The repository expands the Triple-Barrier Method proposed by Marcos López de Prado. It introduces specific enhancements:

• Extension to Multiple Barrier Conditions: Dynamically generates multiple barriers through the n_barriers parameter.
• Centering Capability: Includes a center parameter to center the returns for improved analysis.
• Improved Readability: Utilizes standard data science toolkit with pandas and numpy instead of for-loops for each time-step t to improve readability and efficiency.
• Exporting Intermediate Steps: Allows to view intermediate steps as features for model usage (_check_barrier_crossing() method).
• Convenient Properties: Provides useful properties such as transition_probabilities and signals_pa.

Notebooks:

This repository includes two notebooks:

1. example.ipynb: Provides an example showcasing the usage of the BarrierMethod class.

2. study.ipynb: Generates various labels for a range of n and different barriers and tests them for a normal and uniform distribution.

   For some use cases, the goal might be to have (on average) a new signal each week that is approx. normal distributed (neutral most of the time, rare tails) or uniformly distributed.

Examples:

Triple-Barrier Method

The traditional approach to labeling data involves a fixed-time horizon. However, Marcos López de Prado introduced the Triple-Barrier Method in his book Advances in Financial Machine Learning (Wiley, 2018). This method constructs three barriers for each observation:

• Stop-loss Barrier (-1): Indicates a selling opportunity.
• Neutral Signal (0): Denotes a scenario where the time series doesn't cross any barrier within a maximum holding period of n.
• Profit-taking Barrier (+1): Signifies a buying opportunity.

A visual example:

Installation:

pip install git+https://github.com/nkonts/barrier-method.git