cufolio

cuFOLIO is an instruction skill that drives NVIDIA's `cufolio` Python package to build and analyze quantitative stock portfolios using GPU-accelerated Mean-CVaR optimization. It solves the problem of constructing downside-risk-aware allocations from price data: computing returns, generating KDE scenarios, solving weight allocations with the cuOpt GPU solver via cvxpy, tracing efficient frontiers, backtesting against benchmarks, and running scheduled or drift-triggered rebalancing.

The skill acts as a guardrailed wrapper around the installed package rather than reimplementing math. It targets workflows such as building or optimizing a Mean-CVaR portfolio, producing a weights-by-risk-aversion table, plotting an efficient frontier, backtesting an allocation, and rebalancing on a schedule, across S&P 500, S&P 100, Dow 30, or user-supplied price datasets. It sets defaults (for example the CUOPT solver with PDLP method and CvarParameters), points to references/workflows/agent_recipes.md as the source of truth for exact API shapes, and enforces validation guardrails on user CSVs (a date-like index, numeric ticker columns, at least 60 rows). Notably, it treats the first-run download of public S&P 500 market data through the cufolio/yfinance helper as an action requiring the user's confirmation, and it forbids fabricating or substituting price data. It explicitly scopes out price forecasting, neural-network training, vehicle routing, and non-portfolio optimization.

The intended users are quantitative analysts, portfolio engineers, and researchers with an NVIDIA GPU runtime (cuOpt and cuML installed, a matching CUDA extra, and cvxpy exposing cp.CUOPT). Network access is needed only on first run to fetch the default price CSV.