NUMERAI STOCK PREDICTION CHALLENGE

Author: Tanmay Singh

---

Project Overview

This repository presents a meta-learning-based solution for the NumerAI tournament, involving the design and training of a two-level stacked ensemble model. The goal is to predict financial data patterns effectively, even in the presence of noise, imbalance, and rapidly changing distributions.

Only a subset of successful runs and model checkpoints are stored in this repository for backup.
Note: The latest commit corresponds to the final submission, with a meta-test correlation of 0.018.

---

Key Contributions

• Developed a stacked ensemble architecture comprising 6 expert models and a meta-model (XGBoost, Random Forest, LightGBM, etc.).
• Applied meta-learning techniques with feature neutralization and balanced sampling strategies to address dataset drift and class imbalance.
• Boosted generalization by combining diverse models and fine-tuning evaluation using meta-testing correlation and feature-neutral metrics.

Tech Stack: Python, Pandas, NumPy, Scikit-learn, XGBoost, LightGBM, Imbalanced-learn, Git/GitHub

---

Setup & Usage Instructions

1. Install Dependencies & Download Assets

python pre-installs.py

Function of pre-installs.py:

• Downloads and installs all required libraries.
• Creates:
  • data: for downloading NumerAI data via API.
  • saved_models: to download pretrained models from Google Drive using gdown.

2. Initialize Models

python Models/models.py

• Initializes all expert models and the meta-model architecture.

3. Train Models

# Jupyter
jupyter notebook train.ipynb

# Or Script
python train.py

• Trains models and saves new pickle files to saved_models.
• To revert to using original Google Drive models, re-run pre-installs.py.

4. Validate Trained Models

# Jupyter
jupyter notebook validation.ipynb

# Or Script
python validation.py

• Evaluates performance on the meta-testing (validation) parquet.

5. Generate Live Predictions

# Jupyter
jupyter notebook predict.ipynb

# Or Script
python predict.py

• Generates and saves predictions in a new predictions directory.

IMPORTANT NOTE

• Before every run, execute pre-installs.py to ensure a fresh slate on every iteration.
• If you're using newly trained models, comment out the block that downloads models from Google Drive using gdown.

Repository Structure

.
├── data/                # NumerAI dataset (auto-downloaded)
├── saved_models/        # Trained model pickles (auto-downloaded)
├── predictions/         # Stores output predictions
├── Models/
│   └── models.py        # Model architecture definitions
├── train.ipynb / .py    # Training pipeline
├── validation.ipynb/.py # Validation script
├── predict.ipynb/.py    # Prediction generation
└── pre-installs.py      # Setup script