DeepFolio is a Quadratic Programming-based Python library for large-scale portfolio optimization.
It implements a unified QP framework with an efficient OSQP-ADMM solver (featuring sparse LDL caching and warm starts), and provides full support for real-world constraints such as long/short limits, turnover bounds, and factor neutrality. DeepFolio pioneers ℓ₁+ℓ₂² regularization on mapped factor weights and integrates cutting-edge covariance estimators (AP-Trees, IPCA, RP-PCA, etc.), enabling single-core, high-precision optimization for portfolios with hundreds of assets across multiple rolling-window backtests.
- Flexible data pipeline: Resampling, feature engineering, and custom hooks
- Portfolio Transformer model: Attention-based allocation for multiple assets
- Multiple loss functions: Sharpe ratio, CVaR, and more
- Cost models: Linear, Almgren-Chriss transaction cost modeling
- Portfolio constraints: Gross exposure, leverage cap, and more
- Scenario augmentation: Monte Carlo simulation and scenario generation
- Integrated backtesting: Performance monitoring and evaluation
- Auto re-training & hyperparameter tuning
pip install deepfoliofrom deepfolio import Pipeline, PortfolioTransformer, MultiLossTrainer, Backtester
# 1. Data loading and preprocessing
pipe = Pipeline(hooks=[Pipeline.pct_return])
prices = pipe.load('prices.csv') # Load your price data (CSV with datetime index and asset columns)
returns = prices.values # Convert to numpy array
# 2. Build the model
model = PortfolioTransformer(n_assets=returns.shape[1])
# 3. Train the model with multiple objectives (e.g., maximize Sharpe, minimize CVaR)
trainer = MultiLossTrainer(model=model, losses=['sharpe', 'cvar'])
trainer.fit(returns, epochs=100, batch_size=32)
# 4. Generate portfolio weights for new data
weights = model.predict(returns[-20:]) # Predict weights for the last 20 periods
# 5. Backtest the strategy
backtester = Backtester(model)
results = backtester.run(returns)
# 6. Analyze resultsMIT