Portfolio of sample work I have engaged in as part of my wider learning and experiences
This dashboard is an adaptation of a dashboard I developed during a data analytics internship. It is a Power BI dashboard that visualizes workforce shrinkage using a variety of views, enabling managers to plan jobs optimally. Three data sources are used:
AbsenceData.csvPT63.csvTeamList.csv
Included in the folder are the CSV files mentioned above, the .pbix file (for dynamic viewing), and the .pdf file for static viewing.
In this project, I developed a classification model to predict whether a patient will suffer a stroke, using the stroke prediction dataset from Kaggle. In the .ipynb file, I cover the data extraction, cleaning, feature engineering, and model development. Eleven models were tested and cross-validated, and of these, the best-performing six were hyper-parameter tuned. The best model was the AdaBoostClassifier, which returned the following metrics:
Accuracy: 74%Recall: 76%True Positive Rate: 73%True Negative Rate: 79%
This folder includes the data source as a .csv file and the Jupyter Notebook (.ipynb file) with the data extraction, cleaning, feature engineering, and model development, as well as the results.
This project aimed to test several reinforcement learning models in trading stocks in the first 100 days after its IPO. For the initial test, the Deliveroo (ROO.L) stock was used, particularly due to the negative news coverage around this IPO. To do this, OpenAI's gymnasium was used in conjunction with the gym_anytrading environment. The models that were tested in this environment alongside the gym_anytrading base model, obtained from OpenAI's stable-baselines3, were:
A2C(Advantage Actor-Critic)PPO(Proximal Policy Optimization)DQN(Deep Q-Network)
Summary of performance:
gym_anytradingbase model: Returned a 15.69% lossA2C: Returned a 12.19% lossPPO: Returned a 2.49% profit (this was the best model)DQN: Returned a 55.16% profit - this value is misleading as the algorithm only took long positions after an initial short position (this was the worst model)
Files in this folder include: .ipynb file with testing and comparison of models and renders folder with visualizations of renders of long and short positions taken by algorithms.
N.B: In further commits, the following will be added:
- Hyperparameter tuning for models
- Testing of models on other IPO stocks
- Adding extra features for prediction
This project is a FastAPI app that allows users to load ticker data using Alpha Vantage API and predict volatility using a GARCH model. The project is an adaptation of a similar project I completed as part of the WorldQuant University Applied Data Science Lab.
The app has 2 main functions: fit and predict
- fit: Loads data from AlphaVantage API to SQLite Database, fits a GARCH Model to this data and saves model
- predict: Uses saved fitted to predict future volatility
The Jupyter Notebook in this project also analyses the performance of models using conditional volatilities, residual autocorrelation analysis and walk-forward validation