Machine learning research predicting stock value using ESG (Environmental, Social, Governance) factors. Random Forest with walk-forward backtesting achieves a ~15% relative accuracy lift over a Logistic Regression baseline across 500+ companies over 10 years.
Conducted as undergraduate research (Nov 2023 – Mar 2024).
| Category | Features | Signal Type |
|---|---|---|
| ESG Composite | Average of environmental, social, governance ratings | Fundamental |
| Momentum | 5-day price percentage change | Technical |
| Trend | 10-day rolling mean of stock prices | Technical |
- Walk-forward backtesting — 5 sequential time-ordered folds
- Always train on past, test on future (no data leakage)
StandardScalerfit only on training fold, then applied to test fold
| Model | Role | Approach |
|---|---|---|
| Random Forest | Primary | Regression → classification (above/below median) |
| Logistic Regression | Baseline | Direct classification |
- Random Forest achieves ~15% relative lift in classification accuracy over Logistic Regression
- Consistent outperformance across all 5 walk-forward folds
- Per-fold accuracy and RMSE values are generated at runtime (dataset is not redistributable)
Run all cells in the notebook to reproduce the backtest accuracy chart and per-fold metrics.
- Data: Place
my_esg_stock_data.csv(stock price + ESG ratings, 500+ companies) in the repo root - Dependencies:
pip install -r requirements.txt - Execution: Run all cells in
esg-driven-stock-value-prediction.ipynb
The dataset was sourced from a university research database and is not redistributable.
MIT