Portfolio optimization is a process of allocating funds into financial assets with the goal of maximizing returns over risk. This research is our attempt at utilising machine learning methods to create an optimized portfolio that will perform well into the future. This was achieved by utilising 2 different Random Forest models to forecast stock prices, namely:

•Model 1 – uses technical indicator features.
•Model 2 – uses technical indicator features combined with sentiment features.

Having predicted the stock prices, we then calculate the relevant expected returns. From these we apply an optimisation technique, using a custom mean-variance loss function that optimises the returns over the portfolio risk. Each model repeats this process 500 times before proposing the best suited portfolio as per the results.

• @KeithArogo

Packages used;
•numpy
•pandas
•sklearn
•torch
•SciPy
•seaborn
•os
•matplotlib.pyplot

•Principal Component Analysis.
•Random Forest Regression.
•Optimisation.

•R
•Python

The data set used in this project can be accessed and imported directly into each model, from the folder. The data pre-processing for each model is done in its respective Jupyter notebook. The suggested order to view the Jupyter notebooks begins with:

• Model 1.
• Model 2.
• Graphs.

•PyTorch

‘Model 1’ and ‘Model 2’ in the ‘Models’ folder, can be used to iteratively predict future stock prices and optimize each the 5-asset portfolio’s weights depending on the results. The dataset expected is a 2d DataFrame, with dates as rows and different metrics for stocks in columns. These can be used by simply running all cells in the notebook.
‘News extraction model’ demonstrates the extraction and manipulation of the said sentiment features (news headlines).
‘Comparison metrics’ takes results from ‘Model 1’ and ‘Model 2’ and displays graphs and visuals detailing their differences and similarities to draw conclusions