Iceberg Detection via Synthetic Limit Order Book Simulation

Overview

This project explores whether iceberg orders (large hidden institutional orders broken into smaller visible chunks) can be detected using only observable order book behavior. Since real exchange-level order book data is expensive and often inaccessible, our approach is to build a synthetic limit order book simulator, manually inject iceberg-style execution behavior, and train a model that outputs a probability that an iceberg is present at a given price level.

---

Goal

Build a system that can:

1. Simulate a realistic limit order book (LOB) with limit orders, market orders, and cancellations.
2. Inject iceberg orders manually (known ground truth).
3. Extract microstructure-inspired features (resilience, refills, volume absorption, etc.).
4. Train a lightweight ML model (logistic regression / random forest / small NN) to output:

P(iceberg exists at this price level | recent order book behavior)

---

Key Project Idea

We simulate a market where most orders behave normally, but occasionally a large participant enters the market and executes using iceberg-style logic:

• parent order size: large (e.g., 100,000 shares)
• visible tip: small (e.g., 200 shares)
• whenever the tip is filled, it replenishes automatically
• optional randomization in timing and tip size

Because we injected the iceberg ourselves, we have labels, enabling model evaluation.

---

Model Inputs (Features We Track)

• Price-level resilience (how much volume is absorbed before price moves)
• Refill frequency at a given level
• Persistence under aggressive flow
• Volume traded at a level without depletion
• Order imbalance (bid vs ask pressure)
• Trade clustering at a level
• Queue depletion vs replenishment asymmetry

---

Modeling Approach

We start with simple, explainable models:

• Logistic Regression (baseline probability model)
• Decision Tree / Random Forest (nonlinear patterns)
• Optional: small neural network (only if needed)

---

Evaluation Plan

Since we have ground truth labels (we know where we injected icebergs), we can measure:

• Precision / Recall

• ROC-AUC

• False positive rates under high-noise regimes

• Model performance under different market conditions:

  • high vs low volatility
  • many vs few cancellations
  • multiple icebergs vs none
  • different iceberg refill strategies

---

Deliverables

By the end of the project, we aim to have:

• A working LOB simulator

• A reproducible dataset generator

• A feature extraction pipeline

• A trained classifier that outputs iceberg probabilities

• Visualizations showing:

  • order book heatmaps over time
  • iceberg vs non-iceberg behavior
  • probability spikes when hidden liquidity is present

---

Haven't Figured it Out yet

We can pivot to these ideas below if the current project is too hard:

• simplify to tracking only top-of-book instead of full depth
• simulate fewer order types
• switch from ML to purely statistical detection
• treat iceberg detection as anomaly detection instead of classification

---

Tools / Tech Stack

• Python 3
• NumPy / Pandas
• scikit-learn
• matplotlib
• Jupyter notebooks (for experiments + visualization)

Runs locally on a laptop.

---

Final Pitch (One Sentence)

We simulate realistic order books, manually inject iceberg-style hidden liquidity, and train a lightweight ML model to estimate the probability that an iceberg exists based only on observable order book dynamics.