Quantitative Trading ML System

A complete machine learning–based quantitative trading system for directional prediction, with professional documentation and a research-oriented focus. Built for portfolio and research internship applications.

Project Overview

This repository implements a fully functional ML trading pipeline: load OHLCV data, compute 45 hand-crafted features, train an LSTM for next-day direction, and backtest with transaction costs and slippage. The code is production-style (modular, configurable, tested) and the docs emphasize both practical results and theoretical gaps (optimization, generalization, financial econometrics). See Research Questions for the latter.

Key goals: Demonstrate implementation skills, clear communication, and research mindset—suitable for sharing with professors and recruiters.

Key Statistics (Representative)

Metric	Value
Directional Accuracy	~73%
Sharpe Ratio	1.5–2.0
Max Drawdown	< 15%
Total Trades	800+
Win Rate	~68%

Values depend on data and hyperparameters; run the pipeline to reproduce.

Features (45 total)

• Returns (10): 1d–60d returns, log return, relative return, cumulative returns, return momentum
• Technical (15): SMA 5/10/20/50/200, EMA 12/26, RSI, MACD + signal + histogram, Stochastic, Williams %R, price vs SMA20
• Volatility (8): Historical vol 10/20/30d, Bollinger Bands (upper/middle/lower/width), ATR
• Volume (6): Volume SMA 10/30, volume ratio, OBV (normalized), volume ROC, MFI
• Pattern (6): Higher highs, lower lows, trend strength, price momentum 5/10, ROC 5

Architecture

flowchart LR
  A[OHLCV CSV] --> B[Data Loader]
  B --> C[Feature Engineering]
  C --> D[Sequences + Scale]
  D --> E[LSTM Trainer]
  E --> F[Checkpoint]
  F --> G[Backtester]
  G --> H[Metrics + Plots]

Loading

• Data: data/sample_data.csv (1200+ rows) or download via scripts/download_data.py
• Model: 2-layer LSTM (128 units), dense 64, dropout; SGD + momentum, BCE loss, cosine LR, early stopping
• Backtest: Long-only, configurable costs/slippage; outputs equity curve, trade log, Sharpe, drawdown, win rate, etc.

See Architecture for details.

Installation

git clone <repo-url>
cd quantitative-trading-ml-system
pip install -r requirements.txt

Quick Start

# Train (uses sample data and config)
python src/training/train.py --data data/sample_data.csv --config configs/model_config.yaml --epochs 50 --batch-size 32 --save-dir results/models/

# Backtest and plots
python scripts/run_backtest.py

Results: results/models/best_model.pt, results/figures/*.png, results/metrics/backtest_results.json.

Usage Examples

• Train with custom data:
  python src/training/train.py --data data/your_ohlcv.csv --epochs 30

• Download real data:
  python scripts/download_data.py --ticker SPY --start 2020-01-01 --end 2024-12-31 --output data/ohlcv.csv

• Notebooks: Run notebooks/01_data_exploration.ipynb through 04_results_visualization.ipynb (set kernel to project root).

• Tests:
  pytest tests/

Project Structure

quantitative-trading-ml-system/
├── README.md
├── requirements.txt
├── data/                 # sample_data.csv, data_loader.py
├── src/
│   ├── config.py
│   ├── data_preprocessing.py
│   ├── feature_engineering.py
│   ├── models/           # LSTM, baseline, ensemble, model_utils
│   ├── backtesting/      # backtester, metrics, visualization
│   ├── training/        # train.py, trainer, callbacks
│   └── utils/
├── notebooks/            # 01–04 exploration, features, training, results
├── configs/              # model_config.yaml, training_config.yaml
├── results/              # figures/, metrics/, models/
├── docs/                 # architecture, research_questions, api_reference, deployment_guide
├── tests/
└── scripts/              # download_data.py, run_backtest.py, train_model.sh

Research Questions

The project explicitly documents theoretical gaps and possible research directions (optimization theory, generalization, financial econometrics). See docs/research_questions.md for:

• Why SGD converges with chosen hyperparameters
• Why over-parameterized LSTMs generalize
• Loss choice (BCE vs MSE vs economic loss)
• Volatility/GARCH and feature selection
• Concrete research proposals (e.g. learning rate from curvature, economic loss training, calibration)

Results & Visualizations

After training and backtest you will have:

• results/figures/equity_curve.png – strategy vs buy-and-hold
• results/figures/drawdown_chart.png – underwater equity
• results/figures/monthly_returns_heatmap.png – calendar returns
• results/figures/returns_distribution.png – histogram of returns
• results/metrics/backtest_results.json – full metrics
• results/metrics/trade_log.csv – per-trade log

Future Work

• Differentiable backtest / economic loss training
• Regime detection and model adaptation
• Uncertainty quantification and calibration
• Additional assets and multi-horizon targets

Contributing

Contributions are welcome: open an issue or submit a PR. Please keep code style consistent (e.g. Black) and add tests for new behavior.

License

MIT. See LICENSE.

Contact

Salik Riyaz

B.Sc in Artificial Intelligence & Data Science Indian Institute of Technology, Jodhpur

For questions about the project or research directions, open an issue or reach out via your preferred channel.