Sparse portfolio optimisation
Most positions
are noise.
Find the ones that aren't.
Give it a list of tickers and a date range. Variance-reduced proximal gradient methods return a portfolio concentrated in a handful of positions, everything else driven to exactly zero, not just small.
8 of 140 candidate assets selected, the rest, exactly zero.
Four engines
SPGD, Prox-SVRG, Prox-SARAH, Prox-STORM, pick the convergence/variance trade-off that fits, or take the recommended default.
Two dials
Risk-aversion and sparsity strength, exposed directly, not buried behind a black-box "risk score."
Real data
Computed against actual historical returns, run by a variance-reduced stochastic optimiser with proven convergence guarantees.