Enhancing Trading Strategy Robustness through Anomaly Detection

| October 1, 2023

Project overview

This project is conducted within the Lusis - CentraleSupélec - CNRS Chair (https://chaire-lusis.centralesupelec.fr/) in collaboration with the AI department of Lusis AI (https://www.lusisai.com/), which will actively participate in the project and serve as its final client.

A significant number of discretionary and algorithmic trading strategies rely on the concept of patterns, which involve both raw price sequences and combinations of technical indicators. However, most handcrafted trading patterns are based on cognitive biases rather than robust predictive principles. Indicator-based patterns often fail under backtesting, while purely graphical patterns remain too subjective to be effectively quantified.

Machine learning (ML) approaches provide a way to build predictive models that are free from human bias. The hypothesis of this project is that financial markets do contain predictive patterns, but only a small fraction of them is actually useful for forecasting. These patterns are likely to be too complex for human traders or simple algorithms to identify.

In traditional trading, no single strategy is consistently profitable—a single losing trade can negate multiple winning trades. To address this, the project will explore anomaly detection methods such as One-Class SVM to identify the conditions under which a trading strategy is likely to succeed. Anomaly detection is a special case of classification, where the model is trained only on positive examples, constructing a boundary around them. At inference time, new samples are classified as either inside (normal) or outside (anomalous).

The goal is to determine the applicability conditions of a given trading strategy and prevent its execution if these conditions are not met. The approach will then be extended to analyze a portfolio of trading strategies, ensuring that only those operating under favorable conditions are applied.

Deliverables:

  • A GitLab repository with a modular codebase to reproduce results.
  • A detailed experimental report summarizing findings and performance evaluations.