Skip to content

Please Remember...

This is a fictional example and not real documentation. The purpose is to demonstrate my technical writing.

Important to Know

  • Quant models don’t prove guilt. They highlight trades that look unusual based on known patterns.
  • Further investigation is always needed. The model is a tool to guide your attention, not make final judgments.

← introduction | quant models (explanation)

Trading Fraud and Detection Models

Quant Models Used in Fraud Detection

The table below outlines common types of financial misconduct, a short description, and the commonly used quant models to detect them.

Fraud Type Description Main Model to Know Model Insight Other Models Commonly Used
Insider Trading Trading using non-public information Logistic Regression Great for binary classification (fraud vs. not fraud); interpretable. Behavioural Profiling, SVM
Spoofing Placing then cancelling fake orders to manipulate price Random Forest Captures complex patterns; useful when false positives are costly. Decision Trees, Anomaly Detection
Wash Trading Buying and selling to simulate market activity K-Means Clustering Groups similar trades to detect outliers/unusual patterns. DBSCAN, Isolation Forest
Front Running Acting on early knowledge of large trades Anomaly Detection Flags unusual trading ahead of big moves; often unsupervised. Neural Networks, Logistic Regression
Layering Multiple fake orders layered across price levels Support Vector Machines (SVM) Effective for high-dimensional data and subtle fraud signals. Random Forest, Logistic Regression

How the Model “Knows” What It Knows

Quant models are typically trained as follow:

  1. Start with a set of known fraudulent trades and a clean set of non-fraudulent trades.
  2. Feed a third, mixed dataset into the model to test if it can tell the difference.
  3. Correct the model when it gets things wrong.
  4. Repeat steps 2 and 3 until it consistently gets it right.

How Models Work Together

Models can be run in two ways:

  1. Rule-Based Detection

    • Each model has a fixed condition.
      • Example: “If trade size is greater than 3 times the average, then flag it for possible fraud.”
      • Downside: Often leads to false positives.

  2. Weighted Probability Framework

    • Each model outputs a probability score. These scores are then combined to give an overall likelihood of fraud.
      • Example:
        • Three models each give 0.6, 0.7, and 0.8.
        • The combined risk = (0.6 + 0.7 + 0.8)/3 = 0.72
        • 0.72 means 72%
    • This weighted approach is more flexible and more commonly used in modern systems.

Typical Workflow

  1. A client (e.g., a trading firm) sends post-trade data to Company-X.
  2. The data is compared against historical market behaviour.
  3. Quant models run (individually or together) to evaluate the likelihood of fraud.
  4. The client receives a spreadsheet or report showing flagged trades for further review.

What comes next?