: Smoothed sentiment signal from buyer/seller tags
• Time Pressure: Scaled seconds-in-minute (0 at start, 1 at 59s)

These features help capture microstructural patterns like urgency, trend, and liquidity.

---

🧠 Model Logic

We used both a Multilayer Perceptron (MLP) and a Decision Tree Classifier to predict whether the current tick lies in the lowest 3% ask price of that minute.

To avoid missing trades, we implement a time-adaptive confidence threshold:

• 0–15s: p > 0.95
• 15–30s: p > 0.90
• 30–45s: p > 0.85
• 45–60s: p > 0.80

If no signal is triggered by second 60, a fallback trade is placed at current ask.

---

📊 Backtesting Results (AAPL)

• Average Algo Price: $584.0644
• TWAP Price: $584.0826
• Improvement: –$0.0182 per share
• Hit Rate (lowest price in minute): 63.2% (vs. 55.4% baseline)
• Precision/Recall: Tree = 0.24 / 0.63, MLP = 0.13 / 0.66

Most executions happened in the first 15–30 seconds, validating the effectiveness of our dynamic thresholds.

---

🔍 Key Takeaways

• Feature engineering using LOB is powerful even without deep learning
• Time-dependent thresholds help balance price quality vs. execution certainty
• Real-time decision-making can beat simple baselines with careful design
• Our framework is extendable to sell-side, multi-asset, or cost-sensitive settings

---