Machine Learning Finance · 2025
CNN Volume-Based Trading Strategy
A machine-learning trading research project using technical indicators, CNN modeling, confidence thresholds, and backtesting.
- My role
- Model experimentation, backtesting, visualization
- Status
- Research project
01 · The idea
A full ML research loop, not only a model file
The project covers data ingestion, feature engineering, model training, thresholded signal generation, backtesting, and visual evaluation.
02 · How it works
From market data to signal confidence
Daily SPY data is transformed into lookback windows with price, volume, and indicator features. The CNN predicts next-day direction, and confidence thresholds decide when the model is allowed to act.
03 · The hard part
Prevent leakage and overfitting from dominating the story
Financial ML work can look convincing while being fragile. The important engineering constraints are clean train/test separation, robust preprocessing, transaction assumptions, and honest comparison against buy-and-hold.
04 · What's next
Risk management and richer validation
Future work should include stop-loss logic, dynamic thresholds, broader assets, walk-forward validation, and sharper treatment of transaction costs.
~/outcomes
Results
- Built a complete data preprocessing, modeling, and backtesting loop.
- Expanded from volume-only features to broader technical indicators.
- Added signal thresholds to control confidence and reduce noisy trades.
~/stack
Built with
- Python
- Pandas
- Scikit-learn
- PyTorch
- Matplotlib
- Seaborn