DeepDow

What it is

A Python package that connects portfolio optimization with deep learning by merging market forecasting and allocation into a single, fully differentiable network that outputs weights in one forward pass. Built on PyTorch, it offers differentiable convex optimization via cvxpylayers, clustering-based allocation algorithms, multiple dataloaders, and losses such as Sharpe ratio and maximum drawdown, with CPU and GPU support. Researchers experimenting with neural networks for buy-and-hold weight allocation are the primary audience. It is a research framework rather than a trading product, focused on allocations held over a horizon rather than active trading, so short-term transaction costs are not a primary concern.