This is a repository of my Bachelor's thesis project Reinforcement Learning vs. 1/n and Mean-Variance Optimization In The Portfolio Allocation Problem. In this project I am testing a number of model-free Reinforcement Learning algorithms based on the framework from https://arxiv.org/abs/2011.09607 written on top of OPENAI stable-baselines. Using data from Bloomberg of the broad equity U.S and EU listed securities I am trying to answer the main research question:

Based on price data and features constructed wherewith, do model-free RL algorithms outperform the Occam's Razor 1/n portfolio and the different linear specifications optimized using a convex objective over a convex set of constraints?

This is an evolution of the weights suggested by the algorithms in the S&P 500 Momentum portfolio:

The project and the notebooks have been developed on python = 3.6.0. All of the dependencies can be installed using:

pip install -r dependencies.txt